AI can write an email.
It can also help decide who receives that email, choose the offer, generate the landing page, answer the customer's questions and help the customer decide whether to buy from someone else entirely.
I think we underestimate the change when we treat AI as a collection of tools for getting our existing jobs done faster. There is plenty of value in that, obviously. I would quite like fewer hours spent moving things between spreadsheets, as I'm sure we all would.
However, the bigger change involves the whole relationship between businesses and their customers: how we understand people, what we put in front of them, and how they choose to deal with us.
From conversations I've had with friends and colleagues, and the amount of time I spend on Twitter, I definitely get a sense that basically everyone feels like they are being left behind. Which doesn't make sense of course, not everyone can be left behind, but I get it.
This is my attempt to map the subject of marketing to AI in a useful way.
I'm not saying I have all the answers here, and that I'm some kind of expert. But I do spend a lot (probably way too much) of time staying up to date with AI, and I do think a lot about how computers impact our lives. What I have done here is spend time trying to understand each discipline of marketing and how it is changing, for both a mental documentation exercise for my self and, I hope; a learning exercise for those who read it.
This is kind of a long article, so I've included a contents here. If you work in a single discipline and don't care about the rest, please do skip to that part. If you have a spare 30-45 mins to give to this, I'd be flattered, and I'd also appreciate any feedback or ideas. You can find me on Twitter.
Three ways AI changes marketing
I find it useful to separate three things:
| Lens | The question | A simple example |
|---|---|---|
| The work | How does the team do the job? | Producing and checking twenty versions of an advert. |
| The marketing | What actually reaches the market? | Different versions selected for different situations. |
| The customer | How does someone discover, judge or buy? | Asking an assistant to compare the advertised product with alternatives. |
These changes can happen independently. Making a TV advert cheaper doesn't mean the viewer watches television differently.
An AI answer replacing a website visit changes discovery even if the brand hasn't adopted a single AI tool.
We also need to stop using “AI” as shorthand for generating text. Predictive models estimate things such as purchase likelihood. Classification systems sort enquiries or content. Computer vision interprets images and video. Generative models create material. Agents combine models with tools and permissions to take steps towards a goal.
An agent might investigate a campaign, prepare changes and request approval. A chatbot explaining the campaign is doing a different job. Equally, a scheduled email or an ordinary pricing rule doesn't become AI because someone has updated the sales deck.
Across the guide, I use the same three lenses. The examples are either linked implementations or illustrative applications, rather than claims that every business is already doing all of this.
Strategy and research
Market research and competitive intelligence
Research helps us understand customers, markets and alternatives before committing money. AI can make evidence easier to work with, but it also makes invented evidence exceptionally easy to produce.
The work: A useful application is organising interview transcripts, comparing competitors' claims and finding recurring objections in sales calls. Keep the underlying sources attached so someone can check the interpretation. Asking a model to role-play a customer can help develop questions; it doesn't establish what actual customers believe. Kantar's discussion of synthetic samples explains why validation against real people matters.
Ipsos's February 2026 paper on synthetic data boosting takes this further, setting out checks for statistical similarity, usefulness, fairness and expert validation when extending a dataset. That is a much more considered approach than asking a chatbot what thirty imaginary customers think. The interesting opportunity is expanding what real evidence can tell us.
The marketing: Faster research could support more specific propositions. A software business might find that buyers care less about its long feature list than the effort of replacing their current system. That should change the campaign, demonstration and migration offer, rather than simply add another slide to the research presentation.
The customer: Their alternatives are changing too. They may compare your product with an AI-assisted workaround, an existing subscription or doing nothing. Our internal competitor list is only one view of the market. Research needs to discover the shortlist in their head.
Segmentation and targeting
Segmentation groups customers in useful ways; targeting decides which groups deserve attention. Both become more interesting when behaviour can be analysed beyond a few broad demographic labels.
The work: Teams can use models to identify patterns in purchase history, needs and engagement, then investigate whether the groups are commercially meaningful. An outdoor retailer might distinguish occasional weekend walkers from people replacing heavily used equipment. The model still needs sensible inputs, enough evidence and someone capable of spotting nonsense. A beautifully named segment of eleven people is not necessarily a strategy.
For a concrete implementation, Adobe's Predictive Audiences documentation, updated in May 2026, describes learning from existing audience groups and classifying other visitors using first-party behavioural signals. The marketer supplies the starting groups. That matters: AI can extend a segmentation approach, but it doesn't establish that those were the right groups to begin with.
The marketing: Messages and distribution can reflect the job people are trying to do. Durability evidence may matter to frequent users; straightforward buying advice may matter to beginners. The useful change is connecting a meaningful difference to an appropriate response, rather than generating hundreds of personas because the software permits it.
The customer: Relevance can improve, but people also move between situations. Someone buying for themselves today might be choosing a gift tomorrow. Treat a predicted segment as a working assumption, not a permanent identity. Leave room for discovery and let explicit preferences override a guess. Otherwise personalisation can become a machine confidently misunderstanding the same person for several years.
Positioning, brand and go-to-market strategy
These decisions establish who the business serves, why someone should choose it, and how it will reach them. They determine what all the other activity is supposed to achieve.
The work: AI can compare positioning options, organise evidence, challenge an argument and turn a strategy into draft briefs. The difficult part remains choosing. A model doesn't bear the consequences of pursuing the wrong customer, making an unsupported promise or launching into a market the business cannot serve. The team needs access to customers, commercial information and decision-makers, rather than just a better prompt.
The marketing: A maintained source of product facts, approved claims, distinctive assets and customer evidence can help teams produce more consistent campaigns. For a software launch, that could connect the positioning to sales demonstrations, onboarding and comparison pages. Brand is broader than visual identity; making the logo consistent won't rescue an inconsistent experience.
Kantar's NeedScope AI analyses the emotional signals in imagery and music against its positioning framework. Its May 2026 explanation of NeedScope Positioner also makes the customer input explicit: survey data combined with consumer context. I like the principle here: use a consistent method to interrogate positioning, with evidence behind the choices.
The customer: AI-assisted comparison may expose vague propositions. If an assistant can summarise ten suppliers as “an innovative, customer-focused solution”, nobody has won much. My view is that specific promises with visible proof become more useful here. Decide what you can credibly own, then make it easy for both a person and their assistant to understand.
Pricing and commercial offers
Pricing turns perceived value into revenue. AI can support forecasting and offer selection, but dynamic pricing, personalised pricing and ordinary discount rules are different things.
The work: Models can help estimate demand, explore price sensitivity and identify customers who buy only during promotions. Teams can compare scenarios before changing the offer. For example, a retailer could investigate whether a bundle improves contribution after delivery and returns, rather than celebrating a higher basket value while quietly losing money.
The marketing: Offers can become more responsive to stock, timing or customer circumstances. That doesn't automatically mean charging each person the maximum a model thinks they will tolerate. The CMA's dynamic pricing work is a useful reminder to consider transparency and consumer understanding alongside revenue. Even a technically effective model can create a commercially terrible impression.
In September 2026, Philippine retailer Healthy Options expanded its RELEX partnership to include price optimisation, aiming to replace spreadsheet-led decisions with price-elasticity data connected to forecasting and stock planning. This is an announced deployment, rather than a results study, but it makes the operational change tangible: pricing becomes connected to more of the business.
The customer: Assistants could make total-cost comparisons easier, including renewal prices, delivery and cancellation terms. Some of that research is possible already; widespread autonomous negotiation remains a more speculative prospect. A sensible response is clear pricing and explainable offers. If the attractive headline survives only until someone reads the small print, faster comparison is unlikely to be your friend.
Creative and production
Creative strategy and copywriting
Creative strategy connects a customer insight to an idea people might notice and remember. Copy gives that idea a voice. AI changes how many possibilities we can explore, but quantity isn't the same as range.
The work: Teams can develop territories, test alternative explanations and adapt an approved concept. The useful skill is directing and selecting the work. A 2024 experiment by Anil Doshi and Oliver Hauser found that AI suggestions improved ratings of individual short stories while making the stories collectively more similar. That is a finding about a particular writing task, not proof that every AI-assisted campaign will look identical. It does describe a risk worth watching.
A 2026 study with 98 advertising professionals offers a useful addition: participants given prompt training before using AI generated more ideas than the other groups. Access to the tool alone wasn't enough to produce that result. The finding concerns idea quantity, but it gives creative teams a practical reason to invest in how people use the tools.
The marketing: A brand could explain the same proposition differently to a finance director and a practitioner without starting every execution from scratch. Both versions should still express the same recognisable idea. Otherwise the campaign becomes a collection of locally optimised sentences with no personality holding them together.
The customer: Competent copy may become less distinctive as more businesses can produce it. Surprise, timing, humour and a worthwhile point still matter. I would use AI to explore the territory, then ask whether the result says anything someone would care to hear. “Professional and engaging” is a remarkably low creative ambition.
Graphic design, photography and illustration
Visual production includes everything from exploratory concepts to the images someone uses to judge a real product. Those jobs need different standards of accuracy.
The work: Generative tools can help with moodboards, backgrounds, retouching, layout options and asset adaptation. Adobe's Firefly production announcements show how these capabilities are moving into repeatable production systems, including models customised for a brand. The opportunity is particularly clear in adaptation: taking an approved direction across many sizes and contexts.
Adobe's September 2026 breakdown of its Times Square campaign shows the work involved: existing imagery was expanded and recomposed for huge screens, with new location-specific assets and motion treatments. Firefly supported production; people still coordinated the creative direction, review and approvals. It is worth reading for the workflow, rather than just looking at the finished pictures.
The marketing: Smaller teams can attempt visual ideas that previously required more production budget. An illustrative campaign can explore an impossible scene; a product catalogue needs to represent what arrives in the box. Generating a nicer-looking finish on a sofa is a fairly efficient way to generate returns as well.
The customer: Images help people infer quality, scale, texture and suitability. We should preserve that function. Use real evidence where it matters, and distinguish imagined contexts from demonstrations. The human contribution increasingly includes art direction, selection and knowing when to commission original photography. Access to attractive images becomes easier; having a visual identity people recognise is still a separate achievement.
Video and audio production
Video and audio bring ideas to life through movement, performance and sound. AI can change individual production stages without taking over the whole creative process.
The work: Storyboards, rough edits, transcription, background removal, voice experiments and alternative cuts can become easier to produce. An editor can spend less time preparing versions and more time deciding what makes the piece work. Adobe's production tools also include video and translation capabilities, but a capability on a product page doesn't establish the quality of your finished film.
There are actual campaigns to look at now. In Native Foreign's Paris Hilton/Carl's Jr. work, reported in February 2026, ElevenLabs was used for the audio production. And Adobe's August 2026 release brings music, speech and sound effects into the same creative environment as video. Together, these show how individual production jobs are being connected, without establishing that an entire film can be left to run itself.
The marketing: A product launch could support a main film, short demonstrations, vertical edits and accessible versions within a budget that previously covered less. That creates opportunities to explain things properly, as well as opportunities to upload sixteen versions of something nobody wanted to watch the first time.
The customer: People may encounter more synthetic voices and scenes, but their expectations depend on context. An animated explanation and an apparent customer testimonial make very different promises. Real demonstrations, interviews and performances remain useful evidence. For audio in particular, pronunciation, pacing and emotional appropriateness need listening to, rather than assuming that a clean transcript means a convincing recording. Test the actual experience on the devices people use.
Translation and localisation
Translation changes the language; localisation adapts meaning, references and expectations for a market. AI can expand access, provided we do more than swap the words.
The work: Teams can prepare subtitles, translations, voice tracks and local variants more quickly. YouTube's automatic dubbing updates illustrate how distribution platforms are absorbing some of this production work. Subject experts and fluent reviewers remain important for product claims, technical terminology and the occasional joke that really should have stayed in English.
The marketing: An educational video or useful product demonstration can reach audiences a business couldn't previously justify producing for. Start with material that already helps people, then adapt the proposition, examples and call to action. There is little point translating a campaign beautifully if the payment methods, delivery promise and support service still assume everyone lives in Manchester.
The customer: Better language access can make a business easier to understand and more welcoming. Poor localisation can create false expectations about what is available. Check the whole journey, including prices, measurements, availability and help. The strategic opportunity is entering a market with a coherent experience; the cheaper translated asset is only the first part of that job.
Web design and development
A website combines communication with a functioning service. AI-assisted building lowers the effort of creating a first version, which makes it easier to discover whether an idea deserves a second one.
The work: Marketers can prototype a calculator, comparison tool or landing page and discuss something concrete with designers and developers. This can reduce the distance between an idea and useful feedback. It also increases the need to distinguish a convincing demo from a maintained product. Authentication, accessibility, security, performance and data handling don't disappear because the page arrived quickly.
Wix Harmony, launched in January 2026, combines natural-language site creation with direct visual editing. That combination is useful for marketing teams: describe an idea, get something tangible, then adjust the layout yourself. It also illustrates why control after generation matters as much as the initial “look what I made” moment.
The marketing: Useful interactive experiences become possible for smaller projects. A supplier might build a specification checker that saves buyers an awkward phone call. That could do more for acquisition than another generic guide. Performance and crawlability still deserve attention if people are going to find and use it.
The customer: The best outcome is less effort: clearer choices, faster pages and tools that actually work. The worse outcome is another conversational interface blocking a task a simple form could have completed. Start with what someone needs to do, then choose the interface. A chat box is a design decision, rather than proof of progress.
Content and owned media
Content strategy, thought leadership and original research
Content earns attention by helping someone understand or do something. AI makes the packaging easier, which raises a fairly important question: what information have we actually contributed?
The work: Teams can organise research, analyse datasets, prepare outlines and adapt expert material. A useful workflow starts with evidence the business has earned: customer questions, experiments, specialist experience or original data. AI can help examine that evidence; the team remains responsible for its interpretation and methodology. A fluent explanation of a weak dataset is still a weak finding.
The marketing: A retailer might publish a repairability study based on its workshop records. A software company could explain an industry trend using properly aggregated product data. These assets give other people something to reference and the brand something competitors cannot reproduce by asking the same model the same question. This is the argument behind Write for People, Obviously.
Edelman's own research is a good example of that asset in practice. Its June 2026 report on brand growth draws on 17,688 respondents across 15 countries, giving the agency evidence around which to build a point of view. AI can help analyse or explain findings; it cannot conjure those respondents and turn their imaginary answers into equivalent research.
The customer: Summaries may satisfy basic information needs, so deeper content needs a reason to exist beyond repeating the summary. Show the method, demonstrate the product, share the experience and acknowledge the limitations. I think hero content becomes more valuable when its originality is visible.
Having a human name at the top is not, by itself, the evidence.
Videos, newsletters and editorial programmes
An editorial programme gives people a reason to return. It depends on a recognisable perspective and a worthwhile selection of material, as well as the ability to publish regularly.
The work: One properly researched interview can become an edited video, a transcript, a newsletter and several focused clips. AI can help locate relevant moments and draft the adaptations. Someone still needs to check that a punchy excerpt hasn't quietly removed the qualification that made the original statement true. “Repurposing” shouldn't mean making the expert sound more certain with each new format.
Descript's SaaStr case study describes exactly this kind of operation: the team edits event recordings and interviews in-house, uses AI-assisted audio clean-up, and marks extracts in the transcript to turn into clips. The underlying material comes from real experts and events. The software helps more people encounter it, rather than supplying the expertise itself.
The marketing: Production savings can support a broader range of formats, including captions and shorter explanations. A weekly newsletter could organise material around different reader needs while retaining one editorial position. The opportunity is making useful knowledge easier to consume, rather than treating every channel as another place to paste the same text.
The customer: People may watch a clip, read a summary or ask an assistant to extract the practical advice. Design material that remains useful in those forms and gives interested readers a reason to go deeper. A trusted editor can also save people from the effort of filtering an abundant supply of competent content. Selection is part of the product.
Organic social media and user-generated content
Organic social combines distribution, conversation and public evidence of what a brand is like. AI affects the production of posts and the systems deciding who sees them.
The work: Teams can use AI to sort feedback, prepare formats and adapt approved material. It is particularly useful for turning a substantial piece of work into accessible extracts. But collecting genuine customer experiences and responding appropriately still require attention. A fabricated customer quote is not user-generated content, however convincingly the model adds a typo.
The marketing: More output competes for finite attention. Meta's March 2026 guidance on original content describes rewarding original work and reducing distribution for unoriginal material on Facebook. That doesn't establish a universal penalty for AI assistance. It does undermine the idea that endlessly recycling other people's posts is a durable strategy.
The customer: A real demonstration, useful answer or credible creator can help someone judge a business. Engagement bait and generic comments create activity without necessarily adding trust. I would use production savings to document more actual work: product development, customer questions, staff expertise and things that went wrong. There needs to be something happening behind the content calendar.
Community and customer participation
A community gives people access to one another, not just to the organisation hosting it. Its value often comes from knowledge and relationships that are difficult to obtain elsewhere.
The work: AI can help find unanswered questions, surface relevant discussions and prepare summaries for moderators. Imagine a professional community where a new member's technical question is matched with a useful discussion from six months earlier. The system should preserve attribution and context, with moderation decisions reviewable by someone who understands the group.
Reddit's July 2026 account of its anti-spam work provides a useful example of AI protecting participation. It describes using language models to identify coordinated fake behaviour, alongside human moderation. The platform is using the technology to make human conversations more trustworthy. That is a considerably better community objective than manufacturing more comments.
The marketing: A business can make accumulated community knowledge easier to use and identify recurring needs worth addressing. Those insights might inform an event, a product improvement or a practical guide. This is quite different from populating the group with automated contributors to make the member count look lively.
The customer: People join communities partly for the experience of dealing with other people. Automation should reduce repetition without making it unclear who they are speaking to. Summaries can help someone catch up; an undisclosed bot pretending to share personal experience damages the premise of the space. My expectation is that well-run communities retain value precisely because participation, reputation and relationships require more than generating an answer. That is a strategic judgement, rather than a guaranteed growth forecast.
Search and discovery
SEO and answer engine optimisation
SEO helps people and systems discover and understand useful web content. AEO extends that problem into answers, recommendations and assisted decisions, often before a website visit happens.
The work: AI can support technical investigation, content analysis and research into customer questions. The foundations remain substantial: accessible content, sensible internal links, reliable indexing signals and accurate product information. Google's guidance for AI features explicitly retains core SEO practices; it doesn't require a special AI schema or text file.
The marketing: Brands need evidence that helps a decision: clear comparisons, original research, product details, documentation and credible coverage elsewhere. Being retrieved, being cited and being recommended are different outcomes. My AEO strategy guide goes further into the practical work.
The customer: Answers can reduce the need to click. Pew's study of US browsing in March 2025 found traditional-result clicks on 8% of visits with an AI summary, versus 15% without. It was observational, rather than a controlled experiment. Google reported relatively stable total organic clicks that August. These measure different things: clicks per visit and aggregate traffic can move differently. Neither tells us exactly what happens to your business. Measure your customers, visibility and commercial outcomes alongside traffic.
For a more recent view, Similarweb's September 2026 analysis estimates that worldwide AI referrals more than doubled between its two annual comparison periods, while showing substantial differences between industries. Those are modelled referral visits, not every decision influenced by AI. The useful lesson is to look at your category and the absolute traffic alongside the exciting growth percentage.
App stores and marketplace discovery
App stores and marketplaces organise choices inside someone else's environment. AI adds more interpretation between what a customer asks and which products receive attention.
The work: Product teams need accurate listings, clear capabilities, reliable feeds and a process for learning from reviews. Apple's App Store Connect updates describe AI-generated app tags with human review. It is a concrete example of platforms interpreting product information, rather than relying entirely on the labels a developer supplies.
Amazon's 2026 shopping and advertising update takes this into conversation: sponsored product prompts can lead shoppers into questions about the item. The listing now needs to support an answer, not just a click. Information about compatibility or running costs suddenly has another job to do.
The marketing: Discovery can depend on how well the platform understands a product's suitability for a task. A budgeting app should explain who it serves, which accounts it supports and what it actually does. For physical goods, compatibility, dimensions and availability can be more useful than another paragraph declaring the product “premium”. Keep the listing and the underlying product consistent.
The customer: Conversational comparison can make a large catalogue easier to navigate, but it also gives the platform influence over the shortlist and the criteria used. A brand may be evaluated through a summary before its own page is opened. Ask what evidence would help someone choose accurately, and make that evidence available where the decision happens. Marketplace optimisation remains partly a relationship with a gatekeeper.
Paid digital media
Paid search
Paid search connects an expressed need with a commercial offer. AI is increasingly involved in matching that need, producing assets and choosing the destination, as well as bidding.
The work: Google's April 2026 AI Max announcement describes tools for steering messages and matching, alongside expansion into Shopping and Travel. The specialist's job moves towards defining valuable outcomes, providing good product information, inspecting search behaviour and setting boundaries. Announced rollouts still need checking against the actual account.
The marketing: One campaign can cover a wider range of queries and generate more variations in how the offer is presented. That creates reach, but also more room for inappropriate claims or destinations. A software advertiser should check whether the system is reaching buyers who fit its product, rather than generating cheap enquiries from people it cannot serve.
The customer: The advert and landing page may be assembled around a more specific request. The benefit depends on accuracy and continuity: the offer should remain the same when the person arrives. Feed qualified sales and margin information back into decisions where possible. If the platform is rewarded for form fills, it has little reason to care that the sales team spends Friday explaining the product doesn't do that.
Paid social
Paid social combines discovery with attention, using signals about content and behaviour to decide which advert to show. AI already sits inside that delivery process, rather than merely helping someone write the caption.
The work: Meta's account of its 2026 advertising systems describes developments across creative generation and delivery models. Teams need fewer assumptions about manually matching every advert to a tiny audience, and more discipline around concepts, inputs, exclusions and measurement. Platform-reported gains are useful product evidence, but aren't a forecast for your next campaign.
The marketing: More assets can be combined and selected across placements. An outdoor brand might test a durability demonstration, a practical buying explanation and a customer story. Those are meaningfully different ideas. Changing the adjective in thirty captions is a rather less ambitious test of creative strategy.
The customer: Feeds can become better at finding something relevant, while also becoming more commercially persistent. The brand should consider the experience across exposures: repetition, inconsistent offers and whether an automatically altered image misrepresents the product. Evaluate what people actually see, including mobile crops and generated combinations. The spreadsheet may contain approved assets while the delivered advert tells a slightly different story.
Display, programmatic and native advertising
Programmatic buying automates access to advertising inventory. AI can help predict value and interpret context within that process, but automated purchasing itself isn't a new invention.
The work: Teams can use models to evaluate impressions, classify page content, identify suspicious activity and explore performance patterns. The buyer still needs a view of where ads appear, what fees are involved and how success is measured. Outsourcing more decisions makes those questions more important, particularly when the same supplier sells the inventory and explains how well it worked.
The Trade Desk's August 2026 Kokai Zuma release makes the direction concrete: an assistant for campaign setup and troubleshooting, alongside predictive audience and frequency tools. This brings AI into the buyer's decisions as well as the auction. A helpful recommendation still needs checking against the brief and the actual placements.
The marketing: Contextual understanding can support placements beyond simple keyword matching. An illustrative example would be an equipment brand appearing alongside genuinely useful walking advice, with creative appropriate to the topic. Generative systems can adapt the asset, but a relevant placement won't turn a weak proposition into a persuasive one.
The customer: Native formats and personalised creative can make advertising feel closer to surrounding content. Clear commercial identification helps people understand the relationship. There is also a practical limit to relevance: following someone around the internet after they have bought the product is mostly a demonstration of poor coordination. Test reach, frequency and incremental results, rather than treating a plausible audience description as proof of effectiveness.
Affiliate and performance marketing
Affiliate marketing rewards partners for referred outcomes. Performance marketing makes measurable acquisition central. AI changes both the efficiency of that system and the incentives for producing material around it.
The work: Teams can investigate partners, organise product feeds, identify unusual referral patterns and analyse which relationships bring valuable customers. A useful application would compare partner claims with approved product facts and flag inconsistencies for review. Human judgement remains important when deciding whether a publisher has earned an audience's trust.
The marketing: Product comparisons and recommendations become cheaper to produce. That can help a knowledgeable specialist keep detailed guides current, but it can also make thin affiliate pages easier to multiply. The commercial model rewards an attributed action; that isn't automatically the same as creating additional demand. Existing customers searching for a discount code deserve a different analysis from new buyers discovering the product through a trusted review.
There is useful perspective in Awin's 2026 affiliate trends analysis: across its UK, US, German and French markets, roughly 0.2–0.5% of affiliate sales included an LLM click. That measures a visible referral, not every AI-assisted decision. It is a reason to start learning about these partners without declaring the rest of the channel obsolete.
The customer: An assistant may summarise commercial comparisons without making every incentive obvious. Honest methodology, direct testing and clear affiliate disclosure become useful signals for the person assessing the recommendation. I would prioritise partners who provide information buyers couldn't get from the product feed alone. Otherwise we are paying several people to repeat the same specifications with different buttons underneath.
Television, radio and offline advertising
Television advertising
Television can build broad awareness and make a brand feel familiar. AI changes the economics of producing and buying some TV advertising, while the value of reaching an audience remains a separate question.
The work: ITV launched its GenAI Ads Manager in October 2025, using existing business assets to help smaller advertisers produce commercials. Creative support and the required clearance process remain part of the offer. A quick generated version is not the same thing as a finished, approved campaign.
The marketing: Lower production barriers could make television accessible to businesses that previously ruled it out. Streaming environments also support different buying and targeting options from a traditional broadcast spot. Keep those distinctions clear: a personalised streaming campaign doesn't mean every viewer watching linear television sees an individually generated advert.
A useful 2026 counterpoint is Tesco's accessible cooking campaign with Channel 4. Its research highlighted how inaccurate AI captions can muddle cooking instructions; the campaign included British Sign Language and audio description. This wasn't a showcase for generated ads. It showed why understanding the audience still determines what the production needs to achieve.
The customer: The experience might change very little. They may simply see a local business that could not previously afford to appear there. That is a useful example of AI changing participation in a channel without reinventing consumption. Whether the advert is memorable, credible and worth the interruption still matters.
Cheap production doesn't make the audience's time cheap.
Radio, podcasts and audio advertising
Audio advertising reaches people through broadcasters, streaming services and creators. AI affects production and targeting, but a host's relationship with listeners is a different asset from access to a synthetic voice.
The work: Spotify's advertising automation update describes generative production and automated buying tools. These can reduce the effort of preparing scripts and voice assets; availability needs checking for the market and product. Radio teams can also use assisted production to explore versions before committing to a final recording.
Spotify's 2026 Ads Manager and Ad Exchange update adds a practical connection: advertisers can download generated audio for use through a demand-side platform on Spotify Ad Exchange. Production and media buying can join up more easily, rather than the generated recording sitting in yet another tool waiting for someone to export it.
The marketing: A retailer could create approved local variants with different branch details, or test clearer explanations for distinct buying situations. Podcast sponsorship offers another choice: a creator's own explanation and endorsement. Generating a polished audio spot doesn't reproduce the trust that makes a good host-read placement valuable.
The customer: Listeners need an intelligible message at an appropriate pace, particularly when they cannot see a screen. They also need clarity about who is speaking and whether an endorsement is genuine. Repeated synthetic delivery can be as irritating as repeated human delivery. Test pronunciation, context and frequency, and don't confuse the ability to make more versions with a reason to play more adverts.
Print advertising
Print includes magazine advertising, newspapers, brochures, catalogues and direct mail. The physical format remains the same, but the work required to plan and produce it can change substantially.
The work: AI can help adapt copy, explore layouts, prepare images and check production details. Adobe's InDesign AI Assistant documentation describes beta capabilities for working with layouts and checking issues such as overset text and image resolution. That is useful production assistance, with an obvious need to verify the final artwork before sending thousands of copies to print.
The marketing: A furniture retailer could prepare regional catalogue versions or a direct-mail piece organised around relevant product categories. The useful change is making considered variation more affordable. Printing, postage and distribution still have costs, so digital production speed doesn't remove the need to select an audience carefully.
The customer: A well-designed physical piece can offer a focused experience away from a crowded screen. I wouldn't assume people will automatically value it more because AI exists; that is something to test. Useful curation, credible information and good design give it a purpose. A personalised brochure that gets someone's circumstances wrong is simply a more expensive way to be irrelevant.
Out-of-home advertising
Out-of-home combines a message with a physical location. AI can assist creative production, planning and contextual decisions, while the setting remains central to the idea.
The work: In a 2025 campaign described by JCDecaux, French estate agency network Orpi used The mAIker to produce more than 1,250 locally adapted posters. The creative team developed the messages; automation supported the adaptation. This is a concrete production application, without assuming AI conceived the whole campaign or proved its commercial effect.
Adobe's September 2026 account of its Times Square takeover shows another side of that work: adapting assets to enormous, differently shaped screens, with review and approval alongside generation. The interesting bit is getting an idea to survive the actual environment. A nice image in a laptop window is only the beginning.
The marketing: Digital screens can show different executions according to relevant conditions. A retailer could connect approved creative to weather and stock information. Those triggers might use ordinary rules; there is no need to relabel a temperature check as artificial intelligence. AI may contribute to forecasting, contextual interpretation or asset creation within the larger system.
The customer: People encounter outdoor advertising in a real place, often briefly and alongside other people. Legibility, timing and a strong idea remain important. A clever use of location can outperform elaborate personalisation that nobody notices. The physical experience also limits what can be automated: securing a great site and understanding the surrounding environment remain part of the advantage.
Sponsorship and product placement
Sponsorship associates a brand with something people value; product placement puts it inside entertainment or another experience. Both depend on context and credibility, as well as exposure.
The work: Computer vision can help identify brand appearances and measure exposure across large volumes of footage. Nielsen's sponsorship measurement work combines AI with human analysis. This can make evaluation more systematic, but the estimated media value of a visible logo is not a measurement of incremental sales or a complete account of the relationship.
The marketing: Brands can use better evidence to assess partnerships and plan supporting activity. A sports sponsor might connect an appearance to useful fan content, a retail offer or a live experience. Virtual placement and digitally adapted assets also deserve scrutiny, particularly where the execution could imply an endorsement the person involved hasn't given.
Genius Sports and NBC Sports Regional Networks announced an NBA advertising platform in February 2026, combining AI-powered tracking with branded live insights such as shot probabilities. Here the sponsor can sit inside information the fan wants. Whether that earns attention or becomes distracting still depends on the execution.
The customer: Audiences care about the thing being sponsored. The brand earns a place by contributing appropriately, whether that means funding access, improving an event or making something entertaining. AI can help organise the activation; it cannot manufacture the cultural fit. A highly measurable partnership can still feel completely wrong to the people it is supposed to reach.
PR, communications and reputation
Public relations, media relations and corporate communications
PR helps organisations earn attention and explain themselves to people whose trust matters. AI makes preparation easier, while also increasing the amount of material competing for a journalist's time.
The work: Teams can organise background research, prepare briefing documents and adapt approved facts for different audiences. Meltwater's launch of Mira illustrates AI moving into media intelligence workflows. Finding an apparent opportunity is only the beginning; somebody needs to verify relevance and understand the person they intend to approach.
Muck Rack's 2026 State of Journalism survey provides the other side of the inbox: 82% of respondents used AI tools, but 88% said they delete pitches outside their coverage area. The AI figure includes tools such as transcription and grammar assistance. Journalists adopting AI doesn't make them keen to receive more irrelevant pitches written with it.
The marketing: Original findings, credible spokespeople and a useful story remain the basis of earned coverage. A company could use its own data to explain a real change in customer behaviour, with AI assisting analysis and preparation. That is more useful than generating a press release about its commitment to innovation. I suspect journalists have encountered that particular development before.
The customer: People may meet the story through a publisher, social post or AI summary. Clear, consistent facts help it survive those journeys. Corporate communications also need a recognisable source of truth when something goes wrong. The message should be owned by people who can answer questions, rather than becoming an impressive collection of statements nobody is prepared to defend.
Social listening and reputation management
Listening helps a business understand what people are saying and decide when to act. AI expands the volume that can be examined, which makes interpretation and verification more important.
The work: Classification and summarisation can surface themes, unusual spikes and emerging complaints across media and customer channels. A practical use would connect a sudden rise in delivery complaints with operational data before drafting a response. Sentiment scores need checking: sarcasm, local language and a small number of very active accounts can produce a misleading picture.
Big Valley Marketing's account of using Meltwater gives a refreshingly specific example: its team uses Mira to refine Boolean search queries, alongside dashboards for monitoring coverage and conversations. Better listening starts with collecting the right material. A beautifully summarised pile of irrelevant mentions is still a pile of irrelevant mentions.
The marketing: Faster understanding can support a useful response, a corrected product page or an operational fix. It also creates a temptation to publish too quickly. Synthetic images, false endorsements and misleading clips require verification; automatically escalating or responding to everything that looks dramatic could help spread the problem.
The customer: People want to know what happened and what the business is doing about it. A maintained public explanation, consistent support guidance and a reachable person are practical trust assets. AI can help keep the information aligned. It cannot compensate for refusing to address the underlying issue. There is a limit to how much “we value your feedback” can achieve when the parcel is still in a hedge.
Influencer and creator relationships
Creators offer access to an audience and, in the best cases, a relationship of trust. AI can support finding and managing those relationships, while synthetic creators complicate what an endorsement means.
The work: Teams can examine content, audience fit and partnership history, then prepare more relevant briefs. Treat automated authenticity or fraud scores as leads for investigation rather than final verdicts. Creator selection still needs someone who understands the community and can judge whether the partnership makes sense beyond the follower count.
The marketing: A creator can use AI for editing, translation or production while retaining their own perspective. A virtual character offers a different creative format, but shouldn't be presented as someone with genuine personal experience of using a product. Brand control and human credibility are different benefits; gaining one doesn't automatically provide the other.
The customer: People need to recognise the commercial relationship. ASA research published in February 2026 found that recognising influencer advertising remains difficult and that people want clear labelling. Advertising disclosure and disclosure of synthetic content answer different questions. A label explaining how an image was made doesn't necessarily explain who paid for the recommendation.
CRM, lifecycle and retention
Email, lifecycle communications and marketing automation
Lifecycle marketing helps people progress through a relationship: joining, learning, buying, using and returning. AI adds predictive decisions and generated content to workflows that previously relied on more fixed rules.
The work: Teams can move from manually building every branch to defining objectives, available actions and limits. BrazeAI Decisioning Studio describes using reinforcement learning to select actions for individual customers. This is distinct from simply asking a model to write email copy, and it depends on the quality of the feedback the system receives.
At Braze's September 2026 Forge event, its own lifecycle team described using decisioning for webinar invitations: six adjustable elements created 53,475 possible combinations for 60,000 contacts. This is the supplier's own case, but it makes the practical shift clear. The marketer defines the choices; the system learns which combinations to serve.
The marketing: A retailer might decide between a useful care guide, a replenishment reminder and no message at all. The last option deserves more attention. Optimising the next click can produce a busy programme; optimising the relationship requires frequency limits, appropriate offers and outcomes beyond immediate engagement.
The customer: Communications can arrive when they are useful and reflect what someone has already done. They can also become an exhausting stream of supposedly personal messages. Keep preferences easy to express and make suppression work across channels. Someone who has just contacted support about a broken product probably doesn't need an enthusiastic automated invitation to review it. Context is only valuable if the system acts on it sensibly.
Churn prediction and retention
Retention aims to help the right customers continue receiving value. Predicting who may leave is useful only if the business can do something worthwhile with that information.
The work: Models can examine usage, support history and other appropriate signals to estimate churn risk. Teams then need to investigate causes and test responses. A software customer logging in less often might have solved their immediate problem, changed staff or become frustrated. The same signal can represent very different situations.
Braze's Predictive Churn documentation shows how this works in practice: the marketer defines churn, and the model learns patterns from customers who did and didn't meet that definition. It then scores risk. Choosing a meaningful definition comes before modelling; a quiet week isn't necessarily a customer about to leave.
The marketing: Interventions can be more specific: offer training, resolve a technical issue, explain an unused capability or change the plan. Avoid making discounts the automatic answer. Customers who would have stayed anyway may happily accept them, which makes a retention campaign look successful while reducing margin. Hold back a comparable group where appropriate to assess whether the intervention changed behaviour.
The customer: Useful support can make staying easier. A difficult cancellation process can make leaving harder, which is not the same achievement. Assistants may also help people review subscriptions and spot poor value. I would prepare by making the service worth keeping and the terms easy to understand. A churn model is unlikely to fix a product customers no longer need.
Commerce and conversion
Merchandising, recommendations and conversion optimisation
Merchandising helps people find suitable products; conversion optimisation removes obstacles to choosing and buying. AI supports both, but purchase likelihood isn't the only useful objective.
The work: Teams can analyse search behaviour, product relationships and recurring friction, then prioritise tests. A retailer might connect “doesn't fit” return reasons with missing compatibility information. The useful output is a better recommendation or clearer page, rather than a dashboard explaining the problem in unusually confident prose.
KIKO Milano's work with Algolia is a concrete merchandising example: the beauty retailer uses AI-driven re-ranking to adjust product listings using signals including product attributes, sales and customer behaviour. This brings learning into the shop window itself. The team still needs to decide what deserves visibility, especially when a new product has little history to learn from.
The marketing: Recommendations and page content can adapt to a shopping situation. A beginner assembling a first kit needs different guidance from an expert replacing one component. Product data, availability and compatible accessories become the raw material. Measure margin, returns and satisfaction alongside conversion: selling the wrong item efficiently can make the headline metric look excellent.
The customer: Good personalisation reduces effort and preserves choice. Bad personalisation hides relevant options or makes the site difficult to understand. Let people correct assumptions and browse outside the recommendation. Test basic improvements against complex ones; a clear size guide or a reliable delivery date may solve more than an AI stylist. The point is helping someone buy well, with the technology earning its place in that experience.
Conversational commerce and shopping agents
Conversational commerce lets people explore products through questions. An agent goes further when it can take authorised steps, such as adding items to a basket or completing a transaction.
The work: Brands need reliable product information and clear operational rules: price, stock, delivery, returns and who can authorise what. Shopify's March 2026 agentic commerce update describes expanding product discovery across AI channels. It demonstrates infrastructure being built, rather than establishing that most consumers already delegate purchases. Channel eligibility and checkout arrangements vary.
The marketing: A catalogue becomes input to a conversation. “Find a waterproof jacket for walking to work that fits over a blazer” requires evidence about suitability, rather than merely matching the word “jacket”. Merchants can improve that evidence now, while testing which experiences their actual customers use.
The customer: An assistant can reduce comparison work, but people need to understand its preferences, commercial incentives and authority to act. Researching an expensive purchase is different from making it without confirmation. Wider delegation of routine buying looks plausible to me; the pace is uncertain. Prepare for accurate machine-assisted decisions without treating a platform launch as proof that the familiar shopping journey has disappeared.
Sales, acquisition and customer experience
Lead generation, account-based marketing and sales handover
Acquisition connects the business with potential buyers; account-based marketing concentrates effort on particular organisations. AI can reduce preparation work, but knowing someone's job title is a long way from understanding their buying situation.
The work: Teams can research accounts, summarise interactions and prepare more useful handovers. HubSpot's 2026 prospecting updates illustrate this work moving inside the CRM. A good system should help a salesperson understand the evidence behind an opportunity, including unresolved questions, rather than presenting a confident score without an explanation.
The marketing: A business could coordinate useful material around an account's actual needs: integration guidance for the technical team, commercial evidence for the budget holder and migration advice for the person doing the work. The opportunity is a more coherent buying experience, rather than sending every stakeholder a different automated email pretending to be spontaneous.
The customer: Buyers can use their own tools to research suppliers and prepare questions. They may arrive better informed and less willing to repeat basic information. The sales conversation should build on what they already know. Measure qualified opportunities, deal progress and customer fit alongside response rates. More replies aren't especially helpful if most of them are requests to be left alone.
Customer service and the experience after purchase
Service is where marketing promises meet the product and the organisation behind it. AI can help people answer questions and complete tasks, with a clear distinction between assisting staff and replacing the interaction.
The work: A study of 5,172 customer-support workers, published in the Quarterly Journal of Economics in 2025, found that AI assistance increased issues resolved per hour by 15% on average. Gains varied across workers. This was evidence about assistance in one setting, not a guarantee that an autonomous bot improves every service operation.
The marketing: Better access to product knowledge can make responses more consistent and reveal promises that cause confusion. A repeated support question might justify changing a campaign or onboarding sequence. That connection between service and acquisition is often more valuable than treating the chatbot as a separate efficiency project.
The customer: Convenience matters, but so does getting help when the situation is complicated. Ofcom's September 2026 telecoms research found mixed experiences and continuing demand for human contact. Build a useful handover that carries the history forward. Having to explain the problem again is not much of an upgrade on being put on hold.
Events and experiences
Conferences and trade shows
Events create opportunities to learn, meet people and assess suppliers in person. AI can help attendees make better use of that expensive day away from their normal work.
The work: Organisers can prepare communications, categorise sessions and match interests with relevant opportunities. Cvent's May 2026 releases include AI-assisted session and attendee recommendations. These are concrete capabilities; whether they produce better relationships depends on the event, information and participant choices.
The marketing: An event can offer a more useful route through a large programme. A trade-show visitor researching packaging equipment might receive suggestions based on their production requirements, rather than simply being directed towards whoever bought the biggest sponsorship package. Commercial arrangements should remain understandable when they affect recommendations.
The customer: The benefit is spending less time navigating and more time having worthwhile conversations. Preserve room for unexpected discoveries, too. Part of the value of attending is meeting someone a recommendation system wouldn't have predicted. AI can improve the preparation and follow-up, while the actual encounter remains the product. Nobody travels across the country primarily to admire the quality of the agenda algorithm.
Webinars and experiential campaigns
Webinars make expertise accessible at a distance; experiential campaigns invite people to participate. Both benefit when technology helps someone engage with the subject rather than simply registering their attendance.
The work: Teams can organise questions, prepare accessible versions and create follow-up material. ON24's AI Analytics and Content Engine describes tools for turning event material into further content and tailored follow-ups. Treat those as production and distribution capabilities, then measure whether participants found the experience useful.
Its June 2026 webinar benchmarks, covering activity in 2025, report increasing use of AI to turn webinars into additional content. That is evidence of a production habit taking hold on one platform, rather than proof that AI caused better commercial results. The useful opportunity is extending access to a worthwhile session after the live event ends.
The marketing: A webinar can support different next steps for a beginner and someone evaluating a purchase. An experiential installation could let visitors explore product configurations or contribute to a shared creative work. Those are applications to test, with the interaction designed around an idea. Generating a picture with the logo on it is not automatically an experience worth queuing for.
The customer: Participation should provide something they couldn't obtain as easily from a static page. Access to a knowledgeable person, a practical demonstration or a memorable shared moment can justify the time. Make any generated follow-up relevant to what happened. Signing up for one useful session should not accidentally subscribe someone to a small, autonomous publishing company dedicated to their inbox.
Analytics and marketing operations
Analytics, reporting and dashboarding
Analytics helps a business understand performance and decide what to investigate. AI makes it easier to ask questions of data, which increases the importance of knowing what the data actually represents.
The work: Google's Analytics Advisor announcement shows conversational analysis becoming part of familiar reporting tools. Marketers can investigate questions with less manual report preparation. Someone still needs to verify definitions, date ranges, missing data and whether a proposed explanation follows from the evidence. The interface becoming easier doesn't make the underlying measurement correct.
Google's March 2026 practical guidance shows the intended progression: start with a performance question, then drill into channels or steps in a purchase funnel. That is a more useful habit than asking for a monthly summary and accepting whatever explanation arrives first. Treat the assistant as someone helping investigate, not the person signing off the conclusion.
The marketing: Reporting can move closer to decisions. A team might spot that a campaign attracts customers who return unusually often, then change the offer or acquisition criteria. Ask for the supporting breakdown and competing explanations, rather than accepting a polished narrative about why sales changed.
The customer: Better analysis can identify frustrating journeys and inappropriate targeting. But observed behaviour is incomplete: someone may research through an assistant, ask a friend and buy later through a branded search. Avoid interpreting the last visible step as the whole decision. The purpose of a dashboard is to improve judgement, not supply a reassuringly precise story for every number.
Attribution, experimentation and budget allocation
Attribution assigns credit; experimentation tests whether activity caused a change. AI can support analysis and planning, but it doesn't remove that distinction.
The work: Teams can examine more scenarios and organise tests with less preparation. Google's open-source Meridian is one example of investment in marketing mix modelling. This is statistical modelling of aggregate outcomes, rather than an LLM magically reconstructing every customer journey. It needs suitable data, assumptions and informed interpretation.
The September 2026 Meridian update adds agentic help with data checks and model building, while making GeoX generally available for geographical experiments across advertising platforms. Those experiments can inform the wider model. I like the combination: easier analysis, with an explicit route back to testing what the advertising actually changed.
The marketing: Budget decisions can combine experiments, broader models and channel reporting. A retailer might test the incremental effect of advertising in comparable regions, then use the results to inform planning. Automated creative testing also needs discipline: comparing dozens of combinations after the event makes it easy to find a winner by chance.
The customer: AI-assisted research can take place outside the brand's observable journey. That makes perfect individual-level attribution an even less sensible promise. It doesn't make measurement pointless. Decide which uncertainty matters to the business, then choose a method suited to answering it. Sometimes a controlled test and an honest range are more useful than another model allocating 17.4% of the sale to a touchpoint nobody can verify.
Marketing operations and connected workflows
Operations connects the people, information and systems that let marketing function. This is where isolated AI experiments can become reliable working practices, or a collection of mysterious things only one enthusiastic colleague understands.
The work: Teams can connect research, briefing, production and reporting, with agents handling defined steps. Maintain a shared source of product facts, approved claims, brand guidance and examples. Give it an owner and a review process.
More context is useful only when people know which parts are current and trustworthy.
Amplitude's January 2026 account of its internal AI programme describes connected workflows spanning systems including Salesforce, Jira and Slack, supported by repeated staff training. Its own lesson is that execution and education mattered more than perfect tools. The takeaway isn't to copy its number of agents; it is to budget for people learning how the work changes.
The marketing: Consistent inputs make it easier to coordinate campaigns across channels. For example, a changed delivery promise could trigger proposed updates to ads, lifecycle messages and product pages. Keep a record of the changes, test the workflow and allow someone to reverse a mistake. Generating the updates and publishing them are separate responsibilities.
The customer: Coordinated systems should mean fewer contradictory offers and less repetition. They can also distribute an error much faster. Start with a bounded workflow, clear success criteria and permissions appropriate to the consequences. Smaller classification tasks can be useful here. Not every useful automation needs a general-purpose agent wandering around the business with administrative access.
The changes that cut across everything
Cheaper production changes what is worth paying for
AI can reduce the marginal effort of producing another variation, translation or analysis. The total cost still includes preparation, integration, checking, distribution and mistakes.
A first draft arriving instantly doesn't mean the whole job is free.
My expectation is that competent execution becomes easier to buy, while useful attention becomes harder to earn. If every competitor can produce more material, production volume alone offers little protection. Audience access, original information and the ability to choose a worthwhile idea may absorb more of the value.
This changes the agency conversation. Charging for a process that has become dramatically quicker will invite questions. Agencies can build value around judgement, original work, reliable systems and demonstrated business results. That doesn't require pretending every engagement can be priced against revenue; it does require explaining what expertise and responsibility the client is buying.
A March 2026 Harvard Business School explanation of research at IG Group helps here. AI sped up article creation, but people further from the relevant expertise still struggled to match the specialists' writing quality. It is one bounded study, not a universal productivity forecast. It does give clients a sensible reason to pay for knowledge as well as production.
For internal teams, I would measure time saved and what happened to it. Reinvesting it in customer research is different from filling the newly available hours with more output nobody requested.
Small businesses gain capability; large businesses still have advantages
A small team can prototype, produce and adapt work that previously required several suppliers. That is exciting. It can make a focused business more capable without making its marketing department enormous.
Larger organisations may have richer data, established distribution, trusted brands and more resources for integration. They also have complicated systems and more ways for an outdated document to become everybody's source of truth.
ONS research published in July 2026 puts some numbers around the uneven adoption: 28% of UK businesses with 0–9 employees reported using at least one AI technology, compared with 49% of those with 250 or more. These are adoption figures, not evidence of profit. They are also a useful corrective to the impression that absolutely everybody has already transformed their business.
I don't think either side wins automatically. The useful question is which scarce assets the business can combine with the tools. Those might include specialist knowledge, a loyal community, a distinctive product or years of operational evidence.
My bet is that brand becomes more useful when it helps people make a confident choice among similar-looking options. That advantage has to be earned through recognisable communication and a consistent experience. Attaching a familiar logo to generic generated material is a fairly shallow version of the strategy.
Personalisation needs a reason to exist
The progression from mass communication to segments, predictions and individually generated material is technically interesting. It isn't a requirement to make everything different for everyone.
Some things should remain consistent: the price someone was promised, the product's capabilities, the brand's identity and the way complaints are handled. Other things benefit from adaptation, such as showing a relevant integration guide or remembering a stated preference.
Edelman's June 2026 brand research makes a helpful distinction: relevance includes community, identity and emotional connection, not just utility or topical references. I read that as a useful challenge to the software-led version of personalisation. Feeling understood is a bigger ambition than receiving an email with an unusually specific opening sentence.
I would assess personalisation through three questions: does it help the person, can we explain the decision, and does it outperform a simpler approach? Test who gets excluded or repeatedly misclassified, too. Historical customer data can reproduce the business's existing blind spots.
The customer needs ways to correct the system. Otherwise an inference becomes a permanent constraint, and we end up describing a less useful experience as “deeply personalised” because the software bill says so.
Trust needs evidence, with a little practical housekeeping
I don't think “made by a human” will automatically make something valuable, or “made with AI” will automatically make it worthless. Context matters. A useful translated instruction and a fabricated testimonial should not be judged as equivalent uses of the technology.
Show where claims come from. Keep genuine demonstrations and customer evidence. Content Credentials can record information about an asset's origin and edits, but provenance isn't proof that the underlying claim is true. Nor does missing provenance prove something is fake.
There is some unglamorous work here. Check rights to source material, likenesses and voices, plus supplier terms covering training, reuse and generated outputs. Don't assume ownership or commercial permission simply because a tool produced the file. For customer data and communications, the ICO's direct marketing guidance remains a practical starting point for permissions and appropriate use.
Give somebody responsibility for these decisions. It should be possible to answer “where did this come from?” without convening an emergency meeting.
Platforms offer more capability and keep more control
A platform that helps create the advert, selects the audience, runs the auction and reports the outcome provides a convenient service. It also influences most of the decisions we are trying to evaluate.
That doesn't make its tools useless or its reporting automatically wrong. It means the advertiser needs an independent view of success. Use commercial data, experiments and customer evidence to assess whether the activity helps the business.
As more execution moves inside platforms, the team's responsibilities change. Understanding incentives, inspecting outputs and choosing the right objective matter more. So does retaining access to your own information and relationships.
I would be wary of a marketing strategy whose only explanation is “the platform recommended it”. The platform has a perfectly reasonable business model. It just happens to have its own business to run as well.
Amazon's conversational advertising update is a concrete example of that convergence: product discovery, questions and sponsored prompts sit within the same shopping environment. For a brand, making the product understandable is necessary, but so is understanding the commercial system deciding when that explanation gets seen.
Generalists, specialists and the problem of learning the job
Generalists gain the ability to attempt more work. Specialists gain tools for reducing repetitive execution and investigating more difficult problems. Both still need enough knowledge to recognise when the result is wrong.
There is a real tension for junior roles. Some routine work can be automated, but that work has also been how people learn. The customer-support research above found particularly useful gains among less experienced workers; assistance can help people develop and perform. That doesn't tell us how many entry-level marketing jobs will exist in five years.
Anthropic's January 2026 experiment on learning coding skills found weaker immediate understanding among the AI-assisted group. Participants who asked for explanations tended to learn more than those who delegated heavily. This was a coding study, not a marketing apprenticeship trial, but the distinction is useful: completing the task and learning to do it are different outcomes.
My recommendation is to design the learning deliberately. Let junior staff inspect source material, make a judgement, compare it with an assisted result and receive feedback. Give them ownership of bounded problems. Watching a senior person approve generated output all day is unlikely to replace an apprenticeship.
For experienced marketers, the valuable combination is commercial judgement, subject knowledge, technical confidence and the ability to explain decisions. Prompting is useful, but maintaining evidence, defining a good result and checking it are more durable skills than memorising this month's favourite incantation.
The customer can automate their side too
We have spent a lot of time imagining AI helping brands reach people. People can also use it to compare claims, draft complaints, review subscriptions and avoid some of the effort marketing asks of them.
Ofcom's September 2026 research found that 8% of UK online adults had used AI for telecoms-related tasks. That is a bounded example of emerging behaviour, rather than evidence that everyone has handed their spending to an agent.
The wider possibility matters nonetheless. An assistant acting for a customer could question the renewal price, check compatibility or filter out unsuitable offers before a brand ever receives a visit. It might also make the wrong recommendation, so trust in the intermediary becomes part of the decision.
My working assumption is that businesses will need to serve both people and software acting for them. Clear product evidence, dependable operations and straightforward terms help in either situation. There is no need to wait for a precise forecast of autonomous shopping adoption before improving those things.
What I would do with all of this
The fastest practical changes are in work that is digital, repeatable and reasonably easy to check: adaptation, classification, analysis and parts of production. Consumer behaviour is changing too, but unevenly. A new tool becoming available and a market reorganising around it happen on different timescales.
Activities built around physical delivery, relationships and original experience are less directly replaceable, although their preparation can change considerably. A model can help organise a trade show. It still cannot attend on your behalf, shake someone's hand and work out why their demonstration keeps breaking. At least not in the way I mean.
For a marketing leader, I would start with five decisions:
- Choose a customer problem worth solving. Identify whether AI improves the work, the delivered marketing, the customer's experience, or several of these.
- Establish the baseline. Record current cost, effort, quality and business outcomes before announcing a transformation.
- Prepare the inputs. Fix product information, definitions, permissions and access to evidence. Give the important information an owner.
- Run a bounded test. Decide who checks the work, what the system may do and what would make you stop or expand it.
- Change the organisation around what works. Update responsibilities, training and agency relationships; don't leave the useful experiment dependent on one person and an undocumented prompt.
The opportunity is broader than producing the same marketing faster. We can investigate more interesting questions, make previously impractical experiences affordable and remove genuine effort for customers.
We can also produce a spectacular amount of rubbish.
I am optimistic about the first possibility. I would just like us to measure it carefully enough to notice when we have achieved the second.