78% Followed Advice They Later Suspected Was Wrong

Marketers are delegating more decisions to software than their data can reliably support. Validity found 78% of C-suite marketers had acted on a recommendation they later suspected was wrong because the CRM data underneath it was bad, while only 21% said their CRM data was very well prepared for AI.¹

That number captures the awkward state of AI in marketing this week. The software is getting better at deciding who to target, what to show, when to follow up and where to spend, but the inputs remain full of bot clicks, shifting platform definitions, stale CRM records and incentives that belong to the platform rather than the advertiser. More autonomy makes weak signals more expensive.

Delegation outran data hygiene

Validity’s survey is uncomfortable because the failure did not happen at the model layer. Nearly two-thirds of organisations had increased the number of marketing decisions delegated to autonomous agents over the past year, while 62% said poor CRM data quality had directly cost them revenue.¹ The software could make a recommendation quickly, but the customer record underneath that recommendation could still be duplicated, outdated, incomplete or wrong.

That is a different risk from the familiar fear that a model might hallucinate a sentence. A bad customer field can now influence audience selection, budget allocation, timing and follow-up in one chain. Once software is allowed to act rather than advise, a small data error can travel much further before a person sees it.

The product launches are already assuming more autonomy. Auxia says its platform has crossed 200 billion marketing decisions and has launched Agent Studio to let teams build and govern more of those decision flows.² Groweon is positioning its CRM as a system that can choose and execute next actions within business-set boundaries, while Introhive is making relationship intelligence available to assistants through MCP.³⁴

These products can remove a lot of low-value coordination. A marketer should not have to export a list, reconcile it with another system, paste it into a campaign tool and then rebuild the same logic every week. But once that logic becomes executable, data quality stops being back-office housekeeping and becomes part of campaign strategy.

For small teams, this is especially important. An owner using Instagram marketing AI or an Instagram AI content management tool may have far less data than a global brand, but the same principle applies. A clean list of repeat buyers, lapsed customers, frequent bookers or high-value products can be more useful than a much larger pile of poorly maintained records.

Sixteen per cent of clicks are bots

Omnivery’s email research gives the data-quality problem a more concrete shape. It found that bot-generated activity had risen from roughly 2% of B2C email clicks in 2023 to about 16% in 2026, with higher contamination reported for some major inbox providers.⁵ A metric many teams still read as human interest is increasingly mixed with automated behaviour.

That matters because automated marketing systems are often designed to respond to whatever can be measured. If a click triggers lead scoring, retargeting, a nurture sequence or a bid adjustment, then a bot click is no longer a harmless reporting nuisance. It becomes an instruction.

Clicks still contain information, especially when combined with revenue, qualified actions and verified behaviour. Trouble starts when every observable event receives equal confidence, particularly when the system acting on it can move budget or contact customers without waiting for a person. A contaminated signal can look perfectly legitimate once it has been turned into a score or recommendation.

This is where enthusiasm for end-to-end automation needs more discipline. A system that can generate content, segment an audience and optimise distribution can save time, but it also links previously separate errors. One contaminated signal can affect several downstream decisions, and the speed of the workflow makes the mistake look like efficiency until somebody checks the outcome.

For marketers building an Instagram content strategy, the same caution applies at a smaller scale. Engagement metrics can help decide which formats deserve another iteration, but they should not become the brand brief. AI captions for Instagram business posts can be useful when they start from real product information and a known voice; they are much less useful when every high-engagement phrase gets treated as something to repeat.

When the ruler moves

Google’s Merchant Center reporting change on 24 August is a reminder that sometimes the behaviour did not change at all. Google altered how some YouTube affiliate and organic activity is classified, while historical reporting was restated back to 1 July.⁶ A team could open a dashboard, see organic traffic fall or paid activity rise, and build a convincing explanation around a customer shift that never happened.

This is a mundane problem, but it is exactly the sort of thing autonomous optimisation can mishandle. Software is very good at detecting movement. It is not automatically good at knowing whether that movement came from customer behaviour, a tracking bug, a reporting definition, a privacy change or a platform update.

That distinction becomes more important as the platforms remove manual controls. Microsoft is taking away Max CPC ceilings from several new automated bidding campaigns, while Meta is expanding the ability of Meta AI to analyse campaign performance and recommend changes.⁷⁸ StackAdapt research found that 90% of marketers are comfortable receiving AI recommendations, but only 6% say they almost always act on in-platform recommendations, with many dismissing them as generic or irrelevant.⁹

The hesitation is rational. An ad platform has deep visibility into its own auction, but it does not automatically understand margin, stock constraints, customer lifetime value, brand damage or whether the business would rather accept slower growth than acquire the wrong customers. The platform can optimise the route only after the business defines the destination well enough.

This also explains why more reporting is not always more control. A prettier explanation of a campaign can still be built on a metric whose definition changed yesterday. Marketing teams need a habit of recording measurement changes, annotating baselines and keeping a separate view of the commercial outcomes they actually care about.

First-party data earns its keep

The most encouraging story this week came from the other side of the same problem. HSAD reported that an AI targeting approach built from real CRM behaviour delivered 2.1 times the conversion rate of general-target advertising for a health supplement brand, with average purchase value about 18% higher.¹⁰ One lookalike segment built from lapsed customers converted at roughly seven times the general audience rate.

The interesting part is the input. The system used purchase history, buying cycles, coupon response and wish-list behaviour rather than relying on generic audience assumptions. That is first-party information competitors cannot reproduce by typing a better prompt.

For years, marketers could compensate for weak customer knowledge by buying more reach. That gets less attractive when software can generate endless creative variants and media platforms can optimise distribution on everyone’s behalf. The scarce advantage moves closer to what a business uniquely knows about its own customers and products.

The same logic applies to AI content marketing. Product photographs, customer questions, booking patterns, return reasons, product specifications and previous campaign responses create better material for a model than a one-line request for something "engaging". Distinctive input gives the system something distinctive to preserve.

A small fashion brand does not need a huge data warehouse to benefit. It may only need to know which products sell together, which customers return for new drops and which images consistently lead to product-page visits. A salon may know which services create repeat bookings and which before-and-after posts generate enquiries. A restaurant may know which dishes drive weekday visits versus weekend bookings.

That is a better starting point for how to automate Instagram content creation than asking a model to invent a month of posts from almost no context. A business already contains years of useful signals in its photographs, products, customer behaviour and accumulated decisions. Giving AI access to better material is often more valuable than asking it to produce more material.

There is also a useful distinction between adapting and inventing. Resizing an existing campaign, turning one product shoot into several formats or generating variations around an approved message starts from something the business has already chosen to stand behind. Asking a model to manufacture the identity, claim and creative direction at the same time gives it much less to anchor to.

Context is getting closer to purchase

Other launches this week show where that better information may be used. KERV.ai is expanding shoppable connected-TV experiences with LG Ad Solutions, analysing video context so relevant moments can become interactive or commerce-enabled.¹¹ Rokt is using transaction and first-party data to decide which offer or message appears after a purchase.¹² Both are trying to move decisioning closer to the moment when context is most useful.

These systems are trying to read context at the moment a decision might happen. What someone watched, what they bought and where they are can all be useful signals, and AI can process combinations that would be impractical for a human media planner to evaluate manually. That creates opportunity, but it also raises the cost of acting on a signal too aggressively.

Better context does not mean every moment should become an ad slot. The most useful decision a system makes may be to show nothing. If every surface becomes shoppable and every signal triggers a commercial response, relevance turns into interruption surprisingly quickly.

That restraint belongs in the rules before the software acts. Brand teams need to decide which customer states deserve an offer, which deserve information and which deserve silence. The same principle applies to Instagram content generation: a recognisable content strategy is often built as much from what a brand declines to publish as from what it produces.

This is one reason the current race to automate marketing should not be measured by the number of actions a platform can take. Capability is becoming abundant. Judgement about when to act, what evidence is trustworthy and what the customer should experience is still scarce.

The 78% warning

The lesson from that 78% figure is not that marketers should stop using automated recommendations. It is that delegated decisions need better inputs than human-reviewed suggestions because the cost of a mistake can travel further. The closer software gets to budget, customer contact and commercial action, the more valuable provenance becomes.

That means teams need to know where a recommendation came from, which data informed it, which rule permitted it and what business outcome followed. Those questions are not governance theatre. They are the practical checks that separate useful automation from a faster way to repeat a bad assumption.

For many marketing teams, the next useful AI project may look disappointingly unglamorous. Clean the CRM. Separate verified human activity from bot traffic. Annotate platform reporting changes. Define which outcomes matter more than clicks. Write down the brand rules and approval boundaries that software is allowed to act within.

Once those foundations are in place, the benefits get much more interesting. AI can reduce repetitive production, coordinate workflows and turn customer knowledge into faster action while leaving people with the decisions that still need taste, context and accountability. Good automation should remove low-value decision work and leave more human time for the choices that need judgement.

Sources