AI in Marketing: A Cognitive Processor, Not a Gadget

We use the term cognitive processor for an AI system that helps a company analyse information, compare options and prepare decisions. Used only for an isolated task, the same model may impress in a demo without improving how the company works. The difference lies in the information and rules it receives.

TL;DR. In McKinsey’s survey, 44% of respondents say their organisations have scaled AI across the enterprise, while 37% attribute a positive effect on operating profit to AI. These are different measures: subtracting one from the other does not show how many projects failed to recoup their investment. An MIT study of integrated pilots reinforces the need to measure value separately from use.

One cause arises before the choice of model. When company knowledge is undocumented, AI is asked to perform tasks without the necessary information. The model needs access to that context before it can use it.

What does “cognitive processor” mean and how does it differ from a tool?

A tool used in isolation solves the current request. If you ask for advertising copy without preserving the context and corrections, the next task starts with the same explanations. What you learned has not been organised for reuse.

A cognitive processor uses the company’s relevant, accessible context: information about clients, earlier decisions and working rules. Its output is checked against that context. The practical difference can be tested when a second person requests the same task and both results are assessed against the same criteria. A subscription alone does not ensure that consistency.

This distinction helps set priorities. A subscription can solve a one-off task; work repeated across a team may also justify investment in documenting knowledge, integration and verification. The balance depends on how the company works.

Ethan Mollick, professor at Wharton, describes a practice in which he chooses the task, evaluates the AI’s output and adjusts his approach when needed. Choosing the task and evaluating the result remain human responsibilities, and both require context.

In our methodology this principle carries a name of its own, AI-ca-procesor-extensie. Its statement places AI in the role of a cognitive processor, or a cognitive extension, deliberately away from the brain metaphor, and it continues with a line we keep repeating in delivery work. Quality depends on how precisely you tell it where to think.

Why do companies buy a gadget when they need a processor?

An immediate production need can drive the purchase before the company examines its working practices. Consider a hypothetical team that needs thirty social posts, two landing pages and a newsletter by Friday. A tool promising those outputs is easy to approve, while documenting what the company knows about its customers may appear less urgent.

The pillar that covers this in our framework is called infrastructura-inainte-de-marketing, and it states that the client rarely has a marketing problem. What the client lacks is infrastructure, meaning systems, channels, positioning and a brand foundation. Its conclusion is short. Marketing built on a poor foundation burns money.

An AI subscription can accelerate work built on that weak foundation. A company unable to explain why buyers should choose its products or services may now produce forty pieces of content with the same problem. This illustrative increase in output does not, on its own, clarify what makes the company worth choosing.

Documenting company knowledge can also expose unresolved choices. A founder drafting a service’s positioning may discover that three colleagues describe it in three different ways. Resolving that disagreement requires more attention than buying a licence, so choosing a tool can feel like easier progress.

What does the data show about companies that have already tried?

It shows why the spread of AI use and its economic effects need separate measures.

The McKinsey report published in 2026, built on 1,719 respondents from 97 countries, shows that nearly nine in ten organisations use AI regularly in at least one business function, and 44% have taken it to enterprise scale, up from 38% a year earlier. In the same survey, 37% of respondents attribute a positive effect on operating profit to AI, practically unchanged from the year before. These indicators measure different things. The seven-point difference cannot tell us where investment was lost; each company must check what economic result it can attribute to AI use.

The MIT study published in 2025, based on interviews and a review of public initiatives, estimates that about 5% of the integrated pilots it examined produced measurable value. Its authors caution that the sample and varying definitions of success limit generalisation. That estimate does not mean 95% of all companies using AI gained nothing. Separately, Gartner forecast in 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing data quality, costs and unclear business value. This was a forecast, not a later count of abandoned projects.

Gartner also noted the difficulty organisations faced in demonstrating value from their investments.

BCG offers a rule of thumb for allocating effort in an AI transformation in a guide published in December 2024: 10% of the effort goes into algorithms, 20% into technology and data, and 70% into people and processes. It is guidance on effort, not a measured breakdown of spending or failed projects.

What can cause an AI project to fail inside a company?

In our analysis, three organisational problems deserve attention before the model is changed.

The first is the absence of written context. The company asks the model to produce something without ever having told it who the client is, what has been promised so far and what it is not allowed to say. The model fills the gaps plausibly, and the result is a grammatically correct text that does not resemble the company.

The second is the absence of verification. A generated result can look ready to use even when it contains errors. Without checks tested on bad cases too, mistakes can reach deliverables and damage client trust.

The third is choosing the wrong task. What gets automated is what is easy to automate, meaning content production at volume, instead of what hurts, meaning the time lost rebuilding context at every piece of work. The result is more material, with the same lack of clarity, produced faster.

The common pattern is that all three concern how the company works. BCG’s rule of thumb emphasises people and processes; it does not prove that the choice of model is irrelevant.

Where does AI sit in a company’s architecture?

Our methodology analyses a company through four layers. L1 is performance marketing: campaigns, channels and their measured results. L2 covers revenue, the sales process, pipeline and commercial structure. L3 stores context, applies procedures and coordinates work through the AI Brain. L4 defines positioning, what the company stands for and the category it wants to own.

The AI Brain sits at L3 and supports both execution in L1 and positioning decisions in L4. It can prepare information and options, while people remain responsible for choosing the company’s direction.

The part that is hardest to see from outside is how feedback from daily work becomes part of the procedures. People identify recurring rules, check them and write them down for reuse. This is not the same as training the model on company files. A correction that is relevant beyond one task can become a documented rule; the procedure can then bring it into context for similar work.

Keeping that feedback makes it available for later work. A subscription alone does not establish this process: the company must decide how corrections are recorded, organised and retrieved.

Which tasks are worth giving to AI and which are not?

The rule we apply is that AI receives the processing work and the person keeps the judgement work. The difference between them is tested with a single question, whether the result can be verified by somebody who did not take part in producing it.

A conversation summary can be checked against the transcript, making it suitable for delegation with verification. Positioning involves choosing among strategic directions and accepting their consequences, so responsibility stays with the people running the company.

This split helps explain a possible disappointment with high-volume content production. A team may speed up drafting while still spending time rebuilding context for each task.

Legal obligations also depend on the use of the system and the company’s role. Documentation helps but does not replace checking the applicable requirements. We cover them separately in the article on the EU AI Act.

How do you measure whether it works?

Track the time taken to produce a result a person accepts without rewriting, comparing the same type of work before and after implementation. Output volume alone does not tell you whether the work has become more useful.

The second measure is the rate at which results are rejected during verification. A decline may indicate better adherence to company rules, but may also mean the check has become too permissive. Track the rate alongside tests on known good and bad examples; a constant rate alone does not establish that the check has stopped working.

The third measure is consistency. Two different people ask for the same piece of work on two different days, and the results are compared. If they differ substantially, check the available context, the requests, the model and its settings. The difference between outputs does not establish the cause on its own.

What happens to a company the AI engines cannot see?

It may be absent from an assistant’s recommendation even when its products or services remain strong. That absence can be hard to detect through ordinary marketing reports.

In an audit on 10 September 2026, we checked a medical clinic’s public profile, site files and answers from AI assistants. The profile showed 2,970 Google reviews and a 4.85 rating at the time of the check.

Its robots.txt blocked GPTBot, ClaudeBot and Google-Extended. The clinic was absent from the answers we tested, but those directives alone do not establish why. Google-Extended, for example, does not control inclusion in Google AI Search.

The proposed next step was to build pages answering patients’ questions and check whether the relevant engines can access them. That would provide material an assistant could cite, without guaranteeing inclusion.

An industrial cable manufacturer offered a different example. In one test, an assistant placed it first in an answer and cited an older technical article among its sources. Its site appeared around tenth in conventional search for the term we measured. The citation alone cannot explain the order of the assistant’s answer.

That measurement is ours. We ran it on 2 September 2026 through DataForSEO, asking the question under Romanian language and country settings, and the classic search position comes from the same round, on the central term of the category.

The strategic move there was to shift the unit of value from the product to the economic outcome being protected, which in that company’s language became a single line. We do not sell cable. We sell the thing that is not allowed to fail.

Both cases show why the information available about a company matters. Documenting its knowledge and making relevant pages accessible gives an AI engine material it can use, without guaranteeing inclusion in every answer.

What is different in a market that started later?

The starting point is lower, and that is a risk and a window at the same time. Eurostat data for 2025 puts AI use among European companies with at least ten employees at an average of 20.0%, while the member states at the bottom of the range sit around 5%, meaning a quarter of the European level.

Eurostat reports that 34.7% of EU enterprises using AI apply it in marketing or sales. Separately, McKinsey describes a 16-percentage-point gap between Central and Western Europe in scaling AI across organisations. Its estimate of potential economic value ranges from €280 billion to €700 billion, depending on the scenario; it is not a revenue forecast.

These figures suggest room to differentiate through useful analysis and consistent work, beyond increasing content volume. A mid-sized company documenting its knowledge now may accumulate an advantage over two years that buying a subscription alone cannot reproduce. This is our hypothesis, based on keeping each decision available for later work.

What does AI used as a processor look like in practice?

One example is our own system. Selected working information about clients, delivery work and decisions is kept in versioned files alongside procedures. At the start of a task, the system can load the relevant client documents within the access granted. People add context when information is missing or has changed.

The system handles research, client profiles, conversation analysis, drafts and verification. People retain responsibility for decisions and client relationships because they bear the consequences of each choice, even when the model prepares useful options.

I described the layered structure in the article on the AI Brain a founder actually needs, and how it connects to a marketing strategy in the article on marketing strategy built with AI. The way the stack fits together is laid out in the piece on marketing architecture.

The effect a founder seeks is consistency. Two people requesting the same task should be able to use the same procedure and verification criteria, while the reason behind a decision stays documented alongside it. The files remain when you change the model. The new model’s behaviour still needs testing on the same work before it is used in client projects.

Where does a company that has done nothing so far start?

With three moves, in this order, and none of them requires a software budget.

The first move is the knowledge inventory. Check what the company knows and where the information sits: in documents, inboxes or with the people doing the work. If relevant material is inaccessible for the chosen task, the model cannot use it; it can still help organise material it is given.

The second step is choosing one recurring task. Start with a frequently produced deliverable that takes substantial time, before trying to automate a whole department. Document the full procedure and test it by asking a colleague to follow it. The procedure works when that colleague produces a result the original author accepts.

The third move is the verification gate, calibrated on bad cases too. Run a weak deliverable through the check. If it passes, the criteria need adjusting, however convincing the report looks. Testing only good examples can leave this weakness unnoticed.

You can use a model during the inventory or while drafting the procedure, with human verification. These three moves then give you a basis for comparing results and time spent.

Frequently asked questions

Where do I start if I have nothing written down?

Start by inventorying sources and selecting information relevant to one specific task. In the system described here, we organise checked material in documents and grant access only to what that task requires.

How much of marketing can AI do?

The processing part, meaning research, structuring, drafts and mechanical checks. The decision about what you say and to whom, meaning positioning, stays human work, because it requires choosing what to leave aside.

Do we risk sounding like everybody else?

Yes, if you use the model without company-specific information, the result can remain generic. Documenting your voice, examples and rules reduces that risk when the model receives them for each task. The text still needs to be reviewed.

How do I know whether an AI project is worth continuing?

By asking whether the time to a result a person accepts has fallen, not by the volume produced. A system that produces ten times more material somebody has to rewrite anyway has moved the work rather than reduced it.

Do we need a new department?

Not at the size we are discussing. What is needed is one person who knows the operation and has allocated time to write the procedures, plus a clear decision about what the system is not allowed to touch.

Which model should we choose?

Define the task, context and verification criteria first, then compare models on that work. Keeping knowledge and procedures in reusable files makes switching easier, but differences in model quality, speed and cost still need to be tested.

How long before any of this shows a result?

The first useful check comes when another person follows the written procedure and produces a result accepted against the agreed criteria. Timing depends on the task, the available documentation and the corrections needed. Economic effects need separate tracking through time saved, costs and output quality.

If you want to see what AI would process in your company’s case, meaning what knowledge already exists and what is missing for it to be usable, you can request a positioning analysis. It shows you how you look from the outside now, and from there it becomes clear whether your problem calls for tools or calls for clarity first.