The technology is available. What companies lack is the translation layer.
A company that has described its work so that both an employee and a language model understand it can choose any tool and any supplier. A company that has not done so buys licences and waits for a result.
Eight in ten respondents say that AI has improved their productivity. 37% see an effect on EBIT.
Employees work faster. The income statement does not show it.
Four figures frame the whole conversation about language models in a company. The first says that people work better. The second, that the company does not see it in the result. The third shows where the difference leaks away. The fourth says where Poland stands in that field.
Eight in ten respondents report that AI has improved their productivity.
McKinsey, The State of AI: Global Survey 2026. A survey of 1719 respondents from 97 countries, conducted from 4 May to 8 June 2026.
37% of respondents attribute any effect on EBIT to AI. That is practically the same as a year earlier.
McKinsey, The State of AI: Global Survey 2026.
In the past month 40% of American office workers received a piece of “finished work” generated by AI that had to be corrected. Correcting one such case took 1 hour and 56 minutes on average.
BetterUp Labs and Stanford Social Media Lab, 1150 office workers in the USA, September 2025. The methodology page of the study gives “2 hours”; the precise figure of 1 h 56 min is reported by Axios.
8.4% of Polish enterprises employing 10 or more people use AI technology. The EU average is 20.0%, and in Denmark it is 42.0%.
Eurostat, data for 2025.
The model does not make the difference. The difference comes from a process described so that the model can carry it out. Among the 6% of respondents whom McKinsey classifies as achieving a real return, that is, attributing at least 5% of EBIT to AI and rating its effect as significant, close to three quarters have thoroughly redesigned their workflows. A year earlier the figure was 55%. A tool on its own does not change the way a company works.
Five reasons why licences do not turn into results
Licences that go unused
The company pays for dozens or hundreds of seats. After a month most employees forget about the tool. The cost is certain, the effect is not.
A tool instead of a process
The employee gets an assistant but not a description of what to use it for. The process stays the same, only with an extra window on the screen.
More content, less meaning
Faster emails and longer documents that nobody reads to the end. Communication speeds up and its value falls. The work comes back to the sender as corrections.
Knowledge in two or three heads
Procedures, discounts, customer history and exceptions exist, but not in a form that a new employee or a language model understands.
No measurement
Success is measured by a training evaluation, not by time recovered in a specific process. The board has no figure to ask about at the next meeting.
Every one of these five reasons comes down to a single thing: between the business goal and the tool, the layer in which the work of the company is described is missing.
What the translation layer is
The translation layer is the layer between the business goal and the tool: process, knowledge, rules, data and decisions described so that both a person and a language model understand them. In most companies this layer exists only in the heads of two or three people and has never been written down.
Brand Semantics works with the language that both people and language models understand. That same language decides two things: whether a model can correctly recommend a company to its customers and whether it can do the work of that company. The first direction is described in the brand semantics method. This page is about the second.
Until the layer exists, every new use of models starts from zero and ends with the person who happens to know how it works. Once it is there, the company stops depending on a single supplier: the process description is independent of the tool, so the model can be replaced without losing the work put into it.
The same layer works in two directions. Inward it decides whether a model can do the work of the company. Outward it decides whether a model can correctly recommend that company to a customer. That is why an organisation that puts one of them in order gets the other more cheaply.
The process translator: a capability that is not yet on the organisation chart
A process translator is an employee of the company, not an IT specialist, who understands a process from the business side and can describe it so that a language model carries it out correctly and repeatably. This is not a new post and not a new department. It is a capability granted to someone who already works in the team.
At Brand Semantics we choose the process translator together with the company, train that person and guide them through the first deployments. When the project ends, the capability stays in the organisation rather than with the adviser.
What the role does day to day
- describes the tasks of its own department in a form a language model understands
- maintains a library of proven work patterns
- assesses the quality of answers before they reach the customer
- flags processes ready for automation
What the company gains
- independence: the capability stays in the organisation after the project ends
- quality: less low-value content, fewer corrections, a consistent standard
- pace: new uses deployed in days, not in quarters
- security: data and permission rules applied in every new use, not only in the first
One company semantics, two directions of model work
Inward
How language models help employees do the work of the company: processes described so that a person and a model understand them, knowledge held in the system instead of in individual heads, automation where the process is already named. That is what this page is about.
Outward
How language models describe and recommend a brand to customers, dealers and buying agents: visibility in generative engines, likelihood of recommendation, share of voice, work with hallucinations. That is the subject of the brand semantics method, advisory services and the research tool Semantio.
Both sides feed on the same company vocabulary: entities, concepts, relations and rules. That is why a company that puts one of them in order gets the other more cheaply.
Where we start: a thirty-day readiness diagnostic
We do not start with a deployment and not with a six-month commitment. We start with a diagnostic that ends in a single board decision, taken on data rather than on declarations.
People
Capability and real use of the tools. We measure them with a knowledge test and with data from the systems, not with a satisfaction survey.
Knowledge
Whether the knowledge of the company is available and decodable, that is, whether a language model can use it to answer real questions from employees and customers correctly. We measure this dimension with the Brand Semantics Readiness index.
Processes
Whether a process can be described and measured. Every candidate gets a card with an owner, a metric and a list of exceptions, together with a rating on three axes: value, feasibility, risk.
Each dimension gets its own score and its own list of causes, instead of one aggregate figure. This lets the board see where to start: with people, with putting knowledge in order or with a single process. We always assess the answer quality of a particular model in the context of a particular knowledge source and a particular process, never in isolation.
What stays in the company after thirty days
- a readiness scorecard across three dimensions, together with the causes and a reference point for later measurements
- a list of repeatable processes ranked by value, feasibility and risk
- a map of information friction: where work stalls for lack of data, rules or decisions, and who unblocks it
- a specification of one process ready for a pilot, together with a deployment brief for the IT department
- metrics and a method of measuring before and after, based on data from the company systems
What we need from the company
- a decision-maker on the board
- a process owner in every department covered
- access to tasks, documents and data within an agreed scope
- from twenty to thirty employees available for two workshops and interviews
The diagnostic ends with one of three recommendations: deploy, put things in order first, or stop. We launch later stages only where the diagnostic has shown a process, an owner and a measurable effect. If they are not there, we say so plainly.
What we do, what stays with the company
We are the architect of the method and the facilitator of the change. The company remains the owner of the process and the owner of the result. Technical integration is carried out by its IT department or by a partner it nominates, on the basis of our process description. We do not sell licences, we are not a reseller for any model supplier and we do not maintain infrastructure belonging to others.
| role | who | responsible for |
|---|---|---|
| decision-maker | a member of the company board | the goal, priorities, the decision to scale and the use of recovered time |
| process owner | a director or head of department | the course of the process, the input data, sign-off on the pilot result |
| architect and facilitator | Brand Semantics | method, diagnostic, process description, model selection, brief for IT, measurement |
| technical implementer | the company IT or a nominated integrator | integration, permissions, environment, maintenance |
| approving functions | security, legal, data protection | approval of the scope of data and uses before every pilot |
decision-maker
who
a member of the company board
responsible for
the goal, priorities, the decision to scale and the use of recovered time
process owner
who
a director or head of department
responsible for
the course of the process, the input data, sign-off on the pilot result
architect and facilitator
who
Brand Semantics
responsible for
method, diagnostic, process description, model selection, brief for IT, measurement
technical implementer
who
the company IT or a nominated integrator
responsible for
integration, permissions, environment, maintenance
approving functions
who
security, legal, data protection
responsible for
approval of the scope of data and uses before every pilot
We work with models from various suppliers: cloud, open and locally run. The choice depends on the type of data, the cost per task, the security policy of the company and the language, never on a partnership agreement on our side.
We help to shape standards for the responsible adoption of AI at the AI Business Center Foundation, which works on business education, certification and compliance with the AI Act. The Foundation's president is Grzegorz Miłkowski, CEO of Brand Semantics, who on behalf of the AI Business Center works with company boards on setting the strategic direction of AI transformation in business.
The rules that already apply, and those reported as coming into force later
In July 2026 the news went round the media that the Union is postponing the rules on artificial intelligence. That is true, but it concerns a different part of the regulation from the part that reaches the marketing department.
The AI Act, that is Regulation (EU) 2024/1689, was amended by Regulation (EU) 2026/1744, known as the Digital Omnibus on AI and referred to below as the AI Omnibus, which entered into force on 27 July 2026. The Omnibus postponed the obligations for high-risk systems: for standalone systems under Annex III to 2 December 2027, and for systems embedded in regulated products under Annex I to 2 August 2028. Those are the postponements that were written about, and that is the reason why many companies wrongly concluded that the subject does not concern them.
The transparency obligations under Article 50 have not been postponed. They apply from 2 August 2026, that is, already.
High-risk systems are, for example, AI used in recruitment or in credit scoring, critical infrastructure, biometrics. Marketing and brand communication, by contrast, usually do not fall into that category. Transparency is a separate layer, which concerns entirely different activities of a company.
What obligations does Article 50 place on a company that uses AI in communication?
Three things, simplified and without settling borderline cases.
- 1First, a person is to know that they are talking to a machine. If an AI assistant or an AI chatbot operates on a website, within a mobile application or on a telephone call, the user must be informed of this at first contact, unless it is obvious from the circumstances.
- 2Second, content generated or modified by AI, if it meets the criteria of a so-called deepfake, is to be recognisable as such. This is precisely the transparency obligation that every marketing department should take an interest in from 2 August 2026. The transparency obligation does not cover every published text, nor every graphic or video item. As far as texts are concerned, the labelling obligation covers material informing the public about matters of public interest, so most marketing texts do not have to be labelled. Visual, audio and audio-visual content, on the other hand, if it was created using AI and has the potential to mislead the recipient as to whether the characters, objects, places and events shown in the material are real, must be clearly labelled.
- 3Third, if a company uses emotion recognition systems or biometric categorisation systems, the persons subject to them must be informed of this.
It is worth saying plainly what Article 50 does not require: it does not order the labelling of every sentence written with the help of a model, nor of every graphic or video created or enhanced with AI. A great deal of material in circulation suggests the opposite. A company that labels everything will do itself no legal harm. It will do itself harm in communication.
Employee capability: the rule that changed in the other direction
Article 4, on AI literacy, has applied since 2 February 2025. The AI Omnibus softened it: providers and deployers of AI systems are to support the development of AI literacy among their staff, instead of ensuring a defined level of it.
Softening the obligation does not, however, change what those capabilities are for. The rule describes a minimum, not an ambition.
Timetable for the application of the AI Act
2 February 2025
Prohibited practices (Article 5) and AI literacy (Article 4).
2 August 2025
General-purpose models, supervision and penalties.
2 August 2026
The AI Act applies as a general rule, including the transparency obligations under Article 50.
2 December 2026
End of the transitional period for machine-readable marking of content. AI providers will have to mark every visual item that came out of their tool in a machine-readable format. A new prohibition on creating intimate material with AI without consent also starts to apply.
2 August 2027
Providers of general-purpose AI models placed on the market before 2 August 2025 should bring themselves into line with the AI Act.
2 December 2027
Obligations for providers and deployers of high-risk systems under Annex III.
2 August 2028
Obligations for providers and deployers of high-risk systems under Annex I.
Legal position as at 13 September 2026. The above is information, not legal advice.
Text of the amending regulation: Official Journal of the EU, Regulation (EU) 2026/1744.
What we think about it
This is not a reason to stop training. It is a reason to stop buying training out of fear of a penalty and to start buying it for the result. Marketing material that still threatens sanctions for the lack of a certificate under Article 4 describes the legal position before 27 July 2026.
Five risks that a project like this has to get around
| risk | how we get around it |
|---|---|
| More content, less meaning | a communication standard and verification before sending; we measure corrections, not the number of pages generated |
| Resistance: “faster means more work” | a narrative of recovered time, the participation of managers in the workshops, HR as the owner of the change |
| Data and permissions | an access audit before any agent is launched, the principle of minimum permissions, sensitive decisions always with a person |
| Dependence on a single supplier | a process description independent of the tool, a model chosen to fit the task, the option to switch without losing work |
| Regulatory compliance | documented programmes and measurement, scope agreed with the legal adviser of the company |
More content, less meaning
how we get around it
a communication standard and verification before sending; we measure corrections, not the number of pages generated
Resistance: “faster means more work”
how we get around it
a narrative of recovered time, the participation of managers in the workshops, HR as the owner of the change
Data and permissions
how we get around it
an access audit before any agent is launched, the principle of minimum permissions, sensitive decisions always with a person
Dependence on a single supplier
how we get around it
a process description independent of the tool, a model chosen to fit the task, the option to switch without losing work
Regulatory compliance
how we get around it
documented programmes and measurement, scope agreed with the legal adviser of the company
Who this work is for
We work with companies that
- have bought licences for GenAI tools and cannot say what they get from them
- have several locations or several departments that do not speak a common language
- have a person on the board ready to take responsibility for this change
- want to build the capability in-house rather than buy dependence on an adviser
This is not work for companies that
- are looking for a turnkey deployment without involving their own people
- want a one-off training course on how to use a tool and then close the subject
- expect a declaration of compliance with the AI Act instead of a real programme
In our capability programmes every participant finishes with a personal certificate carrying a token that confirms its validity and authenticity. The company finishes with a documented scope and with measurement before and after. That is the only way to know after a year whether it worked.
How the companies that work with us begin
Readiness diagnostic
Thirty days, a readiness scorecard across three dimensions and a decision: deploy, put things in order or stop.
In-house programme for a company
A scope designed for the specific organisation, after a needs diagnostic and surveys among the participants. From brand visibility in models, through GenAI tools in the daily work of departments, to agents and the capability transformation of teams.
Open training courses and webinars
One day, one industry, your own brand on the workbench. A good way in for someone who wants to see the method live first.
Experts
The programmes are led by Michał Grzebyk, Katarzyna Bienussa and Grzegorz Miłkowski.
Frequently asked questions
It requires neither training nor a certificate. Following the change introduced by Regulation (EU) 2026/1744, Article 4 of the AI Act requires providers and deployers of AI systems to support the development of AI literacy among their staff, instead of ensuring a defined level of it. This is an obligation of diligent action, not an obligation of result. Material that threatens penalties for the lack of a certificate describes the wording of the rule before 27 July 2026.
Not all of it. The transparency obligation under Article 50 of the AI Act, applying from 2 August 2026, does not cover every published text, nor every graphic or video item. In the case of texts, the labelling obligation covers material informing the public about matters of public interest, so most marketing texts do not have to be labelled. Visual, audio and audio-visual content that was created using AI and has the potential to mislead the recipient as to whether the characters, objects, places and events shown in the material are real must, on the other hand, be clearly labelled.
Because the tool lands in an unchanged process. According to the McKinsey survey, eight in ten respondents report that AI has improved their productivity, but 37% attribute an effect on EBIT to it, practically the same as a year earlier. Among the 6% of respondents achieving a real return, close to three quarters have thoroughly redesigned their workflows. Minutes saved at a desk do not turn into money on their own.
It is putting in order the way a company describes its work, so that both people and language models can carry it out, and building in the organisation the capability that maintains it. It covers the readiness diagnostic, capability programmes on real tasks, process descriptions, a standard for working with models and measurement on data from the company systems. It is not a training course on how to use a tool.
A process translator is an employee of the company, not an IT specialist, who understands a process from the business side and can describe it so that a language model carries it out correctly and repeatably. The role describes the tasks of its own department, maintains a library of proven work patterns, assesses the quality of answers before they are sent to the customer and flags processes ready for automation.
The translation layer is the layer between the business goal and the tool: process, knowledge, rules, data and decisions described so that both a person and a language model understand them. In most companies it exists only in the heads of two or three people. A company that has it can choose any model and any supplier without losing the work done so far.
With one process that can be described and measured, not with buying licences for everyone. The order that works: choice of a process with an owner, measurement of the state before, a description of the process in a form understandable to a person and to a model, a pilot with measurement after, a decision on scaling. Buying the tool is the fourth step, not the first.
In our format, thirty days. It ends with a readiness scorecard across three dimensions, that is, the dimensions of people, knowledge and processes, a list of processes ranked by value, feasibility and risk, a specification of one process ready for a pilot and one recommendation: deploy, put things in order first or stop.
No. We are the architect of the method and the facilitator of the change. Technical integration is carried out by the IT department of the company or by an integrator it nominates, on the basis of our deployment brief. We do not sell licences and we are not a reseller for any model supplier, so the choice of model follows from the type of data, the cost per task and the security policy, and not from a partnership agreement of ours.
A tool training course teaches operation. Capability transformation changes how the work of the company is described and leaves in the organisation a person who maintains that description. The effect of a training course is measured by the evaluation of the course. The effect of a transformation is measured by time in the process, the number of corrections and throughput, on data from the company systems, before and after.
A decision-maker on the board, responsible for the goal and for the decision on scaling, and a process owner in every department covered. Without those two roles we do not start, because without them the diagnostic ends in a report rather than in a decision.
Let us talk about a diagnostic for your company
Thirty days, a readiness scorecard and a decision taken on data from your systems. A good starting point is one process in the sales or the customer service department.