ChatGPT for business: what you delegate and what you don't

ChatGPT for business means using OpenAI’s conversational assistant for research, writing, analysis and preparing decisions. For a founder whose team already uses it, the practical question is which tasks to delegate and which decisions to keep with people, whether or not the company has a written policy.

TL;DR. You can delegate tasks to ChatGPT when the output can be checked against a source and mistakes can be corrected before they cause harm. A person remains responsible for the decision and its consequences, because the model cannot take on that responsibility.

Two documented cases show why this boundary matters. In 2024 a Canadian tribunal ordered an airline to pay for what its chatbot had promised a customer. An experiment run by Harvard and Boston Consulting Group found that using AI reduced the likelihood of a correct answer on a task that required judgment. We will look at what happened in each case and what it means for a company.

I wrote separately about the role AI should play in the architecture of a company, in the piece on AI used as a cognitive processor. This article deals with the line itself, task by task, and with the way we apply it in the system we work in every day.

At the end you will find the delegation sheet, a one-hour exercise you can run on your own company’s list of recurring tasks.

What does ChatGPT for business actually mean?

In practice, employees may start using ChatGPT on their own to write, find information and ask for advice. The company can then buy a team subscription with admin controls and integrate the assistant into selected workflows. Employees can start without management approval, but choosing a shared workspace and changing company processes require an explicit decision.

A public study of how people use ChatGPT was published through the NBER in September 2025 by economists from OpenAI and Harvard. The paper, “How People Use ChatGPT”, reports that by July 2025 the product had been adopted by around 10% of the world’s adult population.

The same 2025 NBER study found that non-work messages grew from 53% to more than 70% of all usage. Practical guidance, seeking information and writing together account for nearly 80% of conversations, and writing dominates work-related tasks. The authors conclude that the economic value comes largely through decision support.

That finding changes how you should look at the tool. People already use ChatGPT as an adviser for decisions, and writing is simply the most visible part. That is where the risk sits for a company, because a wrong piece of advice, received with confidence, turns into a decision without passing through anyone who answers for it.

The business version also has an official name. OpenAI’s help page on ChatGPT Business, read on 6 October 2026, states that ChatGPT Team was renamed ChatGPT Business on 29 August 2025. A standard seat cost 25 dollars per user per month at that point, or 20 dollars on annual billing.

Official adoption is still uneven across Europe. According to Eurostat’s December 2025 release, 20% of EU enterprises with at least 10 employees used AI in 2025, from 42% in Denmark down to 5.2% in Romania. The figure only measures what companies report about themselves.

The gap between reported and real use shows up in the 2024 Work Trend Index from Microsoft and LinkedIn. Among people who use AI at work, 78% bring their own tools, and in small and medium-sized companies the share rises to 80%.

Without a shared policy, each employee decides what to delegate to the assistant. The company ends up working under choices that leadership has neither discussed nor assigned responsibility for.

Can I use ChatGPT for my business?

Yes, and you almost certainly already do, at least through your team. The question that changes the result is what kind of work it receives.

ChatGPT is useful for processing information you provide into a summary, an analysis or a draft. Decisions also require judgment about what to choose, what to leave out and who is responsible if the choice turns out to be wrong.

In our methodology, AI as a cognitive processor and extension describes this role in supporting your thinking. The quality of the output depends on the information and instructions the model receives, so give it the context for the task, refine the output and check its claims against sources.

Incomplete information leaves room for invented details, which the model can present just as confidently as verified facts.

A founder who asks for “a proposal for client X” without handing over past proposals, margins and the history of the relationship receives a well-written text about a company that is not theirs.

Before delegating a task, we establish how we will check the result. The fact that ChatGPT can produce an answer does not tell us whether we can use it. Sending it to a customer without checking means accepting responsibility for the mistakes we have not spotted.

Where is the line between processing and decision?

One way to distinguish processing from decision-making is to ask how the output can be checked. A summary can be checked against the document, a translation against the original and a competitor list against the companies’ websites. A price calculation can be checked against the data, but the effects of choosing that price become clear once the offer reaches the market.

To keep the line from depending on instinct, we test it with three questions, asked in order, before a task goes to an AI model. Every question answered with “no” moves the task closer to a person.

The final question makes you name the person responsible. Having a model write the text does not remove responsibility for using it. If you cannot name the person who will check and approve it, the task is not ready for AI.

This is what “human in the loop” means in practice, a name next to the task rather than a vague promise of oversight.

Experimental data shows why these questions matter more than the quality of the model. The study “Navigating the Jagged Technological Frontier”, led by Fabrizio Dell’Acqua at Harvard and published in Organization Science in 2026, had 758 BCG consultants work with and without GPT-4.

On 18 tasks inside the frontier of AI capabilities, consultants using it completed 12.2% more tasks and worked 25.1% faster, with significantly higher quality, according to the same 2026 Dell’Acqua study.

On a management task selected to sit outside that frontier, the same consultants were 19 percentage points less likely to reach the correct solution.

The negative result explains why the authors describe a jagged frontier. Tasks AI handles well can look similar to tasks it gets wrong, so apparent simplicity is not a sufficient basis for delegation. A task that requires judgment needs checking even when the answer looks convincing.

What can you delegate to ChatGPT, task by task?

You can delegate to ChatGPT any task that passes the first two questions, with a named person who checks the output. The table below applies the test to the tasks we see most often in a mid-sized company, from the back office to sales.

The first five tasks produce outputs that another person can check before use. The last four also have consequences that editing a document will not resolve. An email sent to a customer, a price already published or a dismissal may require explanations, negotiations or other action, even if the text is later corrected.

That is why drafting a proposal can be delegated even though deciding on the offer cannot. The model writes from your pricing rules, and the person who approves the proposal checks every figure before sending it. AI saves preparation time while the person retains the decision.

The sequence matters. Someone using AI in a hurry may send the output as soon as it arrives. In a process with proper checks, they compare it with the source, approve it and take responsibility before sending.

What should you never delegate to an AI model?

Keep human approval for decisions that commit the company and whose consequences cannot easily be undone. These include pricing, positioning, communications to real people, publishing, deletions and decisions about people. The model can prepare the supporting material, but the choice stays with someone responsible for it.

In Moffatt v. Air Canada, decided in 2024 by British Columbia’s Civil Resolution Tribunal, the chatbot on the airline’s website told a customer he could claim a bereavement fare after travelling, even though the airline’s written policy said otherwise.

Before the tribunal, Air Canada argued that the chatbot was a separate legal entity responsible for its own actions. The tribunal’s decision rejected the argument in plain terms, and the airline was ordered to pay a total of 812.02 Canadian dollars.

For a company using a chatbot on its website, the case shows that automatically generated replies do not remove its responsibility for the information given to customers.

European law also sets transparency obligations, which I covered separately in the piece on the EU AI Act rules for AI-generated content.

For text, Article 50(4) of the AI Act concerns content generated or manipulated with AI and published to inform the public on matters of public interest. The obligation to disclose AI use does not apply when the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for its publication. This exception concerns that category of text and that disclosure obligation.

The right-hand column covers tasks where mistakes can affect the company’s finances, its reputation or the people it works with. Correcting the text does not automatically undo those effects.

Why is ChatGPT wrong precisely when it sounds most sure?

ChatGPT is confidently wrong because language models are trained and evaluated in a way that rewards guessing. OpenAI explained the mechanism itself in “Why language models hallucinate”, published on 5 September 2025, where it defines hallucinations as plausible but false statements generated by language models.

OpenAI illustrates the problem with a multiple-choice test. A model that guesses an answer has some chance of being right, while an admission of uncertainty receives no points under that scoring system. Across thousands of questions, a model that guesses can rank higher than one that admits uncertainty, so evaluations can reward confident answers over caution.

For a company, the practical consequence is that the tone of an answer says nothing about whether it is correct. A wrong answer arrives phrased as firmly as a right one, with the same grammar and the same structure. The only filter that works is a check against the source, done by someone who knows what to look for.

The human side makes it worse. A Microsoft Research study presented at CHI 2025, “The Impact of Generative AI on Critical Thinking”, surveyed 319 knowledge workers who described 936 real examples of using AI at work.

In the Microsoft study, higher confidence in AI was associated with less critical thinking, as reported by the participants. People confident in their own ability to do the task checked more. The authors note that cognitive effort shifts towards verifying information, integrating the response and overseeing the task.

Together, the two studies suggest a practical risk. A model can state false information confidently, and trust in AI can reduce the effort spent checking. In our process, verification is therefore a mandatory step assigned to a named person before the output is used.

What data should never go into ChatGPT?

No data goes into ChatGPT unless the company has explicitly decided it may be sent to an outside provider. That decision belongs to leadership, because it depends on the type of account, on customer contracts and on the personal data involved.

Without a policy, employees make that decision themselves each time they paste something into a chat window. The company should set its data policy before choosing a subscription.

The 2024 Microsoft report cited above links the habit of bringing your own AI directly to data risk, in a year when cybersecurity and data privacy were the top concern of the leaders surveyed. And a tool brought from home runs on an account the company does not administer.

OpenAI spells out the difference between accounts on its enterprise privacy page, updated on 8 January 2026 and read on 6 October 2026. Data from ChatGPT Business, ChatGPT Enterprise and the API is not used to train models by default, unless the company explicitly opts in to share it.

According to OpenAI’s page on data controls, data from individual ChatGPT accounts may be used for training, depending on the account settings. An employee pasting a contract into a personal account makes a decision for the company, which does not administer that account’s settings.

The rule we recommend has three tiers, easy to remember. Anything public can go into any account. Anything internal goes only into the company account, under admin controls. Anything that belongs to customers, meaning personal data, contracts and negotiated prices, goes in only with a written decision and a legal basis behind it.

How do we split the work between people and AI?

We work every day inside an operating system built on AI models, which we call SIGNAL OS, and the line between processing and decision is written into it as a rule, in three colours. The examples below come from its working files and its log of mistakes, so each one has a date and a written trace.

Green lets the system read, convert and research on its own, and add information without changing existing sources, provided the work can be undone. Amber requires it to propose strategic work first, from positioning to messaging direction, for a person to approve. Red covers actions that require explicit human approval.

The red zone covers deletions, communications to real people, from emails to social posts, and website publishing. Each requires approval because simply reversing the action may not repair its effects.

When the system runs for hours without us, the red actions are collected into a single list that waits for a yes at the end.

The strongest argument for the line came from one of the system’s own mistakes. On 9 September 2026, the system recommended to an industrial film manufacturer, as proof, the case study of a cardiology clinic.

A person corrected the choice because the right reference was a packaging company. In that case, customer value came from protecting the packaged product, beyond the material itself.

We then measured the pattern, on 14 September 2026, across our own audit decks. Of 16 case study recommendations, only 4 linked companies with the same economic mechanism. The reference was chosen by industry, by what looks similar on the surface, while a person chooses it by mechanism, by what actually produces the value.

That is the difference between processing and judgment, measured on our own material.

The second proof concerns writing. On 23 September 2026 we tested whether a model could judge on its own whether a text was clear.

With the strictest instructions, the model caught all 7 bad sentences we had planted, and flagged as problems 11 of the 18 sentences in a text a person had already approved. We dropped the automated checker, and the final read stayed with a person.

The third proof comes from the quality gate every article on this blog goes through, whose rules were written from human corrections. On 3 October 2026, the gate rejected a whole batch of three AI-written articles on first pass, all on the same pattern.

Our own estimates had been written as facts, and the description of a company had been bent to fit the article’s theme, instead of being copied from its deck.

These three cases inform the same production rule. In our audit presentations, a person develops the strategic thesis and the visuals, while the system integrates analysis from public sources.

Since September 2026, every new deck made without talking to the company carries a slide called Basis of the analysis, separating what we know, what we estimate and what has to be validated together.

These corrections explain our pillar of AI that learns from dialogue. We store rules drawn from discussions and observed mistakes, and the system reads them before its next task. A correction can then improve later work without someone having to repeat the same explanation.

Our defensible business model pillar also explains the commercial advantage. We apply our methodologies to each client’s situation and retain decisions and corrections in the system’s memory. Even when a competitor uses the same ChatGPT, the results also depend on accumulated knowledge and the judgment of the people using it.

Which layers of a company does ChatGPT sit on?

ChatGPT sits on all four layers of our methodology, from L1, performance marketing, up to L4, positioning and category, with the AI Brain on layer L3. On each layer the line between processing and decision runs in a different place, and the figure below shows where.

Read the figure from the bottom up. The highlighted layer concerns positioning decisions, which shape how the market understands the company.

L1 · Performance marketing. ChatGPT can produce dozens of headline and ad copy variants, which can then be tested on real performance data. A person decides the budget, taking account of margins, delivery capacity and the company’s current situation, which the model may not have enough context to assess.

L2 · Revenue and sales process. Here the model drafts emails, summarises customer conversations and prepares proposals on written rules. Anything actually sent to a customer goes through the person who owns the relationship, because a promise made in the company’s name remains the company’s promise.

L3 · AI Brain. This is the layer that makes the line possible. Without a written memory of the company, with its procedures and decisions, the model starts from zero on every request and fills the gaps from imagination. I described in detail what an AI Brain is and which layers it is built from.

L4 · Positioning and category. On the top layer, the model handles the mechanical work very well, from collecting the claims on competitors’ sites to the test that shows what anyone could say. Choosing the attribute you own, the thing you want to be remembered for, remains a leadership decision, because it means giving up everything else.

The same layer matters from the opposite direction too, since AI models have started recommending companies. What a model learns about your company depends on what it finds written and confirmed, which we cover in the piece on how AI reads your website.

Is ChatGPT Business worth it for a company?

ChatGPT Business can be useful for a company whose employees already use AI on personal accounts. It provides a company-managed workspace whose data is not used for training by default. The subscription gives leadership more control over use, but the company still needs to decide which data can be shared and who approves the outputs.

Our infrastructure before marketing pillar applies here too. Before choosing a tool, the company needs written procedures, organised data the model can read and clear decision rights. These define what AI may do on its own, what it must propose and what a person must approve.

A subscription without written rules only speeds up the existing chaos. Everyone works faster, each along their own line, and the company ends up answering for decisions nobody officially took.

At the time of reading, a Standard subscription cost 25 dollars per user per month. The Air Canada case illustrates another cost to consider, that of a false promise reaching a customer and ending in a tribunal ruling.

When comparing costs, include the mistakes that could reach customers without a check. The subscription is only one part of the cost of using AI in a company.

How do you build a delegation sheet for your company?

A delegation sheet lists the company’s recurring tasks, assesses each against the three questions and names the person responsible. Set aside about an hour for a first draft covering one department, then review it with the team to check that the rules cover their actual work.

There are four steps, and the order matters. A written sheet replaces each employee’s instinct with a shared rule.

Step one. List the tasks that repeat every week in a department, from reports and proposals to replies to customers. The list starts from what your people actually do from Monday to Friday, because a list of everything AI could do never ends.

Step two. For each task, check whether its output can be verified against a source, whether mistakes can be corrected before they cause harm and whether you have named the person responsible. You can delegate when all three answers are yes. If any answer is no, keep approval with a person and define what AI can help prepare.

Step three. Next to every delegated task, write the name of the person who checks the output and the source they compare it with. This gives verification both an owner and a clear standard before anyone uses the model’s answer.

Step four. Decide which data goes into the model, on the three tiers above, and put the rule in the same document. Then reread it after a month, with the mistakes that came up in the meantime, and correct it. Every mistake caught becomes a new line in the sheet. That is how a system that learns actually gets built.

If you want to see how your company looks from the outside, including in the AI engines that already describe it to customers, you can take the brand diagnostic in Romanian, which gives you a score from 0 to 100 in two minutes.

Frequently asked questions

Can ChatGPT make business decisions?

ChatGPT can prepare a decision by gathering data, comparing options and writing the case for each. The decision itself stays with a person, because the model does not answer for the consequences, and in the Harvard and BCG experiment, on a judgment task outside the AI frontier, consultants using AI were 19 percentage points less likely to get it right.

What tasks should never be given to ChatGPT?

Keep human approval for actions whose consequences cannot easily be undone. These include pricing, positioning, messages sent directly to customers, publishing, deleting data and decisions about people. The model can help prepare them, but a person must approve the result and take responsibility for using it.

Is it safe to put company data into ChatGPT?

It depends on the account and on the data. According to OpenAI, data from ChatGPT Business and Enterprise is not used for training by default, while the versions for individuals can contribute to training. Anything that belongs to customers, such as personal data and contracts, goes in only with a written company decision.

Do I have to disclose that a text was written with ChatGPT?

For text generated or manipulated with AI and published to inform the public on matters of public interest, Article 50(4) of the AI Act requires disclosure of AI use. The obligation does not apply if the text has undergone human review or editorial control and a natural or legal person holds editorial responsibility for its publication. Our article on the EU AI Act explains the scope of this rule and obligations for other types of content.