The work you absorbed without noticing

Some time in the last two years you started doing work that is not your trade. You wrote the privacy notice. You built the pricing model in a spreadsheet you would not show an accountant. You debugged the form on your own website, designed the logo, answered a support ticket about a refund policy you also wrote. Nobody decided you should do these things. There was a job to be done, the tool made a plausible attempt at it, and the attempt was good enough to ship.

That is a real change in what your business is. This guide is a method for finding the crossing points, sorting them by what it costs when the output is wrong, and deciding for each one whether you keep it, put a check on it, or hand it back to somebody licensed to do it. It is written for solo operators, freelancers and owners of teams small enough that there is no second person to ask. If you have a legal department, a controller and a design lead, you have a different problem: getting people to admit what they are already doing. This is about deciding what you should be doing.

Small operations cross over the most

On July 27, 2026 OpenAI published an analysis of more than 800,000 messages from US ChatGPT users, looking at work tasks relative to the sender’s own occupation. It calls the pattern task crossover, defined as work historically associated with one occupation appearing in the AI use of people in another [1]. Across all work-related messages, 16.8% were about tasks associated with another occupation [1].

That headline number understates it, because most work is generic. OpenAI splits usage into shared or generic tasks that turn up in every job and tasks specific to one occupation, and 61.5% of the use is generic [1]. Look only at the occupation-specific messages, and 43.5% of them are about tasks associated with another occupation [1]. Close to half of the specialist-shaped work in the sample is coming from people outside the specialism.

The number also moves with the size of the organization, and it moves against you. The outside-occupation task share is 18.9% for users in workspaces with 2 to 5 seats and 16.3% in workspaces with over 100 seats [1]. That is the expected direction and it is worth saying out loud, because most writing about AI and job design is aimed at the 100-seat end, where a specialist exists and the question is whether to route work to them. At 2 seats there is no specialist to route to. The crossover is not a management choice you are evaluating. It is Tuesday.

The share differs sharply by trade. Among occupation-specific messages, outside-occupation tasks account for 77% for customer experience workers, 75% for designers, 69% in human resources, 56% in legal and 53% in marketing [1]. If your own work sits in one of those, the reasonable prior is that a large part of what you produce with an AI tool is not the thing you were trained to do.

Most of what you cross into is not deep expertise

The temptation is to read those numbers as evidence that you have quietly become a lawyer, a designer and a financial analyst. You have not. OpenAI grouped the data into eight occupation groups, and two task types, financial calculation and technology troubleshooting, each appear among the three most common outside tasks in all seven other groups [1]. That is the connective tissue of running any business, and it is now absorbed by whoever happens to open a chat window.

OpenAI’s earlier study of consumer usage points the same way. It analyzed 1.5 million conversations and found that approximately 30% of consumer usage is work-related and approximately 70% is non-work [7]. Within the whole set, 49% of messages were classified as asking, meaning seeking advice or information, against 40% classified as doing, meaning producing something [7]. The dominant mode is asking a question you would previously have asked a colleague, or not asked at all.

This matters for how you draw the boundary. The crossover that is safe and worth keeping is mostly the shared work: arithmetic, first drafts, troubleshooting, understanding a document before you act on it. The crossover that needs a decision is the narrow band where you are producing a specialist deliverable that somebody relies on. Sorting by occupation label will not separate those two. Sorting by consequence will.

Sort every crossing by what being wrong costs

Put each task you have absorbed into one of three buckets, and ignore whose job it used to be.

The first bucket is work that is cheap to be wrong about. Internal notes, first drafts, code you will run and watch fail, a summary you are reading yourself, a spreadsheet formula you will sanity-check against a number you already know. You are the only person exposed and the error surfaces immediately. Cross into this freely, and stop feeling like an impostor about it.

The second bucket is work that is expensive to be wrong about and that nobody licenses. A price you quote a client. A contract clause you copied from a template. A tax number you put in front of your bookkeeper. An email to 4,000 subscribers. There is no rule against you doing it, and the output will look correct whether or not it is, which is exactly the failure mode. Anthropic’s own guidance on reducing hallucinations is blunt about the ceiling: the techniques significantly reduce hallucinations but do not eliminate them entirely, and you should always validate critical information, especially for high-stakes decisions [6]. This bucket does not need a specialist doing the work. It needs somebody checking the work, on a defined trigger, before it goes out.

The third bucket is work where the license is the point, and the quality of the output does not change the answer. The IRS states that anyone who prepares or assists in preparing federal tax returns for compensation must have a valid 2026 PTIN before preparing returns [4]. A perfect return prepared for a fee by an unregistered person is still an unregistered person preparing a return for a fee. The tool vendors draw a matching line in their own terms. OpenAI’s usage policies prohibit the provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional [2]. Anthropic’s Usage Policy names its high-risk domains one by one: legal interpretation and guidance, healthcare and medical guidance, insurance underwriting and claims, financial decisions including investment advice and loan approvals, and employment and housing determinations. When its products are used for advice, recommendations or subjective decision-making directly affecting individuals or consumers in those domains, a qualified professional in that field must review the content or decision prior to dissemination or finalization [3].

The same policy attaches a second requirement, and it catches more small operators than the first. If model outputs are presented directly to individuals or consumers, you must disclose to them that you are using AI to help produce your advice, decisions or recommendations, at a minimum at the beginning of each session [3]. If you use Claude to help draft advice that reaches a client, disclosure is not optional courtesy under that policy. It is a condition of the terms you agreed to.

The task list you need is already written and public

The hard part of this exercise is not judgment, it is inventory. Most people cannot list what they did outside their own trade last month, because each individual instance was small and felt like nothing at the time.

There is a public reference that makes it mechanical. O*NET OnLine, sponsored by the US Department of Labor’s Employment and Training Administration, covers 1,016 occupation titles and codes and draws on a database of over 19,000 occupation-specific task statements [5]. Open the occupation you suspect you have been crossing into, read its task list, and mark the ones you have personally performed in the last quarter. It is dull, and it is the only part of this that produces evidence rather than impressions.

Then look at how the marks are distributed, because the shape matters more than the count. Six marks concentrated in one adjacent trade is a role that has quietly expanded, and it should be formalized, trained for, or hired against. One mark each in four unrelated trades is the connective tissue from the first bucket, and it needs no decision at all. Without the list, both look identical from the inside, which is why the decision never gets made.

The subscription is not the cost you should be modelling

The tool is cheap. Claude Pro is $20 per month billed monthly, or $17 per month with the annual subscription discount, and a Team standard seat is $25 per month billed monthly, or $20 per seat per month billed annually, for teams of 2 to 150 [8]. No crossover decision turns on that.

The cost that matters is your hours and the review gate. If you spend 6 hours a week on work outside your trade, you are spending most of a working day on tasks you are slow at, and the alternative is not zero, it is a specialist’s fee for a fraction of those hours. Sometimes the arithmetic says keep the work, because a specialist’s minimum engagement is larger than the job. Often it says the opposite. Either way the hours are scattered across the week and never show up as a line item anywhere, so the comparison is easy to skip.

Run the numbers before the next time you decide by default. The calculator below is the whole of the model.

Write the boundary down before it writes itself

The output of this exercise should be one page you can actually find again, in whatever you already use, Notion or a text file or the back of the same document that holds your prices. It needs three things.

A list of what you do not touch. The licensed categories from the third bucket, named specifically, so the decision is made once when you are calm rather than at 11pm when a client asks a question you could plausibly answer. Anthropic and OpenAI have both published their versions of that list, and adopting theirs as a starting point costs nothing [2][3].

A review gate for the second bucket, with a named trigger and a named reviewer. “Anything that goes to a client with a number in it gets read by my bookkeeper first” is a policy. “I will be careful with financial stuff” is not. The trigger has to be something you can notice while you are working.

A date to re-read the whole thing. The boundary moves because the tools move, and the version of this page you write today will be wrong within a year in ways you cannot predict from here.

What still goes wrong

The crossover data describes US ChatGPT users, and it counts messages rather than finished work [1]. A message about a legal task is not a completed legal task, and OpenAI’s own earlier study found that asking is a larger share of use than doing [7]. Somebody sending 40 messages about a contract may have produced a signed contract or may have produced a better understanding of one their lawyer already wrote. Use the numbers as evidence that the pattern is large and normal, not as a measurement of how much specialist output is being produced by non-specialists.

The genuinely dangerous zone is the one this method can only partly fix. It is the work where you now know enough to produce something competent-looking and not enough to notice the specific thing that is wrong with it, and where the error stays invisible for months. Tax positions, contract indemnities, dosage, structural loads, employment terminations. The three-bucket sort helps because it makes you name those categories in advance, but it does not give you the knowledge to review your own output, and no prompt does either. The check has to come from outside you. If you cannot name who provides it, the honest answer for that task is that you are not doing it.

Vendor policy is also not law, and law is not uniform. The OpenAI and Anthropic policies quoted here are contractual terms with the vendor, read on the retrieval dates listed in the sources below [2][3], and they can change. Your actual exposure comes from your jurisdiction’s licensing rules, your professional body if you have one, and your insurance, none of which are covered by agreeing to a usage policy. Treat the vendor’s list as a floor and check the trade you are crossing into.