AI can help you do someone else’s job. It cannot make you the expert.

OpenAI’s analysis looked at what people asked ChatGPT to do, not what happened after the answer arrived. The examples reported by Axios include making marketing material, troubleshooting software, doing financial calculations, and explaining regulations. Customer service, design, and HR workers had some of the highest crossover rates.

The denominator matters. The 44% figure applies to occupation-specific requests after generic tasks were stripped out. It is not a claim that 44% of all work has changed hands, that 44% of workers are doing a second profession, or that ChatGPT completed those tasks successfully.

The source population matters too. These were U.S. business users of ChatGPT. That makes the data useful for understanding one large product’s workplace use, but it is not a census of every worker, company, or AI tool.

OpenAI’s own earlier workplace material points in the same general direction: people use ChatGPT across writing, research, programming, and analysis, while coding and data work increasingly appear outside technical departments. That supports the boundary-crossing story. It still does not supply the missing outcome measure.

Imagine a five-person shop preparing a new product page. There is no in-house analyst, security engineer, or copy desk. An AI assistant can help the team inspect a spreadsheet formula, explain why a domain record is failing, and draft three versions of the product description before lunch.

Without that help, some of the work waits, gets sent to a contractor, or lands on the one person who always becomes the unofficial expert. A decent first pass can make a specialist’s hour more focused: here is the calculation, here is the assumption, and here is the exact point we do not trust.

Axios reports a modest size difference consistent with that use. Nearly 19% of work-related requests among typical users at the smallest businesses crossed occupational lines, compared with roughly 16% at larger firms. That does not prove small businesses became more productive. It does fit the plain reality that small teams have fewer specialists sitting nearby.

Used well, AI lowers the cost of getting oriented. It can translate jargon, expose the shape of a problem, and help someone ask a sharper question. Those are meaningful gains, especially when the alternative is guessing from scratch.

The risk changes with the task. A rough event flyer can survive an awkward sentence. A payroll calculation, benefits notice, security change, or interpretation of a regulation can affect someone else’s money, access, or rights.

In those cases, the AI has not transferred the specialist’s judgment to the person holding the prompt box. It has produced an artifact that may look finished before the user knows what to challenge. Fluency can hide the exact gap expertise normally catches: an outdated rule, an unusual jurisdiction, a missing exception, a formula that works on nine rows and fails on the tenth.

There is also an organizational trap. Once people can produce plausible work outside their role, managers may quietly expand the role without changing the title, training, time, or pay. ‘You can ask the AI’ becomes a cheap answer to every expertise gap. The company saves a handoff; the employee inherits review work and liability they were never prepared to carry.

That is why task crossover should not be read as specialist replacement. The OpenAI study does not measure output quality, productivity, hiring decisions, or whether the request created a new responsibility rather than helping with work the employee already did. Those omissions are not footnotes. They are the boundary between an interesting usage pattern and a labor-market conclusion.

Mara Vale worries about the sentence that comes after a successful demo: ‘Great, you own this now.’ If an HR generalist uses AI to draft a policy note, the new tool should not quietly make that person the legal department. Name the reviewer, the deadline, and who carries the consequence before calling the handoff removed.

Mina Torres is more interested in what the first pass can give back. A parent running a small business may not need an AI to make the tax decision. They may need it to organize the receipts, flag the strange line, and prepare three precise questions so the accountant call takes 20 minutes instead of an afternoon.

Those positions are not opposites. AI can widen what a person can begin while preserving a clear line around what they should not finish alone. The useful design is not ‘generalist replaces specialist.’ It is ‘generalist arrives better prepared, and the specialist can see the path taken.’

First write the decision line: what will happen if this answer is accepted? A draft stays a draft. A calculation may set a price. A configuration change may lock out a customer. A policy explanation may change an employee’s benefits choice. The consequence tells you how much review the work deserves.

Then write the expertise line: what would a competent specialist check that I might not know to ask? Keep that question beside the AI output. For finance it may be the accounting treatment and source period. For IT it may be rollback and access scope. For legal work it may be jurisdiction, dates, and exceptions. For marketing it may be claim substantiation and permission to use an asset.

For low-stakes, reversible work, proceed and inspect the result. For work that touches another person’s money, access, health, legal position, or shared record, use the AI to prepare—not to erase—the expert review. Preserve the source, assumptions, changed fields, and unresolved questions so the next person does not have to reconstruct the chat.

The new research is useful because it catches a change before job titles do. People are already reaching across professional boundaries with AI assistants. Whether that becomes freedom, hidden extra work, or expensive error will depend less on the number of crossed boundaries than on whether teams still know where to stop.