AI For Professional Services: When The Deliverable Stops Proving Effort

AI for professional services has quietly crossed a line. It used to help you write. Now it produces the finished artifact, in your firm's own format, and the polish that used to signal care no longer signals anything at all.

On 3 September 2026, OpenAI announced GPT-6 Astra. Most of the coverage focused on the capability list, which is genuinely long. The line that matters for anyone who sells expertise as a document is much shorter, and it is worth reading slowly.

What did OpenAI actually announce?

OpenAI describes Astra as state of the art on computer use, browsing, software engineering, cybersecurity, science and professional work. That last category is the one to sit with.

The published capability list includes filling out online forms, updating customer records in a CRM, organizing a calendar, conducting research and drafting summaries inside your email or document editor, analyzing data and generating plots, and building a website then running frontend checks on it.

And this: it produces documents, spreadsheets and presentations that follow your templates and instructions, adapting when you add requirements or change direction partway through.

On the OSWorld 2.0 computer-use benchmark it scores 72.6% at roughly 40 minutes per task, against the previous model's 65.7% at roughly 75 minutes. On Mind2Web, task completion is about 1.9 times faster. It is available on the Plus, Pro, Business and Enterprise plans, through the API, and on AWS.

Why the template is the part that matters

Almost every reaction to a model release argues about output quality. That argument is comfortable and mostly beside the point, because quality improves on a schedule nobody controls.

The phrase worth your attention is "follow your templates."

A template is not a file. It is where a professional services business stores its method. The structure of your audit, the order in which findings appear, what gets compared against what, the way the argument narrows toward a recommendation: that shape is the accumulated judgment of everyone who ever did the work.

It exists so the work can be repeated by someone who was not there when the judgment was formed. That is precisely why firms guard the template far more carefully than any single document made from it.

Software that follows the template inherits the encoding without the experience that produced it. And the client cannot see the difference, because the client only ever saw the output.

The deliverable used to do two jobs

For as long as firms have billed for knowledge work, the document has carried two things at once.

• The recommendation. What you think they should do, and why.
• The evidence of effort. The formatting, the depth, the density of the appendix, the fact that it plainly took a long time.

The second job was never written down and never invoiced separately, but it did real commercial work. It is why a thin deck for a large fee feels wrong even when the advice inside it is correct.

That second job is now over. When anyone can generate something that looks like eighteen hours of work, looking like eighteen hours of work stops being evidence of anything.

The collapse of the template is the actual disruption, and it is not about quality. A firm can keep producing excellent documents and still lose the ability to prove, through the document, that excellence went into it.

What survives

What survives is everything that never fit inside the artifact.

The judgment call you talked them out of. The recommendation you refused to make because you had watched it fail somewhere else. The question you reframed before any analysis started, which changed what the engagement was even about.

None of that shows up in the deliverable. It shows up in the conversation around it, and firms have historically undersold it because the document was easier to point at.

Here is a practical way to think about the split.

Professional services tasks by whether they are getting cheap fast or holding their value

Getting cheap fast	Holding its value
Formatting to a house standard	Deciding what the real question is
Producing a first draft	Knowing which recommendation is wrong for this client
Assembling comparison tables	Choosing what to compare and why
Writing the summary	Owning the outcome after the decision
Turning findings into a deck	Being in the room when it lands badly

Nothing in the right column is new. What is new is that the left column used to fund the right column, and it is about to stop.

What "computer use" changes that chat did not

Most firms have already absorbed one wave of this. Someone used a chat assistant to draft an email, tidy a paragraph, or summarize a call. It helped, it saved twenty minutes, and it changed nothing structural.

Computer use is a different category, because the unit of work changes. Chat produces text you then have to place somewhere. An agent operating a computer produces the finished thing in the place it belongs.

Read the capability list again with that distinction in mind. Filling out forms, updating CRM records, drafting inside your document editor, running checks on a site it just built: none of those are text generation. They are task completion, including the assembly and the filing that used to sit around the thinking.

That assembly work is where a lot of junior time goes in a professional services firm, and it is almost never what the client thinks they are buying. It is what makes the thing they are buying exist.

The benchmark figures make the point in a less abstract way. Going from 65.7% at roughly 75 minutes per task to 72.6% at roughly 40 minutes is not a quality story. It is a throughput story. More tasks finished, per hour, with fewer handoffs in between.

The junior training problem nobody has solved

Here is the second-order effect, and it is the one that will hurt firms in three years rather than three months.

The traditional way expertise transfers in a services business is through the tedious work. A junior builds the comparison table, and while building it they learn what gets compared and why. They format the deck, and in formatting it they absorb the order in which an argument has to land. They write the first draft badly, get it marked up, and learn from the markup.

None of that is efficient. All of it is how the judgment in the template got into people's heads in the first place.

If the model does the tedious work, the tedious work stops teaching anyone. The senior people in your firm already have the judgment, so nothing looks wrong for a while. The gap opens quietly, in the cohort that never built anything by hand.

There is no settled answer to this yet, and anyone claiming otherwise is guessing. The firms thinking about it seriously are doing one of two things: deliberately having juniors do some work manually and then compare against the generated version, or moving apprenticeship out of production entirely and into structured review, where the junior critiques rather than assembles.

Both are more expensive than letting the tool do it. That is the actual decision, and it is a decision about your firm in 2030 rather than your margin this quarter.

Does this mean the work disappears?

No, and predictions that it does tend to come from people who have never watched a recommendation get rejected for reasons that had nothing to do with its quality.

Someone still has to be accountable. A model can propose a pricing change, a restructure, a campaign. It cannot be the party that is answerable when the change is wrong, and accountability is a large part of what professional services actually sells.

But accountability has to be priced and made visible, because it is no longer implied by the artifact. That is a real change in how these firms describe themselves, and most have not started.

What to do in the next month

Three things, in order, none of which require buying anything.

1. Audit your invoice, line by line. Mark each line as "hard" or "tedious." The tedious lines are the ones that will quietly shrink first, and they will shrink through scope conversations rather than lost accounts.
2. Move the thinking out of the document. If the only place your judgment appears is inside a deliverable that now looks generatable, you have made your most valuable contribution invisible. Put it in the working sessions, the framing, the recommendations you decline to make.
3. Use the tools on the tedious half yourself. The firms that struggle will be the ones still charging for formatting while their clients generate it internally. The ones that do well will have moved that cost to near zero and repriced around judgment.

The uncomfortable version of all this is short. Your standard deliverable is about to be reproducible by a client with a subscription, and they will not tell you when they first try it.

If a client received a version of your report that followed your exact format and read perfectly well, how long would it take them to notice it was wrong, and would they blame the format or you?