What actually changes when a model can draft anything?
The cheap part of professional work just got cheaper: producing a first draft. But drafting was never where firms spent their time or carried their risk. The work that matters — deciding what's right, checking it against the record, and signing your name to it — is unchanged. So the job shifts from producing to governing judgment.
New capabilities keep arriving. As The Data Ecosystem put it in 2025, "the tech is here" — whether or not any single tool is blocked, the underlying capability is now in the room, and operators have to plan around that reality rather than wait it out.
That plan is not "generate more." It's "govern what gets sent."
Was generation ever the bottleneck?
No. In most professional-services firms, producing the first version is the fast, cheap step. The slow, expensive, risky part is reviewing, correcting, and approving before anything reaches a client. Making drafts free speeds up the part that was never the constraint.
The evidence points the same way. Knowledge work — the reading, drafting, checking, and deciding that fills a professional's day — has always been review-heavy:
- McKinsey's research on generative AI (2023) estimated that current capabilities could automate activities absorbing 60–70% of employees' time — largely routine production tasks, not the judgment that surrounds them.
- The same body of McKinsey work found that even in the functions with the highest automation potential, a large share of value comes from tasks a human still has to verify, because the output feeds decisions someone is accountable for.
When you speed up production but leave review untouched, the review queue becomes the new bottleneck — now fed faster than before.
What happens when you skip the review step?
You ship a confident, plausible artifact that may be wrong — and in professional work, wrong reaches a client, a court, or a regulator. Hallucination — a model producing fluent output that is factually false or invented — is not a rare edge case. It is a documented failure mode with real consequences.
Concrete, named examples already exist:
- In Mata v. Avianca (2023), a New York federal judge sanctioned two lawyers whose brief cited six court cases that did not exist — fabricated by ChatGPT and filed without verification. The court's order is public record.
- Stanford's HAI research (2024) found that even legal-specific AI research tools hallucinated on a meaningful share of queries — the study reported error rates above 17% for some purpose-built legal tools, undercutting the assumption that a specialized tool removes the need to check.
The lesson isn't "don't use the tools." It's that an unreviewed draft is a liability with good grammar.
So what should the job become?
Governance, not generation. The firm's competitive edge moves from who can produce fastest to who can review reliably at scale. That means treating every model output as a proposal — staged for a human to approve — never as a committed action. This is Vantrow's spine: propose, never commit.
In practice, a governed desk — software that drafts and stages work but never sends it on its own — does three things:
- Drafts the routine artifact: the intake summary, the follow-up, the invoice, the memo.
- Stages it for a named human, with the source material attached so review is fast.
- Records the approval on an audit trail — a durable log of who approved what, and when.
Generation gets cheaper. Accountability does not move.
How is this different from an autonomous agent?
An autonomous agent decides and acts on its own; a governed desk drafts and waits. The difference matters most exactly where errors are expensive — client-facing work, filings, money movement. Autonomy optimizes for speed on tasks where the slow part was never the constraint, while removing the review that was.
For professional firms, the trade is bad: you'd be automating the cheap step and deleting the safeguard on the expensive one. Governed software keeps the human in the one place a human belongs — the moment before something is sent.
Where does a small firm start?
Start with one high-volume, client-facing artifact and put a review gate in front of it. Don't measure success as "drafts eliminated." Measure hours saved on drafting while keeping — and timing — the review step, so you can see whether the queue is actually shrinking.
A realistic target: controlled studies have shown meaningful first-draft time savings. An MIT/Stanford field study (2023) of support agents using a generative assistant found roughly a 14% average productivity gain, concentrated among less-experienced workers; a separate MIT working paper (2023) found generative assistance cut writing time on business tasks by about 40% for first drafts. Use figures like these as a drafting-time benchmark — not as permission to skip review.
FAQ
FAQ section
Below are the questions operators ask most about this shift.