OpenAI published a short note last month called "AI progress and recommendations." Most of the coverage went to the forecasts. The useful part for anyone running a firm is three ideas in it that explain why the AI you read about and the AI in your office feel like different things, and what to do about that this quarter.
1. The overhang
The note puts it plainly: the gap between how most people use AI and what it can presently do is immense. Most firms picture a chatbot and a better search box. The systems that exist today read a data room, draft the memo, match a thousand transactions, answer a call at 2am and book the consultation, and check their own work against the firm's rules.
That gap is the overhang. If model progress stopped tomorrow, there would still be a decade or more of work in applying what already exists to ordinary business workflows. Two forces are moving at once: the models keep improving, and the share of their capability that people actually use is catching up. The second force is the one a firm controls.
- A chatbot that answers questions
- Drafting an email faster
- Summarizing a document you paste in
- A pilot someone tried and dropped
- Waiting for the next model
- An intake agent that answers every call, qualifies and books
- A document agent that reads the client's files into working papers
- A close that runs on a checklist the system keeps moving
- A review queue holding only what a person must decide
- A system with an owner, measured in hours recovered
For a professional services firm, the overhang is not abstract. It is measurable. It is the difference between the hours your people spend reading, re-keying, chasing and assembling, and the hours they would spend if a system did the reading and they did the judgement. That difference is exactly what our readiness assessment scores, in four areas, in your own vocabulary. The score is a map of your overhang.
2. Spiky skills
The second idea is spikiness. AI capability is not a smooth circle that grows evenly. It is a jagged shape: far beyond human in some domains, still weak in others, and the shape keeps changing. The note describes systems that outperform the smartest humans at hard intellectual competitions while still having serious weaknesses.
People judge AI by its weakest spike, the miscounted letters, the odd hands in an image. Firms that get value from it do the opposite: they find the workflow that sits on a strong spike and build there first.
Reading and extraction
Long documents, statements, decks, forms. Strong spike. Systems built here pay off fastest and fail most gracefully, because a person reviews a draft rather than reading from zero.
Drafting from a template and the facts
First drafts of memos, letters, write-ups and reports in the firm's own structure. Strong, as long as every paragraph shows what produced it.
Matching and checking
Transactions to statements, invoices to purchase orders, claims to pages. Strong, and verifiable, which is the point.
Judgement calls and relationships
Which client to call, how to handle a difficult matter, what a partner would say. Weak spike. Leave it with people and design the system to hand these cases over.
One more thing the note implies. The labs are pouring most of their effort into making models better at code and research, because that is what makes the next model. The spikes that matter to a law firm or an accounting practice, reading, drafting, matching, are already tall and are improving as a side effect. You are not waiting on anyone.
3. Hours of work without a person
The third idea is the one to take seriously. A few years ago, systems could do tasks a person finishes in seconds. Now they do tasks that take a person more than an hour. The note expects days or weeks soon.
Two things drive that, and neither is "the model got smarter." Speed: the loop you watch in a coding tool, a step every few seconds, is limited by the compute you are given, not by the system. And autonomy: a harness that lets the model act, check, and act again without a person in every step. Even a system no smarter than a competent employee becomes a different kind of resource when it runs a hundred times faster and does not stop.
This is why every system we build has a named owner inside the firm, a definition of done agreed before the build, and a queue where the system hands things to a person. As these systems run longer without supervision, the places where they stop and ask become the most important part of the design, not the least.
What to do this quarter
The note ends without prescriptions. Here are ours, for a firm rather than an economy.
- Measure your overhang. Count the hours in one workflow, reading, chasing, assembling, before anyone builds anything. The assessment does a rough version in four minutes; a two-week look inside the firm does the real one.
- Build on a spike. Pick the workflow that is mostly reading, drafting or matching, and leave the judgement with the people who have it.
- Name the owner. Someone inside the firm, with hours a month, who decides what the system may do alone and reviews what it hands back. The assessment scores this as its fourth area for a reason.
- Run it alongside first. A few weeks of parallel operation, measured against the hours you counted, before it acts alone on anything.
The change the note describes is coming to every firm whether or not the firm plans for it. The firms that do well will not be the ones with the smartest model. They will be the ones that closed their own overhang one workflow at a time, with a person still in the loop where a person belongs.