When AI Speeds Up Everything Except the Company

Lately I keep hearing a very specific complaint.

It is not “AI is useless.” It is something more twisted:

“I'm already fast.”

“The tools are in place.”

“The draft is ready in half an hour.”

And yet the project still crawls forward at its old pace. Some people even start to doubt themselves: maybe I'm using it wrong.

But the more I listen, the more I think they are aiming at the wrong target. Because what got faster is usually just one person. What did not get faster is the company.

Picture an ordinary weekday afternoon. An engineer sits in front of a screen. He uses AI to generate a working version of the code. Thirty minutes. Far faster than typing it out himself the old way.

He opens a pull request, feeling a little relief: today I can move one piece forward.

Then he starts waiting. Waiting for review. Waiting for test feedback. Waiting for someone to say “this can merge.” Waiting for another team to nod. Waiting for someone with more authority to reply.

Two days later, the code is still sitting there. Not because it could not be written. Because after it was written, no one could catch it.

The fastest part is already over. The slowest part is only just beginning.

This does not only happen at software companies. A proposal is drafted in half an hour; the decision takes two weeks. The analysis report is ready by the afternoon; the approval chain runs into next month. A customer reply can be generated instantly, but the contract stamp is still in the inbox of a manager three layers up.

AI accelerated a lot of “doing.” But the company is still stuck on “deciding.”

The ordinary conversation likes to ask: can AI improve productivity?

Future Lab would rather ask: once execution gets faster, why does the organization as a whole often fail to get faster with it? Where exactly is the real bottleneck?

Here is an intuitive hypothesis: AI is lowering the cost of execution. But it is not lowering, at the same rate, the cost of the organization — decision, trust, accountability, coordination, approval.

In the past, companies were slow largely because people were not fast enough, information was not fast enough, and production cost too much. Now those constraints are loosening. And so the things that always existed inside the old structure, but were covered by “everyone is busy,” start to surface: decision, trust, accountability, coordination, approval.

This is not a new disease. It looks more like what Future Lab has always called AI as a Revealing Agent: AI does not necessarily create problems, it makes visible where the old structure was truly stuck.

When coding was the bottleneck, no one could see review. Now that coding is no longer so scarce, review steps into the spotlight. In the future, when review is partly absorbed by tools too, what truly surfaces may be decision — and the person willing to sign for that decision.

We call this time structure — where the result has already arrived but the organization has not yet absorbed it — Organizational Latency.

It does not mean the company has no smart people. It does not mean the company has no AI. It means: the information has arrived, the AI has finished, but the system is still waiting on decision, trust verification, assignment of responsibility, coordination alignment, and approval.

It is a lot like latency in a network. Bandwidth went up. Packets travel faster. But if every round trip has to queue for confirmation, the felt speed still refuses to rise.

The individual's sense that “I got faster” usually measures bandwidth. The company's despairing “why is this still so slow” usually measures latency.

Here we have to stop and admit an objection.

Not in every scenario has execution actually gotten faster.

In 2025, METR ran a randomized controlled study on experienced open-source developers: under the tool conditions of the time, being allowed to use AI actually increased task completion time by about 19%. The twist is sharper still — developers had expected to be 24% faster beforehand, and afterward still believed they had been about 20% faster.

In other words, some “acceleration” is, first of all, an illusion. The time was eaten by prompting, proofreading, and waiting for generation.

So this essay is not asserting that everyone in the world has already been sped up by AI. It discusses another structure that is already happening, and increasingly common: once the execution layer really does get faster — or at least looks faster — why organizational throughput can still refuse to move.

The most persuasive case is not a traditional enterprise that “can't use AI yet.” It is Anthropic itself.

In March 2026, when Anthropic announced Code Review on the official Claude blog, they wrote it plainly: over the past year, code output per Anthropic engineer had grown 200%. Code review had become the bottleneck.

They also heard customers say the same thing over and over: developers were stretched thin, and many pull requests could only be skimmed rather than read carefully.

Cat Wu, the product lead for Claude Code, made the mechanism even clearer in an interview: now one person, with one prompt, can throw out a pull request that looks plausible; the burden shifts to the reviewer, who has to spend enormous time verifying edge cases.

Notice the structure of that sentence. AI did not eliminate the work. It pushed the center of gravity of the work from “production” to “verification.”

Anthropic's response is telling too: they use a multi-agent system to do deeper code review, raising the share of internal pull requests with substantive comments from 16% to 54%.

But they were still explicit: the tool does not approve the pull request. That is still a human call.

That sentence is almost the key to the whole essay. AI can lower the cost of verification. AI can even lower part of the organizational cost. But the organization still tends to keep one human node: whoever bears the consequences of “this can be merged in.”

On the surface, review is about finding errors. One layer deeper, it is about building trust: do I believe this output deserves to enter the mainline. Deeper still, it is about assigning accountability: if something breaks in production, on whose account does it go.

So when someone asks whether the biggest bottleneck of the AI era is intelligence, Future Lab's answer is closer to this: often it is not intelligence. It is accountability — the scarce node willing to answer for the result.

This runs on the same line as the Commitment Economy we proposed in AI Won't Replace People First — It Will Replace the Messengers: once execution gets cheap, what is expensive is the person willing to take responsibility.

This essay does not spin “the accountability bottleneck” up into a new term. It already lives inside the commitment economy. Organizational latency is the scarcity of accountability, developed across time like film.

McKinsey's 2025 research named the enterprise-side phenomenon loudly: the gen AI paradox.

Nearly eight in ten companies report using gen AI, and just as many report no significant impact on profit. For the higher-value vertical use cases, about nine in ten are still stuck in pilot.

Their point: the bigger challenge is often not the technology, but the people — trust, adoption, governance.

On the other side, MIT Sloan's account of “task-chain” research is sharper still: every time work passes from AI to a human, and back from the human to AI, it needs review, validation, adjustment. Those checkpoints drag down the whole system.

In other words, organizations are slow often not because they have no smart people. It is that the smart people are all placed along a very long chain of confirmation. Everyone is “responsible for a piece.” Very few are designed to be “responsible for the result.”

And so a familiar kind of spinning appears — what Future Lab calls a close relative of the Hollow Company: lots of meetings, lots of material, lots of tools, little progress.

After AI arrives, the spinning may get louder. Because generation cost has dropped, the organization can more easily fill the waiting with more documents, more proposals, more “AI first drafts.” It looks busier. Latency does not necessarily fall. Sometimes it even rises.

This is also the organizational echo of Judgment Outsourcing: when people stop making judgments and only perform relaying and confirmation, the system becomes safer, and also slower.

Slowness is sometimes risk control. Slowness is sometimes just that no one wants to be the person who signs. The two look very much alike. They need to be told apart.

You can crudely split a workflow into four layers:

Execution. Writing code, writing proposals, doing analysis, producing designs. The layer AI hits hardest.

Coordination. Aligning goals, syncing progress, handing off across teams. AI may lower the cost of translation and tidying up, and may also manufacture more half-finished work that needs syncing.

Decision. Whether to do it, when to do it, by what standard. A model can offer options; it is rarely allowed to bear the choice on the organization's behalf.

Accountability. Who answers when things go wrong; whose account the win is credited to. This layer is the hardest to automate, and the one that least deserves to be romantically “handed to AI.”

In the past, slow execution masked the three layers behind it. Now that execution has loosened, those three layers start to determine the felt speed.

So the company did not get faster — and that is not necessarily because AI failed. It may be because AI succeeded halfway: it accelerated the part that is easiest to accelerate, and forced the truly scarce part into the spotlight.

That is why, after the eighth essay asked “why do companies still exist,” this one has to ask next: the companies that exist — why can they still be this slow?

If the company is shifting from a production machine and a coordination machine into a container for trust and judgment, then what it manages is no longer only labor, and no longer only knowledge. What it manages is this: whether a trustworthy decision can be made at an acceptable latency, with someone to bear it.

Over the next year, more teams will talk openly about the review queue and approval lag. AI-assisted verification will spread. The human signature will most likely remain.

Within about three years, “execution speed” and “delivery speed” may be seen separately. One kind of company will shorten the handoff chain. Another will use more approvals to digest AI noise — that is the accelerated version of hollowing out.

Further out, the company may look more and more like an accountability-throughput system.

In the industrial age, the company managed labor. In the information age, the company managed knowledge. In the AI era, what exactly does the company manage?

Perhaps this: when intelligence is not scarce, the still-scarce flow of trust, judgment, and accountability.

For the ordinary person, this essay is not urging you to “oppose process.” It is urging you to see clearly which layer you stand on.

If your value lies mainly in execution, AI will keep squeezing the unit price. If your value lies in pushing things all the way to “deliverable,” and being willing to put your name at the key node, then what you handle is latency.

Future Lab has spoken of Personal Optionality: a person should try to keep room to act that is not defined by any single organization.

In an era where organizational latency develops into view, that option does not come only from changing jobs. It also comes from whether you have accumulated this kind of credit: others believe you reviewed it. Others believe you judged it. Others believe that if something breaks, you will not be the first to disappear.

Tools will keep getting faster. Trust does not get faster on its own.

If one day execution truly becomes cheap enough to approach zero, what is the company still managing?

More people? More information? Or fewer, but heavier, nodes of accountability?

When AI speeds up everything, why do some companies still not get faster?