AI Didn't Make Us Faster. It Showed Us Why We Were Slow.

Ask a capable team how long it now takes to build a working prototype of a new idea, and the honest answer is often an afternoon. Ask how long until that prototype is in front of a real user, and the answer is frequently a quarter. Three meetings to discuss it. Four sign-offs to approve it. A wait for the next sprint to test it.

The gap between those two numbers is the most revealing metric in the modern enterprise. And it has nothing to do with technology.

In February, we argued in this newsletter that more data can reduce decision quality, because it overwhelms the person deciding. That piece was about the mind of the decision-maker. This one is about the machinery around them: who is allowed to decide, how many people must agree, and why the answer has barely changed while the cost of trying something has collapsed.

Be precise about what got cheaper

The claim that "AI makes everything faster" deserves scepticism, and the evidence is genuinely mixed. In early 2025, a carefully controlled study by the research group METR found that experienced software developers were actually slower with AI tools, while believing they were faster. By February 2026, METR had marked that result as out of date, concluding that developers are now likely being sped up by newer tools, though it could no longer measure by how much.

So be precise. AI has collapsed the cost of building an option: the prototype, the first-pass analysis, the draft model, the mocked-up customer journey. It has not collapsed the cost of acting on that option. Putting an untested model in front of real customers, real patient data or a regulator is as expensive as it ever was. In some respects it is more expensive. Banks now operate under the PRA's model risk principles. Consumer Duty raises the bar on customer outcomes. The EU AI Act adds new obligations. Some of the new approvals your organisation has introduced around AI are entirely justified.

The problem is not that governance exists. It is that most organisations run one approval process, sized for the expensive part, and apply it to everything, including the work that is now almost free.

The slowness was always there

None of this began with AI. A 2019 McKinsey survey of more than 1,200 managers found that fewer than half believed their organisation's decisions were timely, and 61% said at least half of the time they spent on decisions was used ineffectively. McKinsey estimated that this could cost a typical Fortune 500 company around 530,000 days of managers' time a year, roughly $250 million in wages.

That was before generative AI reached any desk. For years, the slowness hid behind a respectable excuse: experiments were costly, so it was rational to deliberate before running one. A business case took weeks to build because the analysis took weeks. A pilot needed a committee because it needed a budget.

AI has quietly removed that cover. When the analysis takes an hour and the prototype costs almost nothing, the three meetings and four signatures are exposed for what they often are: habits inherited from an era when trying things was expensive.

The doors have multiplied

Jeff Bezos gave this problem its best-known framing in his 2015 letter to Amazon shareholders. Some decisions are one-way doors: consequential and hard to reverse, deserving slow and careful deliberation. Most are two-way doors: if you walk through and dislike what you find, you walk back. His warning was that large organisations drift towards treating every decision as a one-way door, and pay for it in speed and invention.

The idea is a decade old. What has changed is the ratio. When a prototype costs an afternoon, the number of genuinely reversible decisions in a business has multiplied. A test on internal data, a mocked-up pricing page shown to ten customers, a model built in a sandbox and thrown away: these are two-way doors. Yet in most organisations they still pass through governance designed for one-way doors, because nobody redrew the map when the economics changed.

The result is a quiet absurdity. In many firms, the approval process now costs more than the experiment it governs.

The error runs both ways

This is not a licence to tear up the rulebook. The opposite mistake matters just as much. Treating a two-way door as a one-way door costs you time. Treating a one-way door as a two-way door costs you options, and sometimes your licence to operate. A customer-facing model, a change to how personal data is processed, or a pricing algorithm that touches vulnerable customers is not an experiment you can quietly undo.

The aim, then, is not less governance. It is governance whose weight matches the risk and reversibility of each decision. Heavy where the door locks behind you, light where it doesn't.

What to do this quarter

The uncomfortable exercise is simple, and few leadership teams have done it.

Take the last ten ideas your function approved, and the ten it didn't. For each, measure the time from first proposal to first real test, then split that time into doing and waiting. In most organisations, waiting dominates.

Then interrogate every sign-off on the path. What specific risk does it price? Is that risk reversible? Would the outcome change if the approver were removed for experiments that use internal or synthetic data, touch no customers and have a fixed end date? Where the honest answer is no, you have found a gate that exists out of habit, not prudence.

From there, build two lanes. A fast lane for contained, reversible experiments, with the authority sitting with the team. A full lane, with all the rigour your regulators expect, for anything that locks the door behind you. The CFO gets controlled spend. The CTO gets velocity. The Chief Risk Officer gets scrutiny focused where it matters, rather than spread thinly across hundreds of trivial approvals.

The gap ahead

Every serious competitor now has access to broadly the same models. The technology advantage that leaders were promised is being commoditised in real time. What will separate organisations over the next few years is not who has AI, but who has reorganised around what it makes possible.

That logic applies to expertise as much as experiments. It is why B2E built its model around bringing the right specialist in within days rather than after a procurement cycle: the slowest step in most programmes is rarely the work itself.

Your competitors can buy the same tools you can. They cannot buy your approval process. The question for this quarter is whether that is your advantage or your handicap.