A marketing team asks an agent for twelve variants of a campaign email on a Tuesday morning. By eleven o'clock they have them, plus subject lines, plus a version for the lapsed-customer segment nobody had ever found time to write for. The emails go out on the Thursday of the following week. The work took forty minutes. The rest was a queue: brand, then legal, then a director who was travelling and approves things from her phone in batches on Sunday nights.
Nobody in that story did anything wrong. Each check was reasonable, and each used to be a rounding error against the time it took to make the thing. That ratio has flipped, and most companies have not noticed, because nobody measures waiting.
When making is cheap, waiting is the cost
For most of the history of office work, production was the expensive part. Drafting a contract took a lawyer two days; signing it off took twenty minutes of a partner's time. A week in the queue was annoying but proportionate. Approvals were a thin layer of judgement on a thick slab of effort.
Agents have shaved the slab. A first draft of a supplier agreement, a pricing analysis, a set of job adverts, a refactor of a billing module: these now arrive in minutes or hours. The approval layer has not shrunk at all. It is the same director, the same committee, the same request for a version with tracked changes. So the share of elapsed time spent waiting for a human has gone from a sliver to most of it.
Engineers have a name for this shape. Amdahl's law says that speeding up one part of a process only helps in proportion to how much of the process that part was. Make the parallel bit infinitely fast and you are still stuck with the serial bit. In most organisations, the serial bit is a person with a calendar.
Why approvals don't speed up on their own
You might expect approvers to feel the pressure and move faster. Mostly they don't, for reasons that have little to do with laziness.
Approval is how a lot of senior people stay in contact with the work. If the head of procurement stops signing off purchase orders above a threshold, she loses her view of what the company is buying. The signature is partly control and partly a news feed. Take it away and you have removed something she values, even if she would never describe it that way.
Approval also carries blame unevenly. An approver who waves through a bad contract is remembered. One who sits on a good contract for a fortnight is not. The incentives all point towards holding, asking one more question, waiting for the next meeting. That is the principal–agent problem in miniature: the company wants speed, the individual wants not to be the name on the mistake.
And there is now more to approve. When producing a thing costs almost nothing, people produce more things. Three proposal options instead of one. Weekly reports instead of monthly. Every extra artefact lands in someone's inbox. Call it an office-sized Jevons paradox: make something cheaper and you get more of it, not less effort spent on it. The approver who was a mild bottleneck at last year's volume becomes a hard one at this year's.
What the tax actually costs
The obvious cost is time to market. The less obvious costs are worse.
Work goes stale in the queue. A sales team builds a tailored proposal in an afternoon while the prospect is warm. Pricing approval takes eight working days. By then the prospect has spoken to two competitors and the proposal answers last week's questions. The agent's speed was real; the customer never felt it.
People batch to avoid the queue. If every request costs a week of waiting, the rational move is to save requests up and send them together. Batches are larger, harder to review and more likely to be sent back whole. The approver then takes longer, which encourages bigger batches. You can watch a team slide into quarterly releases this way without anyone ever deciding to.
And people learn to route around it. Some will find a way to call the thing a draft or a pilot so it never needs sign-off. Others will quietly stop using the agent for anything that needs approval, because there is no point being fast at the part that was never slow. The cheapest way to make an AI investment look useless is to put it upstream of a slow signature.
Measure the wait, not the work
Most teams can tell you how long a piece of work took to produce. Very few can tell you how long it sat. Start there. Pick one flow that matters, say supplier onboarding or campaign launches, and for a month record two timestamps per item: when it was ready for review and when it was approved. The gap is your approval latency. Put it next to production time and the picture usually draws itself.
Then look at where the gaps cluster. They are rarely spread evenly. One person, one committee or one step tends to own most of the waiting. This is not an accusation. That person is usually overloaded precisely because everyone trusts their judgement. But you cannot fix a queue you haven't located.
It also helps to ask approvers what they actually check. Not the formal policy, the real thing. The answer is often short: "I look at the discount and whether the payment terms are standard." If the real check fits in a sentence, it can be written down and applied earlier, by the person submitting or by the agent itself, with the approver seeing only the exceptions. Approval, looked at closely, is just verification with a job title, and the skills involved are the ones we described in our piece on moving from creator to verifier.
Redesigning sign-off for the speed you now have
A handful of moves tend to work. None of them is "approve everything", and none needs a new platform.
- Tier by risk, not by habit. Most approval rules were set by amount or by department years ago. Redraw them around what can actually go wrong, and let the low-risk tier ship on rules rather than signatures.
- Approve the rules, not each instance. A legal team that signs off a clause library once can let sales assemble standard contracts without review and see only the deviations.
- Set a default with a deadline. If an approval isn't given or refused within an agreed window, the work proceeds. This sounds reckless and in practice concentrates minds wonderfully.
- Review after, not before, where mistakes are small and reversible. Sample a slice of the output each week instead of gating all of it.
- Give approvers their news feed back another way: a short weekly digest of what went through, so losing the signature doesn't mean losing sight.
The last point is the one people skip, and it is why the others fail. Take a signature away from a senior person without replacing what it gave them and they will find a reason to put it back, usually after the first small mistake. That mistake will come, because every system makes them. The negotiation is less about process than about status and reassurance, and it goes better when it is had openly than when the new rules simply appear in a policy update.
A fair test for Monday morning: take the last thing an agent produced for your team and find out where it is now. If the answer is "with someone", you have found your real bottleneck. It has a name, and it has a calendar.
