The Cost of Waiting: precaution, opportunity and the ethics of not adopting AI

Paper six of six in the Ecaveo Working Papers on the higher-order issues in AI adoption, and the one most likely to be misread as an argument for moving fast. It is an argument for symmetry.

Why is deferral treated as the safe option?

Because its costs never appear anywhere. An adoption has a sponsor and a budget line. A deferral has neither. Risk registers record hazards rather than benefits that failed to occur. Audit examines what was done more readily than what was left undone. A director who blocked a pilot that would have succeeded is never questioned, because the outcome is never observed.

The paper put two invented firms side by side, facing the same proposal to give junior staff a drafting and summarising assistant. The first approves within a fortnight, with no testing against its own documents and no guidance on checking output. A trainee sends a client a summary citing a regulatory provision that does not exist, the partner apologises, and the tool is withdrawn. The second refers the proposal to a working group that asks for an assurance the tool will never err. No such assurance exists, so the decision is deferred a year. Trainees draft by hand, seniors spend evenings correcting them, and two competitors advertise faster turnaround at lower fees. Both firms made a poor decision, and only the first will ever find out.

What is the case for precaution, and where does it break?

The case is strong where harm would be severe and irreversible, and the paper gives it properly rather than as a straw man. Collingridge’s dilemma is real: early control lacks information and late control lacks power. The failures of the period were vivid, and a court sanction for filed submissions citing invented authorities arrived five days before the research closed.

It breaks as a default position. Regulatory caution has measurable costs, as the drug-lag literature established with stricter pre-market testing reducing the flow of new medicines without clear evidence of better average quality. Anticipation and resilience are both strategies, and a system that can recover is sometimes worth more than one that tried to foresee. Every option carries risk, including waiting, so precaution cannot serve as the resting state.

What did the early benefit evidence actually show?

Three studies dominated the picture in mid-2023, and the paper was careful about their status, treating all three as working papers or preprints covering a single task or a single firm, with no follow-up over years.

A staggered rollout of a support assistant across 5,179 customer service agents raised issues resolved per hour by 14 per cent on average, with a 35 per cent rise for the least skilled and minimal effect on the most experienced. A preregistered experiment with 444 college-educated professionals found writing time down and graded quality up, with the largest gains going to lower-ability participants. A controlled study of developers found a task completed considerably faster with an assistant, with heterogeneous effects that looked promising for career entrants.

The pattern mattered more than the magnitudes. If the gains fall largest on the least experienced, then timing becomes an ethical question rather than only a commercial one, because the people who benefit most from the tool are the people with the least say in whether it arrives.

How should the decision be made instead?

By asking five questions twice, once of deploying and once of waiting.

1. Severity. How bad is the worst credible harm, and for whom?
2. Reversibility. How easily can the harm, or the choice itself, be undone?
3. Time-sensitivity. Does the benefit decay, or pass to others, during the wait?
4. Distribution. Who bears each risk, and who receives each benefit?
5. Evidence threshold. What evidence would change the decision, and by when?

Evidence demanded should then match how hard the choice is to undo, in both directions. A reversible deployment, such as internal drafting under human review, needs modest evidence and a dated review. A hard-to-reverse deployment, such as one affecting decisions about people, needs strong evidence and independent testing first. A reversible delay needs light justification and a named owner for the review. A hard-to-reverse delay, where skills or clients are lost, needs board-level sign-off. Practice is weakest across that lower half, because a delay rarely gets a form at all.

What should change on Monday?

Record every deferral as a decision, with a named owner, a review date, the criteria that would reopen it, and an estimate of the benefit forgone and who forgoes it. Require adoption proposals to arrive with an assessment of the cost of not proceeding, set out with the same care as the risk assessment. Where a proposal is declined, minute who bears the forgone benefit and what is being done for them in the meantime.

Separate safety controls from the deploy decision, since verification and supervision can usually be arranged quickly and are a better first response than prohibition. Consult the people most affected on both sides, who are typically junior staff whose views rarely reach a board.

Reading it now, the evidence has moved in both directions. Two of the three studies were peer-reviewed and published, so the caution in the tagging was discharged. The skill-levelling finding has not replicated cleanly everywhere, and later work on experienced professionals has found smaller and sometimes negative effects, which means the paper’s own hedge did necessary work. Measured enterprise returns proved patchier than the 2023 tone suggested, which blunts the urgency of the delay argument for organisations that waited. What survives intact is the deferral register with a named owner, and the rule that evidence thresholds should match reversibility in both directions. Neither has become standard practice.

Frequently asked questions

Is this an argument for adopting AI quickly?

No. It accepts that some organisations should not adopt at all for some purposes. The argument is that a decision to wait should be recorded, owned and reviewed in the same way as a decision to proceed.

How can you estimate a benefit that was never taken?

Imperfectly, and the paper admits this is an unsolved measurement problem. A rough estimate recorded at the time is still more useful than the current position, where the figure is absent and therefore treated as zero.

Who bears the cost of a delay?

Usually the people with the least say. If early gains fall largest on less experienced workers, a year of deferral falls hardest on juniors, and a board is unlikely to hear from them.

What is a hard-to-reverse delay?

One where waiting itself destroys something: staff leave, a client moves, a capability atrophies, or a market position closes. Those deserve the same sign-off as an irreversible deployment, and almost never receive it.

Should a proposal be refused when no assurance of perfect accuracy exists?

No such assurance exists for any system, including the manual process being compared with it. The useful question is what evidence would change the decision and by when, which is answerable, rather than whether error is possible, which is not.