The answer is sometimes do nothing.

The easiest way to waste money on AI is to decide that AI must be the answer before anyone has stated the problem.

Pressure arrives from a board, a competitor announcement or a demo that made difficult work look effortless. A team is told to do something with AI. Activity follows: workshops, vendor calls, a pilot, perhaps a new steering group. The organisation becomes visibly busy without becoming more capable.

Kingsbury holds a less convenient position:

Change carries the burden of proof. Doing nothing is a legitimate decision when it is made deliberately, evidenced properly and given a condition for review.

That is not an argument for caution as a personality trait. It is an argument for making the intervention earn its place.

There are three kinds of no

“Do nothing” is too easily heard as a single, permanent refusal. In practice, it describes three different decisions.

Not yet

The opportunity may be real while the conditions are wrong. The data is unreliable. The process is changing. The accountable owner has not been named. The likely regulatory position is unclear. Waiting can be the fastest route to a useful result if it removes uncertainty that would otherwise be built into the system.

A decision to wait must include a return condition: a date, an event or a threshold. Without one, “not yet” is only neglect wearing sensible clothes.

Not with AI

Some apparent AI problems are ordinary service, process or information problems.

A clearer form may beat a conversational interface. A search index may beat retrieval-augmented generation. A deterministic rule may beat a model when consistency matters more than flexibility. Removing a redundant approval may create more value than automating it.

The UK government’s AI Playbook makes the same underlying point: begin with goals and user needs, use the right tool for the job and consider whether a simpler approach is better. AI is a means, not an objective.

NIST’s voluntary AI Risk Management Framework Playbook goes further in operational terms: viable non-automated, partly automated and procedural alternatives belong in the comparison. A responsible AI decision is an intervention decision first.

Not at all

Some work should stop because the outcome is not valuable enough, the risk is not defensible or the proposed change makes an already weak service faster without making it better.

Stopping is not a failure of delivery. It is delivery of a decision. The GOV.UK Service Manual is explicit that stopping after discovery can save time and money. An organisation that cannot stop weak work has not reduced uncertainty; it has merely funded momentum.

Make the intervention compete

Before approving a build, compare the proposed intervention with its real alternatives. One page is enough. The discipline matters more than the document.

1. Name the consequence

What must become observably better for a person, a service or the organisation?

“Adopt generative AI” is not an outcome. “Reduce the time between receiving a complete application and making an accurate eligibility decision” might be. It identifies work that can be observed without prescribing the mechanism.

2. Build the option ladder

Start with the least complex credible response. The options might be:

1. remove unnecessary work;
2. clarify information or responsibility;
3. change the process;
4. add a deterministic digital tool;
5. add a model where judgement over variable material is genuinely useful.

The ladder is not a maturity model. Higher is not better. It prevents technical complexity being mistaken for ambition.

3. Compare on the same terms

Assess each option against the same questions:

• What benefit could it create?
• What evidence supports that expectation?
• What is the consequence of an ordinary failure?
• What will the whole service cost to change and operate?
• How easily can the decision be reversed?
• What capability must the organisation acquire and retain?

Do not give AI a more generous comparator. Include the checking, exception handling, monitoring and supplier dependency around it.

4. State why AI wins

Complete this sentence: “AI is the right intervention because…”

The answer should name a material advantage: the variability it can handle, the judgement it can assist, the scale it can absorb or the previously impossible service it enables. “Because we need to learn about AI” is a reason to run a small learning exercise, not to alter a live service.

AI should win the comparison, not be exempt from it.

5. Choose and set the return condition

Record one decision: proceed, simplify, wait or stop. If the answer is wait or stop, state what new evidence or changed condition would justify another look.

If the answer is proceed, the next task is to design a bounded test. That is where baseline, threshold, ownership and exit belong. Kingsbury’s position on that work is set out in If a pilot cannot be killed, it cannot be trusted.

A small example

Imagine a shared inbox that receives several thousand requests in inconsistent language. The proposed answer is an AI triage agent.

The test above might expose three different interventions. Better form design could remove ambiguity at entry. A set of explicit rules could route the majority of requests. A model might then assist with the remainder, where language varies and the cost of a human checking the suggestion is acceptable.

This is not less ambitious than deploying an agent across the whole inbox. It is more exact. Each component is used where it earns the right to exist.

The decision record

A defensible “do nothing” decision should fit on one page:

• the outcome sought;
• what currently happens;
• the options considered, including a non-AI option;
• the evidence available and what remains unknown;
• the decision: proceed, simplify, wait or stop;
• the person accountable;
• the condition that would cause the decision to be reviewed.

Publish it internally. Revisit it when the condition changes. A recorded refusal is useful organisational memory; an unrecorded refusal becomes folklore.

What this position rules out

It rules out buying technology to create the appearance of movement. It rules out treating novelty as evidence. It also rules out reflexive scepticism: when an AI intervention survives the same comparison and produces the strongest case, proceed decisively.

The objective is not to do less. It is to make change happen where change is warranted, and to leave everything else alone on purpose.

Sources and further reading

• The UK government’s AI Playbook on starting from user needs, using the right tool and managing the full AI lifecycle.
• NIST’s voluntary AI RMF Playbook on comparing AI with non-AI, partly automated and procedural alternatives.
• The GOV.UK Service Manual on how the discovery phase works and why stopping can be a valid evidence-led outcome.