Build it, or rent it forever

The licences are bought. The usage dashboard looks acceptable. Very little about how the work actually gets done has changed.

For every 500 people on Copilot or its equivalent, you are already spending around £150,000 a year on licences. Scale that to your headcount.

That money is committed. The question isn't whether to spend on AI. It's whether the spend you have already made produces anything.

A licence-only rollout has a predictable second act. When the tools don't deliver on their own, the next purchase is implementation: specific agents, built by consultancies, for specific tasks.

Current market rates for an enterprise agent build run £80,000 to £160,000 each, and £4,000 to £16,000 per month per agent after that, for maintenance and monitoring. Four agents is most of a million pounds committed before anything is running, and a recurring bill that does not stop.

Three things are wrong with that path, and only one of them is the price.

The people building it don't do the work it's meant to do. Specifications get written by people a step removed from the task, which is how you end up with an agent that technically works and nobody uses.

Every change goes back to the supplier. You have bought a capability you cannot modify. The first time the process shifts, you are raising a purchase order to alter something your own team could have adjusted in an afternoon.

Your organisation is no better at this than it was before. You have bought output. You have not bought competence, and competence is the thing that would have let you make the next decision without buying anything.

That third one is the expensive one, because it is the one that repeats.

Build the capability internally first, so that when you do buy tooling you can specify it properly, judge whether it works, and change it later without a procurement cycle.

This is not an argument against buying software. It is an argument against buying the ability to think about it.

Assessment. A measured baseline of where AI adoption in your organisation actually stands: where it's real, where it's being performed, where it's stalled, and why. Usage data cannot tell you this, which is why the return question is currently unanswerable.

Internal champions. One or two real workflows, redesigned by the people who run them rather than for them. A pilot ends in a decision about whether to roll it out. This ends with people who can already do the thing, and colleagues who can learn it from them without another programme.

Readout. Findings in your board's language, with the gap costed.

£8,000 per month. £24,000 for the ninety days.

Licence spend, per 500 people	~£150,000 a year
One consultancy-built agent	£80,000–£160,000, plus £4,000–£16,000 a month
Ninety days building capability internally	£24,000

The comparison is not £24,000 against nothing. It is £24,000 against finding out the expensive way.

• Evidence about whether the licence investment is producing anything, in a form you can put in front of a board
• One or two workflows that measurably changed, and the people who changed them still in the building
• Enough internal capability to specify the next thing properly, or to build it

Approve ninety days. £24,000. Reassess at the end with evidence in hand.

If it's working, it continues in ninety-day sprints. If it isn't, we stop. You keep the baseline either way.

No annual commitment. No renewal anyone has to be talked out of.

Jonas Haefele, Slow Works. Twenty years as the supplier: brought into organisations to run the training, deliver the system, hand over whatever had been signed off. It is a good vantage point on why bought capability so often fails to land.

An MSc in organisational psychology, an AI-adoption measurement instrument built from that research, and a continuing set of interviews with people doing this work inside real organisations.