Read enough AI business cases and they rhyme. Handling time down. Headcount avoided. Cost per case reduced. A sensitivity analysis, a payback period, an implementation risk register. It is a competent document and it is half a document, because every figure on it describes what the deployment takes out of the system and none of it describes what the deployment puts back into the people inside it.
This is not squeamishness about efficiency. It is an accounting complaint. A ledger with one column is not a conservative ledger — it is a wrong one, and it produces decisions that look excellent for four quarters and expensive for the next eight.
Anyone can use AI to cut. Almost nobody counts what it gives back.
What the missing column contains
Four entries, all of them countable, none of them currently on your template.
- —Hours given back. Time returned to the people who did the work, with a named destination — not time reabsorbed as throughput.
- —Skill kept alive. The capabilities the system now performs, and the plan for the humans who supervise it to keep practising them.
- —Human contact created. For anything sitting between people: conversations started, visits made, relationships kept.
- —Agency preserved. Whether a person can be reached, and can overrule the system, when it is wrong about something that matters.
Together they produce the net human score: what the deployment gives back minus what it takes. One figure, same page as the ROI, same meeting, same signature.
Why the second column changes the decision
Not by blocking things. In practice it does three specific jobs.
The objection, and the answer
The objection is that this is unmeasurable, soft, a values exercise dressed as analysis. Fair, if it stays abstract. It does not have to. Hours are on a rota. Skill is on a training record. Contact is in a caseload. Agency is in an escalation path you can walk through yourself with a stopwatch. Every entry in that column already exists somewhere in your organisation as data — it has simply never been asked for by anyone with signing authority.
The other objection is timing: we will add the human measures once the deployment is bedded in. That does not work, because the baseline is gone. You cannot count what people got back if nobody recorded what they had. Which is why Guardrail 01 puts the human goal next to the business goal at the point of design, on the same page, at the same board meeting.
Where to start
Take the next AI business case that comes to you and send it back with one question written on it: what do the people in this process get back, and how will we know? Not a policy, not a framework. One question, once, on a document that has an owner.
You will learn a great deal about your organisation from how long it takes to get an answer.