Roughly 95% of generative-AI pilots never make it past the demo. The comfortable explanation is that the technology wasn't ready. The truthful one, most of the time, is that the workflow was wrong from the start — someone automated the thing that demos well instead of the thing that pays.

So before we build anything, we write the no-list. Four kinds of work go on it, almost every time.

1. Work where a mistake is expensive and rare

Our whole model runs on earned autonomy: an agent proposes, a human approves, and wider limits are earned through a track record you can inspect. That model has maths behind it — and the maths fails when the work is fragile but infrequent. An agent that acts once a quarter never builds a track record worth trusting, and the one mistake it does make costs more than the automation ever saved. If it's rare and it's dangerous, a person should do it, with a checklist.

2. Work you can't describe

If nobody in the business can write down how the job is actually done — not the org-chart version, the real one — an agent doesn't fix that. It amplifies it. The Blueprint exercise of mapping how work really moves is often the first time anyone has written it down, and sometimes the honest conclusion is: describe it first, automate it next quarter.

3. The work that is the relationship

Some conversations are the product. The renewal call, the apology that keeps a client, the awkward conversation about scope. Clients can smell delegation in these moments, and they should — showing up is the point. Agents can brief you for those conversations brilliantly. They shouldn't have them.

4. Anything below the governance line

In our own company there are things no agent does, ever, at any track record: move money, make legal commitments, touch production data. Not because the models aren't capable — because the failure mode is unbounded and the audit trail is an autopsy. Every business has its own version of that line. Finding it is a Blueprint deliverable; respecting it is what makes everything above the line safe to hand over.

What's left is plenty. Repeatable, describable, high-volume, reversible work — support queues, reconciliations, drafts, monitoring, chasing. That's where agents pay for themselves quickly and visibly. The no-list isn't caution for its own sake. It's what makes the yes-list trustworthy.