Blog

Every agentic demo you have ever been impressed by was greenfield. A clean repository, a well-scoped problem, a fresh prompt, and an audience watching live. The agent writes the feature, the tests pass, the room nods. It is genuinely impressive, and it is genuinely easy. A capable model on an empty canvas with a human steering in real time is close to a solved problem.

That is not the problem you have.

Every services firm runs on the same hidden equation, whether or not anyone has written it down: revenue is a function of headcount. To grow, you hire. To take on more work, you staff up. To protect margin, you push utilization. The whole apparatus of an agency, a consultancy, or a professional-services firm is a machine for converting human time into invoices, and the better you run it, the more cleanly that conversion happens.

The change was forty lines. An AI agent wrote it to fix a proration edge case in the billing service — mid-cycle plan upgrades were rounding a day wrong. A senior engineer reviewed it the way good engineers do, line by line. The logic read cleanly. The variable names were sensible. The diff did what the ticket asked. She approved it, it merged, it shipped on a Thursday afternoon.

There is a quiet assumption buried in most enterprise AI adoption: that the path to better output is a better single prompt. Refine the wording, add more context, raise the stakes in the instruction, and the model will eventually produce the answer you wanted. That assumption is responsible for a lot of plateaued AI programs. It treats a probabilistic system as if it had one best answer waiting to be unlocked by the right key.