A pilot exists to replace opinions with numbers. Two weeks, one process, your documents and a small group of your people. At the end you know what an agent does with your work, and you have the figures to decide whether a full rollout is worth it.
Week one: the agent gets your material
- We go through your documents as they are — policies, templates, procedures, the spreadsheet nobody admits to maintaining — and structure what the agent needs to read.
- I build the search index over that material and wire up the employee or customer context the answers depend on.
- Together we write down the questions that must be answered correctly, with the right answer next to each one. This becomes the test suite the agent is measured against, and it is yours to keep.
- First answers get reviewed by whoever knows the real rules, and the gaps that always turn up in the documents get resolved.
Week two: your people use it
- The pilot group asks it real questions for a week — not test questions, the ones they would have asked a colleague. For an HR assistant in a company of a couple of hundred, 10 to 20 people is enough for the common questions to surface.
- I read the logs daily and fix what the logs show: a policy the agent could not find, a phrasing it misread, a question it should have escalated instead of answering.
- The test suite is re-run after every change, so nothing that already worked quietly breaks.
- We watch the numbers that matter: how many dialogs closed without a human, how often it fell back, what a dialog cost.
What I need from you
Documents in their current state — cleaning them up is part of the work. One person who can settle "what is the actual rule here?" in a five-minute conversation. The pilot group, and an hour of their manager's attention to get them started. Then the material for the case at hand: if we are parsing CVs, 30 to 50 real ones and the fields your ATS expects; if we are reading contracts, 20 to 50 typical ones from your archive and a note of the positions you consider normal.
What you have at the end
- A working agent on your data, running in the channel your people already use.
- Test results on your own questions — not my benchmarks, yours.
- The measured cost per unit of work on your real traffic, next to what that work costs today.
- A written plan of what a full rollout needs: which systems to touch, what stays manual and what it takes to run.
- Everything built during the pilot, including the index and the test suite, in your hands whatever you decide next.
What a pilot deliberately does not include
No writing into your production systems, no migrations, no touching payroll. The agent reads and answers; anything that changes your records stays a human decision until the process has proven itself. That keeps a pilot to two weeks instead of a quarter.
After the pilot
Three honest options. Run it managed, on my EU infrastructure, paying per unit of work. Move it in-house on your own hardware — see self-hosted. Or stop: you keep the artifacts, and you have paid for a decision made on evidence instead of a guess.
Book a call and we will pick the process worth two weeks.