A useful AI agent starts with one task.
You see a possible role for AI in your business. We examine a real task, available information and acceptable errors to decide what an agent can take on.
A specific use.
Document search
Check available sources, their quality and consultation rights.
Response preparation
Define what the agent proposes and what a person must approve.
Actions in your tools
Identify required access and situations where the agent must stop.
What you get from the brief.
A defined use case
Inputs, expected result and limits on the agent’s responsibilities.
A written specification of the task, naming the exact point where the agent hands back to a person.
An evaluation protocol
Representative examples and criteria for quality, time and cost.
A scored test set of fifty real examples, in a spreadsheet you keep and can run again in six months.
A feasibility assessment
Findings from a focused prototype and unknowns to resolve before delivery.
A report giving the measured pass rate and the cost per run, with the prototype code in a repository you own.
How long it takes.
Two to four weeks. Building the test set takes most of that; the prototype itself is often a few days. You see the first measured results in week two, and they are usually lower than anyone expected.
What we need from you.
Fifty real examples, with the answers
Past cases and what a competent colleague actually replied to each one. Without them, the assessment has nothing to measure against and turns into an opinion.
Someone who can say what a good answer looks like
One person who knows the work, two or three hours in total, to grade the cases we disagree on. Accuracy is a judgement before it is a number.
A decision on what the agent may touch
Read access is quick to grant. Writing into your tools is a separate conversation with whoever owns them, and opening it early avoids a three week wait at the end.