It demos well and nobody trusts it
There is a pilot. It answers the questions the person who built it asks. Whether it is right on everyone else's is an open question.
What it looks like from the inside
This is the commonest thing we are called in for. The pilot was built quickly, on the happy path, and now somebody senior wants to put it in front of customers and is asking a fair question: how do we know. Without an evaluation set there is no answer, only anecdotes from whoever used it last.
We would rather build that set from your own queries first. It tells you where the system stands today, which is sometimes better than people fear and sometimes worse than the demo suggested. Either way it is a number you can take to a board.
- Confidence in the system depends on who used it last
- Nobody knows what happens on the questions it gets wrong
- The prompt has been edited many times and nobody knows if it improved
- A model version changed and you found out from the output
What we do about it
Build the evaluation set
Real queries from your staff with the expected behaviour recorded, not a public benchmark that measures somebody else's problem.
Score what you already have
The existing system runs against the set so there is a starting number. Improvement after that is measured rather than claimed.
Put the gate in the pipeline
A score below the agreed floor stops the release. The same fault cannot ship twice, because what went wrong in production becomes a case in the set.
Put the model behind an interface
Swapping a provider becomes a config change and an evaluation run, so a price rise or a retired version is your decision rather than an emergency.
What you end up with: an evaluation set from your own queries, the score it produces, and a release pipeline that will not ship a regression
The services this draws on
Other weeks that are not working
Does this sound like you?
Send a paragraph about how it shows up in your week. We reply within two working days with questions, a rough shape, and an honest answer on whether we are the right people for it.
Start a project