Use IA4AI
Investigations and conclusions
How a period of your AI use becomes findings, conclusions and one next step, and how you answer them.
An investigation is a full look at one period you choose: what your AI did, what it cost, what was found, and what to change. You get one per period you ask about, under Investigations in your workspace.
#Ask for one
Open Investigations, pick the period (UTC days, the way providers bill) and, if you like, say what to look at: why did Monday cost twice as much? or is the cache working for our nightly job?
- On the free plan, this sends a request to your analyst, who opens the investigation for you. It shows as Waiting for your analyst until it is ready.
- On the paid plan, it opens and runs at once, and you read it a moment later. It is labelled Automatic, so you can always tell it apart from one Reviewed by an analyst.
From code: POST /workspace/client-cases with days, or start and
end.
#How one is built
Each investigation runs as a fixed sequence of steps, each a plain, repeatable calculation on your usage records:
- Scope: the period and what was asked.
- Spend: totals, per model, per day, and unusual days.
- Agents: each agent's calls, cost, cache use and failures.
- Jobs: kinds of work, ranked by cost, and work done twice.
- Recommendations and plan: where the spend could be acted on, and cheaper models to test.
- Telemetry: what your agents' own telemetry reports, such as Claude Code's.
- Findings: everything above, as findings, each marked measured, estimated or untested.
No step uses a language model for arithmetic. An analyst can see each step's output, add notes, and re-run a step on today's data (the steps before it keep their results). Every step records the version of the code that produced it, so a later change to IA4AI can never silently change a past result.
An investigation an analyst runs stops for review before you see it. The analyst can leave a finding out (the reason is kept in the evidence log) or add a note to it that you see.
#Findings
Each finding says what was seen, what it costs, and how sure it is:
- Measured: read from records, such as the cost of a day or the share of input served from the cache.
- Estimated: worked out with a stated method, such as a cost at list price.
- Untested: a candidate, such as a cheaper model. It carries no saving until it is verified on your own calls.
You react to each one: accept, reject, or ask a question. Your analyst sees it and replies on the finding. A rejection never starts a new run, or any spend, by itself: the analyst decides what to re-run.
#Conclusions
At the top of each investigation are up to five conclusions in plain words, each with what to do, and the single most useful next step. They are written from the findings you can see, and checked before they are shown:
- every conclusion cites the findings it rests on, and those findings exist;
- every number in it appears in a finding it cites;
- a conclusion that rests on an estimate or an untested candidate says so.
If a check fails, the writer gets one more try with the problems listed; if it fails again, a template writes the conclusions from the same findings instead. Each set says which wrote it.
Who writes them. By default, a template on IA4AI's server, so nothing leaves it. A language model writes them only if you ask us to turn that on for your workspace. Then only the findings' sentences (numbers and short labels) are sent to it, never prompts, traces or keys; the model is DeepSeek's, called with IA4AI's own key, and it can be switched to one we host ourselves.
One chart, drawn from measured data, never by the model: spend per day with unusual days marked, or the most expensive tasks when the period is shorter than three active days.
When a finding is left out or put back, the conclusions are written again, so you never see conclusions that do not match the findings shown. A note or a reaction does not change them.
#Read them from your assistant
Your assistant can list your investigations and read their conclusions through the MCP connector:
list_investigations and
investigation_conclusions. It is told to relay the numbers
as given and not to recompute them.
#After a change
When a finding leads to a change (a cheaper model for one job, say), the outcome shows whether you switched,
whether the saving happened, and whether quality held, measured on your live traffic
(GET …/outcome).

