Poisson LabsData workflows and pipelines
Know what evidence it takes to take a workflow off manual review.
An agent drafts answers, resolutions, flags and reroutes, and a person approves every one. Poisson says which action types a change made worse, and how many cases it would take to clear a threshold you choose.
Results come from a synthetic copy of a payments back office, calibrated to public comparables.
600 actions per version
1 of 4 action types got worse.
| Action type | Before → after | Verdict |
|---|---|---|
| RFI answers | 0.0% → 0.0% | no significant change |
| Break resolutions | 0.0% → 0.0% | no significant change |
| Late-payout flags | 0.7% → 0.0% | no significant change |
| Reroute drafts | 4.0% → 22.7% | worse |
01The threshold
What would it take to stop approving by hand?
Cases needed for the interval's lower bound to clear a threshold you choose, if the rate so far continued. "Not at this rate" means the agent has to get better first.
| Action type | Right so far | Past 90% | Past 95% | Past 98% | Past 99% |
|---|---|---|---|---|---|
| RFI answers | 94.7% | 159 cases | not at this rate | not at this rate | not at this rate |
| Break resolutions | 100.0% | already | already | 189 cases | 381 cases |
| Late-payout flags | 87.3% | not at this rate | not at this rate | not at this rate | not at this rate |
| Reroute drafts | 83.3% | not at this rate | not at this rate | not at this rate | not at this rate |
02Where a bad number enters
Wrong answers: 43 → 24, of 300.
Trace names the first step that carried a wrong figure, so the fix goes where the error entered.
| The agent | Before | After |
|---|---|---|
| did its own work, its query or its arithmetic, with no planted wrong figure read along the way | 33 | 18 |
| ran a query that left the cancelled payments out of the failed total | 10 | 6 |
03Where it runs
Inside your own infrastructure.
Poisson Cloud runs in your own cloud account. Keys, prompts and outputs stay on your side, and sending run records to us is a switch you control.