Keelway
Platform · Trust & Control

Every autonomous AI decision, logged and reviewable.

AI that runs in operations only works if the operator trusts it. Trust comes from being able to see what the AI decided, why it decided it, and what to do when it gets one wrong. Keelway's agent log captures every classification, every drafted reply, every rate-con extraction, every remittance match — with one-click Good / Bad / Undo per row.

Every
AI decision logged
Classify · Pre-filter · Drafts · Rate cons · Remittances
Good/Bad/Undo
Feedback loop per row
Closes back into model tuning
Filterable
By agent, state, time, mailbox
Friday quality review in 15 min
Capture

Input, output, confidence, timestamp

Each agent decision: the inbound email or remittance, the AI's classification or extraction, the model's confidence score, the timestamp. Operator can drill to the source and the resulting state.
Feedback

Three buttons that matter

Good reinforces correct behavior. Bad marks the decision for tuning review. Undo immediately reverses the AI's action — no engineering ticket required.
Filter

By type, state, time, mailbox

Show only classifications. Show only marked-Bad entries. Show only the last 24 hours. Show only the dispatch@ mailbox. The operator's Friday quality review runs off these filters.
Tuning

Feedback closes back into the model

Marked-Bad entries feed an internal review queue. Engineering uses the feedback to update prompts and classifiers. The AI improves on the carrier's actual data over time, not on a generic benchmark.

Frequently asked questions

What does the agent log capture?+

Every autonomous AI decision the platform makes — email classifications (priority tier and category), drafted replies (the proposed text), rate-con extractions (the parsed load record), remittance extractions (the proposed invoice match), pre-filter decisions (the noise the AI suppressed). Each entry carries the input, the output, the model confidence, and the timestamp.

What's the feedback loop?+

Three buttons per row: Good (the AI got it right; reinforce), Bad (the AI got it wrong; mark for tuning), Undo (revert the AI's action). The Good/Bad signal feeds an internal review queue that drives prompt and classifier updates. Undo lets the operator immediately reverse an AI mistake without waiting for a fix.

Can I filter the log?+

Yes. By agent type (Classify, Pre-filter, Drafts, Status replies, Rate cons, Remittances), by feedback state (no feedback yet, marked Good, marked Bad, Undone), by time window, by mailbox account. Useful for periodic AI quality reviews — typically a 15-minute Friday cadence at most carriers.

Why expose this to the operator?+

Trust. AI in operations only works when the operator can audit it. The agent log is the explainability layer — every decision is visible, the reasoning is traceable, and the feedback closes the loop without requiring an engineering ticket. Carriers that deploy AI without this transparency lose trust and turn the AI off; the agent log keeps it on.

AI making decisions you can't audit?

Every decision logged. Feedback closes the loop.

Talk to sales

Related