Correction Capture
When a person fixes something the AI got wrong, that fix is saved and studied — so the same mistake gets engineered out instead of being corrected forever.
Think of a family recipe card with notes in the margin — less salt, ten more minutes in the oven. Corrections only help if they're written down where the next cook will see them. Correction capture writes down every fix people make to AI output, in a place where it actually changes the system.
The problem, in plain words
A teammate reads the agent's draft, sighs, rewrites two paragraphs, and hits send. Fixed — for that one ticket. Next week the agent makes the exact same mistake, and someone else fixes it again, a little more annoyed this time. The knowledge of what keeps going wrong lives in people's heads and in closed tickets nobody will ever reread. Your team is paying the cost of every correction and collecting none of the value. Whether the system needs fixing once a day or fifty times a day, nobody actually knows — it's an impression, not a number.
What we set up
Every time a person changes what the agent produced — edits the text, changes the action, reroutes it to someone else, or rejects it outright — the system captures the before, the after, and optionally the why, along with the original trace (the recorded play-by-play of what the agent did). Corrections that look alike are clustered into patterns, using the same forensics pipeline we use for bad conversations. Each cluster gets a recommended remedy: add an eval case (a saved test example), update the knowledge base, change the prompt (the agent's instructions), or change a tool. Then we track whether each cluster actually gets closed over time.
How it works, step by step
- A person fixes AI output
An edit, a changed action, a reroute, or a flat rejection — the normal work your team already does.
- The fix is captured automatically
Original version, corrected version, optional reason, and the full trace — with no extra paperwork for the person fixing it.
- Similar fixes are grouped
Twenty scattered corrections turn out to be four patterns. Now the problem has a shape.
- Each pattern gets a remedy
A new test case, a knowledge update, a prompt change, or a tool change — a concrete recommendation, not a vague to-do.
- Closure is tracked
Did the remedy ship? Did that kind of correction stop showing up? Both are visible, month over month.
What changes for you
Before, corrections vanished into individual tickets, and the same fix got made over and over by different people. After, every correction becomes an improvement candidate, and the correction rate itself becomes a number you watch: how often does the system need a human fix, and is that number going down? What it won't do: it won't apply the remedies on its own. A person still decides which fixes ship — the capture just guarantees that no correction is wasted.