Skip to content

AI Remediation

Every failure ends with an owner and a date

See what happened, understand the likely cause, and build a practical plan to test the change.

Incident · inc_4a82Ready to review
Expected
Exchange the order
Actual
Refunded the order
Key moment
The user changed the request, the workflow kept the first intent
Context
Conversation · Trace · Tool calls · Outcome

The flow

From outcome to validated change

  • See the whole failure in one place

    The system can stay online, return a valid response, and still do the wrong job. Connect the conversation, state, retrieval, tools, timing, and final outcome.

  • Know where the path changed

    Trace the issue to a prompt, missing source, retrieval problem, stale state, tool choice, parameter, guardrail, or workflow step.

  • Move from explanation to a practical change

    Review suggested prompt, retrieval, Knowledge Base, guardrail, or workflow changes. Your team decides what to apply — and the incident, owner, and validation result are preserved.

Validate the change

The fix is not done until the situation passes

Run the same case again, compare the result, and keep the scenario for future releases.

Validation · SCN-1847Passed
Before
Refund started · Failed
After
Exchange started · Passed
Kept as
Regression scenario

The backlog

A finding with an owner, a fix and an expected number

This is the whole difference between monitoring and remediation. Every row below arrived from a cluster of failing sessions and left carrying a suggested change, a named team, a date, and the specific metric it is expected to move — so it can be argued with, scheduled, and checked afterwards.

Remediation backlog · top 6 of 14 openordered by severity

Every row names a team and a date and states the number it expects to move. Item 3 already shipped and hit its target — 7.1% timeout down to 2.2%. That is what separates a backlog from an alert feed.

Remediation backlog
#Issue and evidenceSevSuggested fixOwnerExpected impactStatus
1Fee-exception answers missing tier exception clause214 failed sessions · 4-turn avg loopHIGHUpdate KB article, add retrieval tags, revise answer templateKnowledge EngFee-exception reliability 63% → ~80%In progressJul 11
2Dispute escalation routed to generic help flow68 low-sat sessions · 167-question clusterHIGHAdjust workflow routing, add escalation branchProduct EngIntent reliability 68% → ~82%PlannedJul 18
3Account verification tool timeout7.1% timeout rate on verification spanHIGHRetry logic + timeout threshold updatePlatform EngTimeout 7.1% → 2.2% observedShippedValidated
4Policy refusal over-fires on benign fee questions146 false refusalsMEDRefine refusal criteria, validate against replay setAI EngRecover ~120 completions/monthPlannedJul 15
5Latency spikes on document lookupp95 rose to 5.4s in one intent clusterMEDCache top retrieved chunksInfra Engp95 5.4s → <3.5sPlannedJul 22

Each item links to its evidence set: the clustered sessions, the traces, the failed evaluations and the root-cause classification that generated it. When it ships, the validation result attaches to the same record — so the item history reads as an audit trail from detection to proof.

Figures from a sample FinVault reliability report. Design-partner data, synthetic and illustrative of real deliverables.

The gate

Nothing reaches production without human approval

4,050 failed sessions → 7 human approvals → 5 shipped changes. PRISM finds the failure and proposes the fix. A human decides what ships.

  1. 4,050

    Failures clustered

  2. 3,822

    Root causes classified

  3. 14

    Backlog items opened

  4. 9

    Fixes proposed

  5. 7

    Human approved

  6. 5

    Shipped & validated

Knowledge and guardrails

Review the change, the affected sessions, and the evidence before approving.

Prompts

Every prompt change is versioned, diffable, and reversible.

Code

PRISM drafts the change and the evidence. You open the pull request. Nothing merges without you.

The proof

Re-measured on the same cohort that found the problem

A fix is only Validated when the target metric improves in production over a minimum seven-day window, measured against the same cohort definition used to detect the failure. Anything that misses its primary target is marked partially validated and opens a follow-up automatically.

  • Knowledge · Jun 9

    Refund policy KB update

    Session success rate

    BEFORE61%
    AFTER79%

    Validated

  • Code · Jun 14

    Verification retry · PR #241

    Tool timeout rate

    BEFORE7.1%
    AFTER2.2%

    Validated

  • Prompt · Jun 21

    Dispute prompt v2.3.0

    Average turns to resolve

    BEFORE3.9
    AFTER2.4

    Partially validated

One item, opened

From a cluster of failing sessions to a merged pull request

The record is continuous. Detection, cause, proposed change, approval and validation are four states of one object rather than four tickets in four systems that nobody reconciles.

Detecteda cluster, not an alert
patternAccount verification tool timeout
sessions648 · 16% of all failures
span7.1% timeout rate on verification
Diagnosedcause, then fix layer
root causeWorkflow / tool failure
fix layerCode PR
proposedRetry logic + timeout threshold
Approvedby a named person
mechanismGitHub PR #241OPENED
approverPlatform Eng lead
mergedJun 14, after review
Validatedmeasured, not assumed
before7.1% timeout rate
after2.2% · 14-day span analysisVALIDATED
knock-onp95 latency 4.6s → 3.7s

Backlog item #3 — the verification timeout — from 7.1% to 2.2% and closed. The same record carries all four states.

The evidence never detaches

Open the item at any point and the clustered sessions, the traces, the failed evaluations and the classification that produced it are still one click away. That is what makes the recommendation arguable rather than something to accept.

The expected lift is stated before the work

“Timeout 7.1% → under 2.5%” was written when the item was opened, not after it shipped. A projection you record beforehand is the only kind that can be wrong, which is the only kind worth having.

A partial result opens the next item

The dispute-flow prompt change reduced turns but missed its resolution target. It was marked partially validated and generated the follow-up automatically — which is how the residual issue got identified as workflow logic rather than prompt copy.

Product boundary

PRISM supports root-cause investigation, failure explanation, remediation recommendations, a remediation backlog, before and after validation, and human review and approval. It does not generate autonomous pull requests, make unattended repository changes, merge automatically, deploy automatically, or perform autonomous production remediation.

Questions

Common questions

Does AI Remediation edit my code?

Not by default. It brings the failure and likely cause into one place and recommends a next action. Your team controls the change.

What kinds of changes can it suggest?

Recommendations may cover prompts, retrieval, Knowledge Base content, guardrails, workflow state, tool use, and the validation plan.

How is a failure validated?

Run the same scenario or representative trace after the change and compare the expected and actual result.

Can I keep a remediation history?

The product should preserve the incident, recommendation, owner, status, evidence, validation result, and relevant audit history according to plan and retention limits.

Is this only for production issues?

No. The same flow can begin from a synthetic scenario, an imported conversation, an evaluator result, or a live production trace.

Start from a failure you already have