Quick answerRun every disputed SLA credit claim through a fixed adjudication sequence before anyone approves a payout: first check whether the agent was reachable and responding (an infrastructure question, answered by uptime logs), then check whether a reachable agent gave a wrong or harmful answer (an agent-reasoning question, answered by replaying the exact prompt and tool calls it had). Most contracts already define credits for the first category; most do not clearly say who is on the hook for the second. Build a named adjudication owner, a fixed evidence checklist, and a default resolution rule for genuinely mixed-cause incidents, so a payout decision does not turn into a weeks-long argument with a customer holding an invoice.
The gap this closes
Most AI agent SLA contracts, and the internal question of what actually counts as downtime in the first place, are already well specified. Separately, what accuracy and credit terms to offer customers is a contract-design question most vendors have now worked through: service credits, not cash refunds, tied to measurable thresholds. What is missing in both of those is the operational step in between: once a customer files a claim, who actually decides whether this specific incident is the kind the contract covers, and how fast do they decide it.
That gap matters because the two credit categories usually have different owners inside your own company. Availability credits are an infrastructure question your platform or SRE team can answer from monitoring data in minutes. Accuracy or behavior credits are a product and applied-AI question that requires reconstructing what the agent actually saw and did, which can take hours or days, and which the SRE team has no authority to sign off on alone.
Build the adjudication sequence before the first disputed claim, not during it
- Triage on reachability first. If the agent was down, unresponsive, or timing out, this is an availability incident. Pull it from uptime monitoring, apply the contracted credit, and close it. No further adjudication needed.
- If the agent was up, reconstruct what it saw. For a disputed "wrong answer" or "bad action" claim, replay the same query against the exact source data and tool responses available at the time. This is the same reconstruction step used in hallucination-versus-bug triage, and it is worth reusing that same runbook here rather than building a second one.
- Classify the confirmed cause into one of three buckets, not two: agent reasoning error (the agent had correct information and used it wrong), upstream data or tool error (the agent correctly used information that was itself wrong), or ambiguous mixed cause. Only the first two have a clean, contract-defined owner in most current SLA language.
- For mixed-cause incidents, apply a default rule instead of negotiating case by case. A reasonable default: if the agent's own output materially amplified or failed to catch an upstream error it had the tools to check, treat it as an agent-side failure for credit purposes. Document this default in the contract itself, not just in an internal runbook, so a customer's legal team can see the rule before a dispute happens, not after.
- Name a single adjudication owner with authority to close a claim, typically a customer success or support operations lead with a documented escalation path to product and legal for genuinely novel cases. A claim that sits unassigned while three teams point at each other is worse for the relationship than a wrong-but-fast decision followed by a correction.
Why this is not the same question as what terms to offer
It is tempting to think a well-written contract makes this process unnecessary. It does not, because contract language describes categories and a real incident has to be sorted into one of them by a person looking at logs, often under time pressure from a customer who does not care which internal team is responsible. The credit-terms question and the adjudication-process question are genuinely separate problems, and skipping the second one means even a well-designed credit policy sits unused the first time a claim is actually disputed.
FAQ
Does the customer need to know this internal process exists? The default resolution rule for mixed-cause incidents should be visible in the contract or SLA exhibit, since it directly affects what the customer is owed. The internal triage steps and team assignments do not need to be customer-facing.
What if the reconstruction step itself is inconclusive? Set a maximum adjudication window, for example five business days, after which an inconclusive case defaults in the customer's favor for the disputed credit. An open-ended "we're still investigating" response erodes trust faster than a modest credit paid on an ambiguous call.
Should engineering or support own the final sign-off? Support or customer success should own the claim end to end and pull in engineering only for the technical reconstruction step. Routing ownership through engineering tends to produce accurate but slow decisions, since credit adjudication is not their core job and will always lose priority to production issues.