Blog · ← All posts
How Long Should an AI Agent Incident Investigation Take? Benchmarks With and Without a Ledger
By the Ingenious AI team · Published: November 14, 2026 · Reviewed: November 14, 2026 · 6 min read
The short answer: With a complete, verified action ledger, a typical single-agent incident investigation runs: detection in minutes (anomaly alerts), scoping in under an hour (interrogation plus downstream reconciliation), root-cause identification the same day (replay to the first divergence, converted into a regression case), and evidence production for customers or auditors in hours (the export bundle). Without a ledger, the same phases measure in days to weeks, dominated by reconstruction — assembling a partial timeline from chat exports, vendor dashboards, and memory — and the conclusions carry 'probable' where the ledger's carry citations.
Replay a recorded agent incident free → Step through a real recorded run in the replay view, or paste ledger records into the public verifier and get a VERIFIED/BROKEN verdict in seconds — no signup.
'How long will the investigation take?' is the first question leadership asks and the one teams answer worst, because the honest answer is 'it depends on the records' — which sounds like evasion and is actually the whole truth. Investigation time is not a property of the incident's severity; it is a property of the evidence's readiness. Two companies can suffer the identical failure and differ by a factor of twenty in time-to-truth, purely on recording.
This guide benchmarks the phases separately, with and without a ledger, so the dependency is visible — and so a team can predict, and then improve, its own numbers. The figures are composites from the patterns this product's incident literature describes; treat them as planning benchmarks, not guarantees.
What is time-to-truth? the elapsed time from an incident's detection to the moment the organization holds a verified account of what happened — complete action set, first divergence, bounded impact — sufficient to remediate and to answer external questions. Time-to-truth, not time-to-fix, is the metric records determine: fixes can be fast and wrong when truth is slow, and most incident cost accumulates in the gap between the two.
Phase 1: Detection — minutes versus the customer's phone call
Ledger path: the anomaly layer — bursts beyond three sigma, first-seen tools, after-hours spikes — fires during the incident, typically within minutes of the behavior's start; detection is a property of the system, and the alert arrives carrying the ledger range to inspect. Non-ledger path: detection waits for a symptom to surface in a human channel — a customer complaint, a reconciliation mismatch, a finance review — with a median measured in hours to days and a worst case of 'at the audit'. Every later phase inherits this head start or deficit, and the damage total is usually set here, because an undetected agent keeps acting.
Phase 2: Scoping — the enumeration hour versus the archaeology week
Scoping asks: exactly which actions, customers, and amounts did this touch? Ledger path: interrogate the action class over the window, reconcile against the downstream system, and the bounded statement exists within the hour — the first-hour playbook budgets twenty minutes for it. Non-ledger path: the same question launches archaeology across chat platforms, tool dashboards with divergent retention, and database forensics, consuming days and producing an estimate whose error bars nobody states. The difference is not effort; it is that one path queries a designed record and the other reconstructs an undesigned one. Scoping quality also caps every later decision: remediation, notification, and disclosure are all only as complete as the enumeration beneath them.
Phase 3: Root cause — the replay versus the narrative contest
Ledger path: walk the replay to the first divergence — a specific record with inputs and stated reasoning — classify the failure family, and the root cause is a citable fact the same day. Non-ledger path: root cause is negotiated. Engineers propose mechanisms, each consistent with the fragments; the prompt gets tightened on a theory; the incident closes with a plausible story whose confidence nobody quantifies. The measurable consequence is recurrence: divergence-based fixes carry regression cases built from the exact failing step, while narrative-based fixes carry hope, and hope's recurrence rate within the quarter is the industry's quiet scandal.
Phase 4: Evidence production — the export versus the correspondence course
When the incident has an external audience — the affected customer, their assessor, a regulator, an insurer — the investigation is not done until the evidence is produced in the audience's form. Ledger path: the bundle (scoped export, timeline, verification, certificate) assembles in hours from artifacts that already exist. Non-ledger path: production is a project — data gathered under deadline from systems that purge on their own schedules, formatted by hand, caveated by counsel — measured in weeks, with each exchange generating new questions the fragments cannot answer. This phase is where the non-ledger incident's cost concentrates, because external clocks (chargeback windows, regulatory deadlines, the customer's patience) do not negotiate.
The compound effect, and the honest exceptions
Multiplied across a year of incidents, the phases compound into a reliability reputation: the ledger team's postmortems cite records, its customer reports land inside the week, its recurrence rate falls because its fixes are aimed at located steps. The exceptions deserve naming: ledger investigations slow down when the incident is multi-agent with missing causal links (the graph has holes to work around), when the failure lives in a tool's internals the ledger can only see from outside, and when the records themselves are implicated — a chain break converts the investigation into two investigations. Those are the cases the completeness and integrity disciplines exist to keep rare; the benchmark assumes them, which is another way of saying the benchmark is earned monthly, not bought once.
Five actions you must take today
Your investigation speed next quarter is being decided by this quarter's habits:
- Measure your last incident's phases. Detection lag, scoping time, root-cause time, production time — written into the postmortem as numbers, so the benchmark becomes yours.
- Attack the detection lag first. If detection arrived via a customer, the anomaly layer is the fix — bursts, first-seen tools, after-hours — before any other investment.
- Time a scoping drill. Pick a random action class from last month and enumerate its complete set against a downstream witness; if it takes more than an hour, the path needs work in peacetime.
- Require divergence citations in postmortems. Root cause stated as a record, or stated as unknown — the discipline that keeps narrative contests from closing incidents.
- Pre-build the evidence bundle. Export shapes, timeline generation, verification attachment — assembled once, so production is a formality during the incident rather than a project after it.
Worked example: one failure, two clocks
The same defect — an agent misreading pasted text as a prior commitment and issuing unauthorized discounts — struck two fictional retailers in the same season. At ledger-equipped Callisto Home: the burst alert fired at minute 6 of the spree's second occurrence (the first, smaller one had already produced a postmortem and a watch); scoping enumerated 6 affected orders totaling $612 within the hour; replay found the divergence step the same afternoon; the regression case gated the fix; affected customers held timelines within two days. Total elapsed time-to-truth: under 24 hours; total engineering time on facts: about 5 hours.
At Northwind Traders, without a ledger: detection came from a customer's email on day 3; scoping consumed eleven days and never fully bounded (the concession strategy assumed the widest plausible set); root cause closed as 'prompt ambiguity' after a week of debate; the failure recurred in seven weeks. Time-to-truth, generously defined: never fully achieved — the file closed on estimates. Engineering time on facts alone: 34 hours, plus a $4,800 concession.
Severity identical, budgets similar, teams comparably skilled. The clock difference is the records difference, itemized — which is why investigation benchmarks belong in every logging decision, stated as plainly as storage costs.
FAQ
Are these benchmarks just best-case marketing?
They are achievable-case, with the assumptions named: complete recording, verified chain, causal links present, bundle pre-built. Strip an assumption and the corresponding phase degrades toward the non-ledger numbers — which is itself the useful reading, because it prices each discipline in hours. A team can locate itself on the spectrum honestly by timing its last incident against the phases; the benchmark's purpose is the self-measurement, not the advertisement.
What dominates investigation time in multi-agent incidents?
Join reconstruction: establishing which agent's output caused which action, across systems whose clocks and identities disagree. With caused_by links recorded at handoff, the graph walk replaces the join work and multi-agent investigations approach single-agent times for scoping and root cause. Without links, expect the archaeology multiplier to apply per agent boundary — the strongest argument for enforcing causal links before the fleet grows, not after.
How fast is fast enough for regulators and customers?
Faster than their clocks: chargeback representment windows, breach-notification deadlines (some measured in days from awareness), and contractual incident-notice periods (often 24–72 hours for a first substantive notice). The practical target is a bounded factual statement inside the notice window — possible only when scoping is an hour-scale activity — with root cause following inside the week. Investigations that cannot bound impact inside the window are forced into worst-case notifications, which are costlier in every currency.
Where do teams waste the most time even with a ledger?
In the hand-offs around it: waiting for access to the downstream witness's data, convening the people authorized to pause or to notify, and re-litigating the timeline in meetings the replay already settled. The records compress the evidence phases; the organizational phases compress only with pre-delegation — named incident authority, rehearsed playbooks, pre-drafted communications. Measure those too; the ledger makes their cost visible by removing the excuse that the facts were slow.
Should investigation time be an SLO?
As an internal reliability metric, yes: time-to-truth per incident, tracked across postmortems, with the phase split published to the team. It behaves like MTTR but measures the part records control, and it trends beautifully when the disciplines land — verification habits, coverage, causal links, pre-built bundles each shave a visible phase. What gets measured in the postmortem gets funded in the roadmap.
Start logging your agents free → The free tier gives you the full hash-chained ledger for 1 agent and 1,000 actions a month, with chain verification and replay included.