Seven in ten can't name the agent that failed

Enterprises are spending on monitoring that says 'something is wrong' — and failing to build the one capability that says 'which agent did it.' Kore.ai's 2026 survey of 408 enterprises running agents in production found 82% report agents autonomously executing consequential actions without sufficient human oversight, 79% of those actions required manual reversal (93% of reversals rated costly and disruptive), and — most striking — 70% of enterprises can detect a failure but cannot determine which agent in a multi-agent environment caused it. Observability and attribution are different capabilities, and most enterprises invested in only one. Without identity-bound action evidence, containment, reversal, and accountability are all impossible.

The data: we know it broke, we don't know who broke it

The Kore.ai Agent Productivity Index 2026, fielded by Propeller Insights with 408 IT and engineering leaders, all actively running agents in production (US-based, 1,001-5,000 employees, 100% completion; vendor-commissioned and self-reported — see the source note below), makes four points that form the core evidence of this article.

First, autonomous execution is the norm, not a pilot. 82% of enterprises report agents autonomously executing consequential actions without sufficient human oversight — financial transactions, approval/denial decisions, data migrations and system updates. Second, reversal is both common and expensive. 79% of consequential autonomous actions ultimately required manual reversal, and 93% of those reversals were rated 'costly and disruptive' — the corrective act itself compounded the original failure. Third, detection is slow. Only 50% detect a malfunction within 1-4 hours, 33% need 4-8 hours, and 15% rely primarily on end users to report problems — the customer often knows before the company does. Fourth, attribution is missing. 70% of enterprises can identify that a failure occurred but cannot determine which agent in a multi-agent environment caused it. (Sources: Kore.ai report page and the PDF)

The consequences land on the P&L: 42% of enterprises suffered measurable revenue loss directly from an agent failure, 31% reported SLA violations attributed to agent malfunctions, and 28% experienced customer churn. Yet most of this is still classified as an IT incident — 54% of enterprises treat agent failures as IT problems only, so the risk never reaches the boardroom.

Observability is not attribution: two confused capabilities

What makes this survey worth a board's attention is the distinction it draws: observability answers 'is the system broken'; attribution answers 'which agent did it.' The Kore.ai conclusion puts it directly: 'Most organizations know when something has gone wrong, but seven in ten cannot identify which agent in a multi-agent environment caused it. Observability and attribution are different capabilities, and the data shows enterprises have invested in one without building the other.'

Why is this distinction fatal? Because attribution is the precondition for containment, reversal, and accountability. You can detect that an agent misbehaved, but if you cannot say which one, which tools it called, and when it crossed the line, you cannot precisely stop it, cannot decide which results to roll back, and cannot define the responsibility boundary for auditors. The 62% of enterprises in the survey that delayed deployment over governance and observability concerns are aiming at the wrong fix — the problem is not 'not seeing enough,' it is that no amount of seeing can assemble an accountable identity.

Our judgment: attribution needs identity binding at the action boundary, not log archaeology

Our judgment: treating observability as attribution is the most expensive mistake in agent governance; attribution must be generated at the action boundary with identity binding, not reconstructed afterward from logs. Here is the counterargument, so it can be tested: 'more logging and better monitoring dashboards will solve this.' The Kore.ai data says that path is closed — enterprises did invest in monitoring (over 92% report increased AI/bot traffic against APIs, most have dashboards); what they lack is the layer that binds an action chain to a specific agent identity. Post-hoc reconstruction is always slower than the event, and it is corrupted by agents rewriting their own records, context compaction, and incomplete logs.

Three implications follow. First, register identity at the action boundary: every agent instance gets a verifiable identity before its first action, and every subsequent action carries that identity — the physical precondition for attribution, which naming conventions or log fields cannot provide. Second, make action evidence tamper-evident: tool calls, parameters, accessed objects, return values, and approval records are hash-locked at the instant they happen, turning 'reconstruct later' into 'locked at the moment.' Third, automate attribution: on anomaly, the system returns a verifiable 'agent X performed action Y at time T' conclusion instead of making engineers piece together multiple log systems by hand — 93% of reversals are costly precisely because the assembly is too slow.

What it means for buyers

AI App leads: put 'attribution' into your evaluation criteria for agent platforms and observability tools — not 'how many alerts can it emit,' but 'can it automatically return the complete, tamper-evident action chain of the responsible agent on failure?' In your procurement matrix, attribution belongs alongside monitoring, not as a feature of it.

CISOs and compliance leads: when 54% of enterprises treat agent failures as IT incidents, liability and contractual risk are being underestimated. Require agent governance solutions to provide identity-bound action audit records — this is the evidence base for audits, insurance, and regulatory response. Without identity binding, logs are data, not evidence, in any accountability framework.

C-suite: 79% of consequential autonomous actions require manual reversal and 42% suffer revenue loss — agent failure is a P&L event. Add 'multi-agent attribution time' to your governance KPIs: from failure to identifying the responsible agent should be minutes, not hours or 'we can't tell.'

Your action plan

Three things this week. One, run an attribution drill: pick a random production agent anomaly and see whether your team can state 'which agent, which tool, which moment' within 30 minutes — if it cannot, your observability has not become attribution. Two, inventory identity binding: check whether every production agent's action logs carry a verifiable identity and are tamper-evident; log the gaps for remediation. Three, update procurement: add 'automatic attribution to a specific agent + tamper-evident action chain' as a scored dimension, on par with monitoring, not subordinate to it. A system that can hand you a verifiable chain of the responsible agent's actions within minutes of a failure is what your production environment deserves.

References: Kore.ai Agent Productivity Index 2026 (report page) · Report PDF. This survey was commissioned by Kore.ai and fielded by Propeller Insights (408 US IT/engineering leaders, 1,001-5,000 employees, all running agents in production). It is vendor-commissioned with self-reported figures, not independent third-party research — calibrate citations accordingly.