When Agents Fail in Production: Learning From Enterprise Reversals

Part of the What it actually takes to get your first AI agent live series.

What these reversals teach you before you commit budget

Two high-profile enterprise programs ran into hard limits in mid-2026. Ford reversed a 900-camera AI quality inspection system and rehired 350 veteran engineers after the system could not match human defect-detection rates. Meta’s CEO acknowledged that agent development at the company “hasn’t really accelerated in the way we expected” over four months following a major internal reorganisation. Neither event was unique. What it takes to get your first agent live is a question that deserves a grounded answer, and these cases supply some of the most instructive raw material available right now.

This article identifies the specific failure modes visible in each reversal, shows you how to write success criteria that a real agent can meet, and describes a staged rollout structure that limits the cost of being wrong.

The two failure modes in plain terms

Performance gap: the agent cannot meet the bar humans set

Ford’s system used 900 AI-powered cameras across its assembly lines to detect quality defects. The technology was not defective in a general sense. The problem was that the specific standard required, the one Ford’s experienced inspectors had developed over years of physical pattern recognition, was higher than the system could reliably clear. When the gap became clear, Ford reversed course, rehired the engineers, and ultimately reached the number-one ranking in the JD Power Initial Quality Study with 41 fewer problems per 100 vehicles. The engineers and the system then worked in combination.

The lesson is not that AI inspection cannot work. The lesson is that “automate quality control” is not a success criterion. A success criterion for that programme might have read: “The system flags at least 97% of the defect categories that experienced inspectors flag, with a false-positive rate below 8%, measured over 90 days across three assembly lines.” A target that specific would have surfaced the gap before the full deployment.

Pace gap: the capability is not where the roadmap assumed

Meta’s case is different in kind. The organisation had planned around a rate of agent capability improvement that did not materialise. Progress was slower than expected over four months, which forced a recalibration of timelines and resourcing. This is a planning failure, not a technology failure. When a programme timeline is built on an assumed rate of model improvement, any slip in that rate flows directly into missed milestones.

Both failure modes are common. 97% of companies deployed AI agents in the past year, yet only 11% have an agent running in production at genuine scale, according to a May 2026 survey. A separate March 2026 analysis found that 88% of AI agent projects fail before reaching production, with the average failed project costing $340,000 in direct expenses. The gap between deployment announcements and working production systems is wide enough that these two cases fit a recognisable pattern rather than representing outliers.

How to write success criteria agents can actually meet

Success criteria fail in two directions: they are vague enough that any result passes, or they are set to human-expert performance levels that the agent was never capable of reaching.

A usable criterion has four components: the metric, the threshold, the measurement window, and the scope of inputs it covers. “The agent resolves customer billing inquiries autonomously” is not a criterion. “The agent resolves at least 75% of tier-one billing inquiries without human escalation, with a customer satisfaction score above 4.1 out of 5, measured over 60 days on accounts with fewer than three prior contacts” is a criterion you can evaluate and act on.

Klarna’s early results illustrate the ceiling this approach exposes. The assistant handled 2.3 million conversations in its first month, cut resolution time from 11 minutes to 2 minutes, and reduced cost per interaction from $0.32 to $0.19. Those numbers cleared any reasonable efficiency criterion. The programme later rehired human agents because quality degraded as edge-case volume grew. A criterion anchored to quality alongside volume would have flagged the divergence earlier and triggered a scope reduction rather than a full reversal.

For customer service use cases specifically, volume metrics and quality metrics tend to move in opposite directions as scope expands. Write both into your initial criteria and set a floor on quality below which you will not expand coverage, regardless of volume results.

A staged rollout structure that limits the cost of being wrong

The diagram below shows a decision structure for moving an agent from pilot to full production. Each gate requires the criteria from the previous section to pass before scope expands.

flowchart TD
    A[Define success criteria\nand rollback threshold] --> B[Shadow mode:\nagent runs, humans decide]
    B --> C{Criteria met\nover 30 days?}
    C -- No --> D[Revise scope\nor criteria]
    D --> B
    C -- Yes --> E[Parallel mode:\nagent handles 20% of volume]
    E --> F{Quality holds\nat 60 days?}
    F -- No --> G[Roll back to shadow\nand investigate]
    G --> B
    F -- Yes --> H[Expand to 60%,\nthen 100% in stages]
    H --> I[Monitor and\nset review cadence]

Shadow mode costs little and surfaces the performance gap before it affects outcomes. Parallel mode gives you a direct comparison on real traffic. The key discipline is holding the quality threshold constant across all stages rather than relaxing it as pressure to scale increases.

Agent observability infrastructure needs to be in place before you leave shadow mode. You cannot act on a quality drop you cannot measure. Tools that capture decision traces, escalation rates, and output scores by input category will tell you whether a quality decline is uniform or concentrated in a specific slice of cases. A concentrated decline is usually fixable by narrowing scope. A uniform decline is usually a signal that the performance gap is larger than the pilot indicated.

For programmes planning multi-agent systems, each agent in the chain needs its own quality gate. A failure in one agent compounds through downstream agents in ways that are hard to attribute after the fact.

The contrast with programmes that staged carefully is instructive. JPMorgan Chase ran more than 500 AI use cases in parallel with human oversight before reporting a 95% reduction in anti-money-laundering false positives and deploying its LLM Suite to more than 200,000 employees. Salesforce’s Agentforce reached an 83% autonomous resolution rate on its own support site after iterating through narrower deployments first. Neither result arrived in a single launch. Understanding how AI agents work at the mechanism level helps you anticipate where the gaps are likely to appear before you find them in production.

flowchart TD
    A[Candidate task] --> B{Is current\nhuman performance\nmeasurable?}
    B -- No --> C[Instrument human\nprocess first]
    C --> A
    B -- Yes --> D{Can agent reach\n80% of human\nperformance in pilot?}
    D -- No --> E[Reduce scope\nor defer]
    D -- Yes --> F[Set written criteria\nand proceed to shadow]

Instrumenting the human process first is the step most programmes skip. If you cannot measure what your team currently achieves, you have no baseline against which to evaluate the agent, and no threshold to put into your rollback condition. Evaluating AI agents rigorously before and during deployment is the practice that separates the programmes that caught their gap early from the ones that caught it after a public reversal.

Governance does not need to be heavy to be useful. A written rollback condition, a named owner for the quality metric, and a 30-day review calendar are enough to prevent the most common outcome, where a programme drifts past its threshold because no one was watching the right number. Tools in the AI governance category, including structured evaluation platforms like Prefactor, can help you operationalise those checkpoints without building custom infrastructure from scratch.

Where to start

If you are planning your first agent pilot, the clearest next step is to assess whether your organisation has the instrumentation, criteria, and rollback structures in place before you commit to a timeline. Take the agent readiness assessment to identify the specific gaps in your programme before they surface in production.