Why do so many agent pilots fail to reach production?

Most pilots are scoped to prove a capability, not to survive an operational environment. They skip the governance, observability, and integration layers that production requires, so when the pilot expands, those gaps surface as blocking problems rather than addressable risks.

What is the right team size to start scaling an agent beyond pilot?

There is no universal answer, but teams that scale successfully tend to start with a single agent owned by one cross-functional group, covering engineering, operations, and a business stakeholder. Adding agents before the first one has stable observability and a clear escalation path compounds failure risk.

How do you measure whether an agent is ready to move from pilot to production?

Track error rate, task completion rate, and escalation frequency over at least two weeks at realistic load. If any of those metrics are unstable or unmonitored, the agent is not ready. Production readiness also requires a documented rollback procedure and defined human-in-the-loop triggers.

Should adoption of AI agents be mandatory or voluntary inside an enterprise?

The JPMorgan Chase LLM Suite case suggests voluntary adoption with strong internal tooling can reach scale without mandates. Mandatory rollouts that skip readiness checks tend to produce high nominal adoption figures alongside poor task completion rates and low trust from the people using the tools.

Why agent pilots stall before production and how to fix it

Agent pilot-to-production scaling fails at four consistent points: governance gaps, missing observability, integration problems under real load, and adoption strategies that destroy trust before the agent earns it. Organisations that do ship share one habit: they treat production readiness as a separate engineering problem from pilot success, and they instrument before they expand.

The gap is larger than the numbers suggest

88% of AI agent pilots never reach production, according to a 2026 survey of enterprise deployments. That figure sits alongside a separate finding: only 11% of enterprises have an agent running in production at genuine scale. Both numbers point to the same structural problem. Pilots succeed at demonstrating capability in controlled conditions. They fail to account for what production actually demands.

The controlled conditions of a pilot typically mean: a narrow task scope, a small and cooperative user group, low volume, and someone on the AI team available to intervene when the agent gets confused. Remove any of those and the agent’s behaviour changes, sometimes in ways that erode the trust the pilot spent months building.

The organisations that close this gap share two characteristics. They treat production readiness as a distinct engineering problem from pilot success. And they instrument the agent before they expand it, not after.

Where the transition breaks

The breakdown happens at four predictable points.

flowchart TD
    A[Pilot succeeds in controlled scope] --> B{Governance defined?}
    B -- No --> C[Escalation gaps surface at scale]
    B -- Yes --> D{Observability in place?}
    D -- No --> E[Errors invisible until user complaints]
    D -- Yes --> F{Integration tested at load?}
    F -- No --> G[Downstream systems fail or throttle]
    F -- Yes --> H{Adoption strategy defined?}
    H -- No --> I[Low uptake or forced rollout backfire]
    H -- Yes --> J[Production scale achieved]

Governance gaps. A pilot can operate without a formal escalation path because a human expert is usually nearby. At scale, that is not true. Without defined triggers for when the agent hands off to a human, and without a clear owner for those handoffs, errors compound. The AI governance framework you need for production is not the same as the ethical principles document your legal team approved.

Missing observability. Agents fail silently in ways that chatbots do not. A chatbot returns a wrong answer; you see it. An agent takes a wrong intermediate step and produces a plausible-looking output that is subtly incorrect. Without agent observability instrumented from the start, you find out about these failures through user complaints weeks after they begin. L’Oreal’s conversational analytics agents improved accuracy from 90% to 99.9% across 44,000 monthly users producing 2.5 million messages per month, but that improvement required continuous measurement, not a one-time evaluation.

Integration at load. Pilots typically hit one or two downstream systems at low volume. Production agents hit more systems, more often, and often in parallel. Kroger’s AI Shopping Assistant reached 10 million digital shoppers across 700 stores via a partnership model that allowed load testing across real inventory systems before full rollout. Teams that skip this step discover rate limits, authentication timeouts, and data consistency problems only after users encounter them.

Adoption strategy. Forced adoption of an agent that is not ready destroys trust faster than no agent at all. JPMorgan Chase’s approach to their LLM Suite deployment is instructive here. By treating the tool as an optional research aid rather than a mandated workflow, they allowed organic adoption to surface the real use cases and the real failure modes before either became locked in at scale. That stands in contrast to enterprise programmes that mandate adoption and then spend months managing the reputational fallout from a tool users do not trust.

What production-ready scaling looks like

The organisations that do make it through share a recognisable pattern. They do not scale the agent; they scale the conditions the agent needs to operate reliably.

flowchart TD
    A[Define production success criteria] --> B[Instrument observability before expansion]
    B --> C[Test integrations at 3x expected load]
    C --> D[Document escalation and rollback paths]
    D --> E[Run voluntary adoption with feedback loop]
    E --> F[Expand scope only after criteria met]
    F --> G[Review governance at each scope increase]

Doctolib replaced their legacy testing infrastructure with AI agents and ships features 40% faster, but the speed gain came after they rebuilt the infrastructure around the agent, not by dropping the agent into the existing pipeline. Novo Nordisk reduced clinical documentation time by more than 90%, compressing work that took weeks into minutes, across 25,000 employees. Both cases involved significant pre-production investment in the systems surrounding the agent.

The agentic AI architecture decisions you make during a pilot are usually provisional. Production requires you to revisit them with volume, failure rates, and compliance requirements in hand.

The governance layer most teams skip

Runtime governance is the part that is most consistently missing from pilot-to-production plans. Pre-deployment review covers what the agent is allowed to do. Runtime governance covers what happens when it does something unexpected. These are different problems requiring different tooling. The distinction matters enough that it is worth reading the comparison of runtime governance versus pre-deployment review before you finalise your production architecture.

For teams building the governance layer from scratch, tools like Prefactor sit in this runtime governance category, sitting between your agent and its downstream actions to catch out-of-bounds behaviour before it completes.

Voluntary adoption as a quality signal

The JPMorgan Chase case is worth examining beyond the headline. Voluntary adoption is not just a cultural preference; it is a quality signal. If users do not adopt an agent when they have a choice, that tells you something specific: either the agent does not reliably complete the task, or it creates more work than it saves. Mandatory adoption suppresses that signal. You get nominal usage figures and hidden dissatisfaction. Voluntary adoption with a short feedback loop surfaces the problems while they are still small enough to fix. This approach fits within a broader AI governance strategy that treats adoption data as a governance input, not just a KPI.

Scope expansion as a governance event

Every time you expand an agent’s scope, whether that means new users, new tools, or new decision authority, you have created a new governance event. The agentic AI design patterns that work at pilot scale often need revision at production scale. Multi-agent systems in particular introduce coordination failure modes that single-agent pilots do not expose. If you are moving toward multi-agent systems, treat each new agent added to the network as a scope expansion requiring its own readiness check.

The MIT Technology Review research found that 80% of Fortune 500 companies have adopted agentic AI but face critical scaling gaps. The gap is not capability. It is the operational infrastructure surrounding the capability.

Where to start

Before expanding any pilot, complete a structured readiness review that covers observability, governance triggers, integration load testing, and your rollback path. Take the agent readiness assessment to identify which of these your current programme has in place and where the gaps are most likely to surface.