What this article covers
Most agent pilots produce something useful in a controlled setting. The harder question is whether that agent is ready to operate autonomously at enterprise scale, where the cost of a wrong action compounds across thousands of decisions. This article covers the maturity indicators worth tracking, the checkpoints to clear before production deployment, and how to instrument pilots so you can identify what separates agents that scale from ones that stall.
The gap between pilot and production is not a technology problem
According to data published in March 2026, 78% of enterprises have AI agent pilots running, but fewer than 15% reach production, and only 14% have successfully scaled. That gap is not primarily caused by model quality. It is caused by the absence of structured evaluation: teams have no agreed criteria for what "ready" means, so deployment decisions default to gut feel or executive pressure.
The production environment differs from a pilot in three concrete ways. The action space is larger, meaning the agent encounters inputs it was never tested against. The consequences are real, meaning errors affect customers, systems, or records. And human oversight is thinner, meaning the agent has to be correct more often, with less correction.
Evaluating readiness means measuring whether the agent can handle those conditions before you expose it to them, not after.
Maturity indicators to track across pilots
Before comparing pilots, you need a consistent measurement vocabulary. Four indicators are load-bearing.
Task completion rate under distribution shift. Measure how often the agent completes its assigned task correctly when the input is slightly outside the training distribution. A pilot that scores 95% on curated test cases and 60% on novel edge cases is not ready for production, regardless of the headline number.
Escalation calibration. An agent that never escalates is not confident; it is uncalibrated. Track whether escalations happen at the right moments: when the agent is genuinely uncertain, when the action is irreversible, or when the stakes exceed a defined threshold. CalPrep's calendar agent, launched in September 2026, made human confirmation mandatory before any calendar action. That is a deliberate escalation policy, not a fallback, and it reflects a conscious decision about where autonomous action is appropriate and where it is not.
Step-level accuracy, not just outcome accuracy. An agent that reaches the right answer through a flawed reasoning chain will fail on harder inputs. Measure accuracy at each step in the task, not only at the final output. This is covered in more detail in step-level evaluation: measuring agent actions before failure compounds.
Behavioral consistency. Run the same task multiple times with minor input variations and measure output variance. High variance signals that the agent's behavior is sensitive to prompt phrasing rather than grounded in a stable policy.
Evaluation checkpoints before production
These are the four gates a pilot should clear before autonomous deployment is considered.
Gate 1: Defined action schema. Every action the agent can take should be enumerated, with expected inputs, outputs, and side effects documented. If you cannot describe the full action space, you cannot evaluate coverage. Activity schema validation gives you a baseline to detect when behavior drifts from the intended envelope.
Gate 2: Failure mode inventory. Run deliberate adversarial tests: ambiguous instructions, missing context, contradictory tool outputs, and out-of-scope requests. Document how the agent fails, not just whether it passes. eSentire's threat analysis agent, deployed in mid-2026, compresses expert threat analysis from five hours to seven minutes and achieves 95% alignment with senior security analysts. That alignment figure is meaningful only because eSentire had a reference standard to measure against. Without one, 95% is an unanchored claim.
Gate 3: Scored quality and risk baseline. Before production, establish a scored baseline across a representative sample of tasks. The baseline should include quality scores (task completion, step accuracy, output correctness) and risk scores (rate of high-stakes actions, escalation frequency, out-of-policy behavior). Prefactor records spans at the action level and scores both quality and risk against this baseline, so deviations in production are visible against something concrete rather than a subjective standard.
Gate 4: Human review at defined thresholds. Specify which action types require human approval before execution, and which can proceed autonomously. This is not a binary choice between full automation and full oversight. It is a matrix of action types crossed with confidence levels. Our earlier guide on agent behavior validation covers how to define those thresholds before deployment rather than after an incident.
Instrumenting pilots to understand what scales
Comparing pilots without shared instrumentation produces anecdotes, not evidence. You cannot identify what makes one agent scale and another stall if each pilot is measured differently.
Instrument every pilot with the same span structure: record the trigger, each tool call with its inputs and outputs, any escalation events, and the final resolution. This lets you compare pilots on identical dimensions. Observability of agent workflows is the foundation, but observability alone is not evaluation, as we have written about separately in evals vs. observability.
The pilots that tend to scale share three structural properties. First, they have a narrow, well-defined task scope. Thomson Reuters' CoCounsel platform, serving one million users across 107 countries as of April 2026, is built for legal document analysis and research, not general assistance. The scope constraint is a feature, not a limitation. Second, scalable pilots have reference data for quality measurement, a corpus of known-good outputs the agent's production behavior can be compared against. Third, they have an escalation policy defined in advance, not improvised when something goes wrong.
L'Oréal's conversational analytics agent, operating at 99.9% accuracy across 44,000 monthly users as of June 2026, demonstrates what structured scoping produces. The agent handles data queries, not open-ended analysis. That boundary is what makes the accuracy figure defensible and reproducible.
According to reporting published September 2026, 80% of Fortune 500 companies have adopted agentic AI but face critical gaps when scaling from pilots to production. The gap is not unique to any sector. It reflects a common pattern: adoption decisions outpace evaluation infrastructure.
For teams running autonomous background agents or multi-agent orchestrators, the evaluation burden is higher because the action space is larger and human oversight is harder to insert at runtime. Both agent types benefit from tighter pre-deployment gate criteria, not looser ones.
Where to start
Pick one pilot that is close to production consideration and run it through the four gates above. Document what you cannot answer, because those gaps are exactly what your evaluation infrastructure needs to close. Start evaluating your agents or read the docs to see how span recording and scoring fit into the process.