Moving an AI pilot to production requires more than better model output. Evidence from McKinsey, the US Census Bureau, and UK ONS shows that most organizations remain in experimentation, piloting, or limited scaling. Successful progression requires a defined workflow, baseline metrics, production data, human-validation rules, accountable ownership, user adoption, and a measured scale-or-stop decision.
Operation
Moving an AI pilot to production requires more than better model output. Evidence from McKinsey, the US Census Bureau, and UK ONS shows that most organizations remain in experimentation, piloting, or limited scaling. Successful progression requires a defined workflow, baseline metrics, production data, human-validation rules, accountable ownership, user adoption, and a measured scale-or-stop decision.
Key findings
• McKinsey’s 2025 survey classified 32% of AI-using organizations as experimenting, 31% as piloting, 30% as scaling, and 7% as fully scaled.
• AI high performers were more likely to aim for business transformation, redesign workflows, deploy in more functions, define human-validation processes, and show senior-leader ownership.
• US Census research found 57% of adopting firms used AI in three or fewer functions, and 65% limited worker-task use to three or fewer tasks.
• UK ONS found only 10% of AI-using businesses reported extensive use in June 2026.
• Breadth is not the first goal. A narrow, controlled workflow with measurable results is a stronger foundation than several unowned pilots.
• Each stage should have entry evidence, production controls, and an explicit decision gate.
Methodology and definitions
This playbook translates observational adoption research into a conservative operating sequence. The sources identify patterns associated with depth and value, but they do not prove that following the sequence guarantees success.
McKinsey’s 2025 State of AI report surveyed 1,993 participants between June 25 and July 29, 2025. The Census Bureau’s AI diffusion working paper uses nationally representative US employer-firm data from the November 2025 to January 2026 reference period. ONS provides official UK evidence about technology count and extensive use.
For this article:
• Experiment tests capability with examples or synthetic data.
• Pilot tests a bounded workflow with real users or real cases under heightened supervision.
• Production means the workflow handles eligible live cases under defined controls, ownership, and service expectations.
• Scaling expands eligible volume, users, actions, or functions while maintaining performance.
• Fully scaled means the system is deployed and integrated across the intended organizational scope, not necessarily every department.
Production is a control state, not a launch date.
What adoption research says about the gap
In McKinsey’s phase classification, only 7% of AI-using organizations were fully scaled. Thirty percent were scaling, 31% piloting, and 32% experimenting. The distribution shows that access and trial activity had become common while complete integration remained rare.
The report also identified a small high-performer group: 109 of 1,993 respondents said AI produced significant value and more than 5% of enterprise EBIT. Half of these high performers intended to use AI for transformative change, compared with 14% of other respondents. They were three times more likely to strongly agree that senior leaders demonstrated ownership and commitment.
High performers were also more likely to have defined processes for deciding when model outputs needed human validation. That finding is operationally important. Human review is not merely a temporary workaround for weak models; it is part of how organizations control consequential outputs.
These differences are correlations. Larger budgets, stronger data, better management, and more favorable workflows may contribute to both practices and outcomes. The evidence supports testing these patterns, not promising their return.
The Census paper adds a representative denominator. Eighteen percent of firms used AI in a business function in the reference period, rising to 32% when weighted by employment. Among adopters, 57% used AI in three or fewer functions. At the worker-task layer, 65% used AI for three or fewer tasks.
The study found a positive correlation between commercial performance and breadth of AI integration, operational investment, and worker-task use. It does not show that adding more functions mechanically improves performance. A firm may broaden AI because it is already performing well.
ONS reported in July 2026 that 35% of UK businesses with at least 10 employees used one or more AI technologies. Only 10% of adopters called their use extensive. The average number of technology categories per adopter increased modestly from about 1.4 in 2023 to 1.6 in 2026.
The consistent message is that the hard transition is from qualifying use to embedded operation.
Gate 1: Choose a workflow, not a tool
A production candidate needs a stable unit of work. Document:
• Trigger and eligible case.
• Inputs and source systems.
• Decisions the workflow makes.
• Output and destination.
• Human owner.
• Current handling time, volume, quality, and cost.
• Exceptions and prohibited actions.
“Use an AI agent for sales” is not a workflow. “Classify inbound demo requests, enrich the company, draft a response, and create a CRM task while a rep approves the message” is.
The first candidate should have repeated volume, observable output quality, recoverable errors, and limited permissions. Avoid a workflow whose success depends on unstructured judgment no one can score.
Gate: Do not begin a pilot until the manual baseline and owner exist.
Gate 2: Validate capability on representative cases
Use historical cases that represent routine work, edge cases, missing data, conflicting inputs, and high-risk exceptions. Separate:
• Correct output without revision.
• Usable output after review.
• Incorrect output caught by review.
• Incorrect output that would escape review.
• Refusal or failure to complete.
Measure the full workflow, including retrieval, tool calls, formatting, and system updates. A model evaluation alone misses integration errors.
Set thresholds before reviewing results. Examples include minimum classification accuracy, maximum unsupported-claim rate, maximum review time, or zero tolerance for an unauthorized action.
Gate: Proceed only if the workflow meets quality and risk thresholds on representative data.
Gate 3: Run a controlled live pilot
Start with read-only, recommendation-only, or draft-only permissions. Route live cases through the workflow, but preserve a simple fallback to the manual process.
Log the trigger, input references, prompt or rule version, output, tool action, reviewer decision, exception, and final outcome. Ask users to record why they changed or rejected an output. That feedback is more useful than a generic satisfaction score.
Track adoption as a denominator: eligible cases processed by the workflow divided by all eligible cases. A pilot can look accurate while users avoid it on difficult work.
Gate: Production requires stable quality, manageable review, meaningful adoption, no unresolved high-severity failure, and an owner prepared to operate it.
Gate 4: Establish production controls
Production needs:
• Least-privilege system access.
• Human approval for consequential actions.
• Version control for prompts, rules, and schemas.
• Monitoring for failure, latency, volume, and cost.
• An exception queue with a named response time.
• A rollback or pause procedure.
• A user-support and change process.
• A review schedule after model or system changes.
The control level should match impact. An internal summary needs less control than a customer message, refund, payment, pricing decision, employment action, or contract change.
Gate: Do not increase autonomy until the action can be reconstructed and reversed where appropriate.
Gate 5: Prove captured value
Compare the production period with the baseline. Include:
• Net handling time after human review.
• Quality and rework.
• Cycle time and backlog.
• Adoption across eligible cases.
• Exceptions and incident cost.
• Software, usage, maintenance, and owner time.
• Revenue, retention, or conversion only when attribution is credible.
Capacity is not automatically savings. State what happened to recovered time: more cases handled, faster response, reduced overtime, avoided hiring, or new work completed.
Use the 30-day AI automation roadmap to structure a bounded first cycle, but do not force a scale decision before enough representative volume has accumulated.
Gate: Scale only when the conservative value case remains positive after all operating costs.
Gate 6: Scale one dimension at a time
Scaling can increase volume, users, actions, systems, or departments. Change one dimension where possible, then re-evaluate quality, exceptions, review time, and cost.
An agent that drafts messages should not simultaneously gain send permission, a second CRM, and a new geography. That makes failures difficult to attribute.
Maintain a promotion record: what permission changed, which evidence supported it, who approved it, and what would trigger rollback.
Practical SMB operating model
One business owner should own the outcome, and one technical owner should own reliability and access. A small review group should include the people who handle exceptions, not only leadership.
Review weekly during the pilot and immediately after material changes. Once stable, move to a cadence based on risk and volume. Keep a sampled human review even when most low-risk cases run automatically.
The best first production result is not “autonomous.” It is dependable, adopted, measurable, and governable.
Limitations
McKinsey outcomes are self-reported and its high-performer practices are correlational. Census measures broad AI use rather than a specific pilot methodology. ONS “extensive use” is self-described and not identical to production scale.
The recommended gates synthesize evidence and operating practice. They have not been tested as one randomized intervention across all industries.
What would change our view
We would simplify the gate model if representative evidence showed that less controlled deployments reached durable value without higher incidents or rework. We would strengthen it if longer-term studies found that review, ownership, or monitoring failures consistently predicted production loss.
FAQs
When is an AI pilot ready for production?
It is ready when representative live cases meet predefined quality and risk thresholds, users adopt it, review burden is manageable, controls are operational, an owner is accountable, and conservative value remains positive.
What is the difference between a pilot and production?
A pilot is bounded and highly supervised. Production handles eligible live work under defined permissions, monitoring, support, incident response, and service expectations.
How long should an AI pilot run?
Run it long enough to observe representative volume, edge cases, user behavior, and operating cost. A calendar target can structure work, but the evidence threshold should control the decision.
What should stop an AI pilot?
Stop or redesign when high-severity errors escape controls, users avoid the workflow, review cost eliminates value, required data is unreliable, or the project cannot meet a predefined outcome threshold.
Should an SMB run several AI pilots at once?
Usually not for a first implementation. One owned workflow produces clearer learning, stronger adoption, and lower integration overhead than several uncoordinated pilots.
Get a 20-Minute AI Workflow Audit
AI Operator can score one pilot against the workflow, evidence, control, adoption, and value gates required for a production decision.