Enterprise AI From Experimentation to Production: The Playbook

Most enterprise AI programs do not have a pilot problem. They have a second-workflow problem.

If your second production workflow costs as much to ship as your first, you have not moved from experimentation to production. You are running a series of expensive experiments.

Below: why pilots stall, the four shared pieces that make workflow #2 cheap, who owns what, and how to fund and staff the move.

Why enterprise AI stalls between experimentation and production

MIT’s Project NANDA, in its July 2025 report The GenAI Divide, looked at task-specific enterprise AI tools. 60% of organizations evaluated one, 20% reached a pilot, and 5% reached production. The authors say these figures come from interviews and are directional. In S&P Global Market Intelligence’s 2025 survey of more than 1,000 respondents, the share of companies abandoning most AI initiatives rose to 42%, from 17% a year earlier. The average company scrapped 46% of its proofs of concept before production, per CIO Dive’s March 2025 report. And Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value, or inadequate risk controls.

The surveys do not say why in architectural terms, but here is the failure mode I would bet on. Each pilot is built as a standalone project. It gets its own service account, its own logging (or none), its own test spreadsheet, and a security review that starts from zero. The pilot budget pays for all of it, then ends at the demo. To go live, every one-off piece must be rebuilt to a production standard that nobody budgeted for. Data adds a second tax. A February 2025 Gartner release says 63% of organizations lack, or are unsure they have, the right data management practices for AI.

A caution: these surveys measure different things (custom tools, initiatives, proofs of concept, agentic projects), so none of them is “the failure rate.” Read together, they say many initiatives stall or are abandoned before production: MIT counts stages reached in its interviews, S&P counts proofs of concept scrapped in its 2025 survey, and Gartner forecasts agentic cancellations through 2027. The named reasons are mostly cost, value, risk, and data. Model capability does appear: Gartner said in 2025 that models then lacked the maturity and agency to reach complex business goals or follow nuanced instructions over time. But it is one reason among several, and the only one you cannot fix yourself. For stage definitions, see PoC vs pilot vs production. Stuck now? That is pilot purgatory.

An early test: ask your current pilot team, “What will we reuse for the next workflow?” If the answer is “some prompt patterns and lessons learned,” the pilot is a dead end even if it works.

The real test: is workflow #2 cheaper than workflow #1?

This is the one number I would track. A hypothetical: say a company ships an accounts payable exception agent as its first production workflow. The effort, in engineer-weeks:

Without shared pieces, workflow #2 costs 17 weeks, or 94% of the first. With them, it costs 9.75 weeks, or 54%. Say building the shared versions adds 8 engineer-weeks to workflow #1. Each later workflow saves 7.25 weeks (17 minus 9.75). Workflow #2 recovers most of the 8, and workflow #3 finishes paying it back. By workflow #5, you are 21 weeks ahead (four workflows times 7.25, minus 8).

Workflow logic and integrations do not shrink. A platform removes the tax, not the work.

McKinsey’s 2026 State of AI survey, fielded May to June 2026 with 1,719 respondents, points the same way. 44% said AI is scaling across their enterprise, up from 38% a year earlier, yet only 37% attribute any EBIT impact to AI, about the same as in 2025. The roughly 6% it calls high performers are far more likely to have redesigned the work: nearly three-quarters report fundamentally redesigning workflows, against one-quarter of other respondents. They are also twice as likely to have defined processes to measure the impact of AI initiatives. That second practice is decided once, for every workflow.

Decision rule: if the estimate for your next workflow is more than 70% of what the last one cost, stop starting new pilots. Spend the next sprint on the shared pieces. If the estimate is under 60%, the platform is working. Add workflows.

The four shared pieces: identity, evals, logging, release gate

CNCF’s platforms white paper, completed in March 2023, says the essential goal of a platform is to reduce the cognitive load on product teams. It also advises the thinnest viable platform layer over managed services. For AI, that means four pieces before workflow #2, and little else.

1. Agent identity. Every agent gets its own non-human identity, scoped to one workflow, with a named human owner and an expiry date. No agent runs on an engineer’s personal token or a shared “ai-service” key. Vendors have moved this way. Microsoft Entra Agent ID is now generally available, and Microsoft’s June 2026 guidance gives every agent a sponsor, accountable for its purpose and retirement, and an owner, responsible for its configuration. Keep an inventory: the NIST AI RMF (January 2023), in GOVERN 1.6, calls for mechanisms to inventory AI systems. Warning sign: security cannot list every agent and what it can write to within an hour.

2. Eval runner. One runner, one report format, many case sets. The platform team owns the runner and the CI hook. Each workflow team owns its cases. My starting point is 30 to 50 cases from real historical data on day one, so a new workflow adds cases instead of building a framework. More in AI agent testing strategies and agent evals in CI.

3. Logging and traces. One trace schema across all agents: run ID, workflow, agent identity, model and prompt version, tool calls, tokens, cost, outcome, and any human override. OpenTelemetry’s generative AI semantic conventions, which moved to their own repository in 2026, define spans, metrics, and events for model calls, agent operations, token usage, and MCP. The spec is still marked Development status, not stable, so pin the version you adopt. Compare tools in AI agent observability tools.

4. Release gate. One tiered checklist for every workflow. Tier 1 is internal work with read-only access or human-approved writes: the tech lead approves once evals pass. Tier 2 is any agent that writes to systems of record or touches customer data: add security and risk sign-off and a shadow period. Tier 3 is regulated decisions or unreviewed external actions: add a formal risk assessment and executive sign-off. Most requests should land in tier 1 and get a yes in days.

Do not centralize prompts, workflow logic, eval cases, or data mappings. Those belong to workflow teams. A platform team that starts building one “general agent” for everyone has lost the plot.

Who owns what: a RACI for the move to production

Pilots skip this because the pilot team does everything. That works until the first incident at 2 a.m., when nobody knows who owns the agent. NIST’s AI RMF says roles and responsibilities should be “documented and are clear to individuals and teams throughout the organization” (GOVERN 2.1). It also puts decisions about AI risk on executive leadership (GOVERN 2.3).

My starting RACI: R does the work, A is the one person with final say, C is consulted, I is informed.

Three rules. First, one A per row. If two people are accountable, nobody is. Second, the business owner is accountable for eval cases. Only the people who do the work know what a wrong answer costs. MIT’s report found successful organizations let budget holders and domain managers surface problems, vet tools, and lead rollouts. Third, the platform team holds A only on its own pipeline row. Accountable for workflow outcomes, it becomes the bottleneck. Check quarterly: if the platform team is A on more than one row, fix it.

How to fund the move

Pilot budgets are project budgets. Production costs recur: on-call time, eval upkeep, model version changes, and token spend that grows with volume. Unowned, the agent drifts until someone turns it off.

BCG’s Applied AI Index 2026, a survey of 1,330 CxOs and senior leaders released September 30, 2026, found corporate AI spending doubled in under a year to 3.3% of revenue. More than 80% of it now sits outside the enterprise IT budget. McKinsey’s 2026 high performers are more than twice as likely as others to spend over 15% of their enterprise-wide information and communication technology budget on AI. My read: if most AI money already sits in business budgets, the workflow line below belongs there too.

There is also a zombie problem. In Deloitte’s January 2025 report, based on a survey of 2,773 leaders fielded July to September 2024, 70% said they need at least 12 months to resolve ROI challenges. 76% said they would wait at least 12 months before cutting investment in an initiative that misses its value targets. That patience fits a platform. For a single workflow, a year of waiting without a written stop rule is a year of spend that nobody decided to keep making.

My recommendation is two budget lines. The platform line is central, funded for 12 months, and judged by the cost of workflow #N against workflow #1. After year one, show its cost back to the workflows that use it. The workflow line belongs to the business owner, is released in stages, and gets a 90-day review on the hero metric and cost per case. Write the stop rule first, in a scale-or-stop charter.

Measure cost per completed case, not token spend. The FinOps Foundation’s FinOps for AI overview recommends unit metrics such as cost per inference and per token, plus tagging and quotas. Hypothetical: an agent handles 6,000 cases a month. Model and hosting cost $3,600, or $0.60 a case. Humans review 20% of cases at 4 minutes each, which is 80 hours. At $45 an hour loaded, that is another $3,600, or $0.60. A platform showback of $2,400 adds $0.40. The total is $1.60 per case. Model and hosting are only about 38% of it.

How to staff it

My minimum year-one shape (a judgment call, not a benchmark) has three groups.

The platform team is 2 to 4 engineers who own identity integration, the trace pipeline, the eval runner, and the release pipeline. One of them should come from security or IAM, because identity is the piece a pilot can fake and production cannot. Each workflow pod is one business owner at about a day a week, one or two engineers, and one reviewer from the team that does the work today. A named risk partner in compliance knows the tiers and turns tier 2 reviews around in days.

Build or partner? In MIT’s sample of 52 organizations, external partnerships reached deployment about 67% of the time, against about 33% for internal builds. The authors warn this may reflect organizational capability rather than the approach itself. My read: a partner can speed up workflow #1, and hurts you if it keeps the platform.

Decision rule: if you partner, write four conditions into the contract. The agent identities live in your tenant. The traces land in your store. The eval cases sit in your repo. Your platform engineers pair on the build. If a partner cannot work that way, your second workflow will cost as much as the first.

Hero-engineer check: if one person’s vacation would stop releases, you are not staffed for production.

A 90-day transition plan with pass/fail gates

Run these gates in order. None of the tests needs a meeting.

Pick workflow #2 around day 60, so the last estimate rests on a real scope.

Bottom Line

The move from experimentation to production is a cost curve, not a launch date.

Build four shared pieces: agent identity, a shared eval runner, one trace schema, and a tiered release gate. Put one accountable person on every row. Fund the platform and the workflows on separate lines.

Next step: estimate your next workflow against your last one. If it is above 70%, you know what to build first.

FAQ

What percentage of AI pilots make it to production?

It depends on what is counted. MIT’s 5% figure above covers task-specific tools only. In Deloitte’s 2026 State of AI in the Enterprise survey of 3,235 leaders, fielded August to September 2025, only 25% said they had moved 40% or more of their AI pilots into production. Track your own conversion rate per workflow instead of borrowing one.

How long does it take to move an AI pilot into production?

MIT NANDA’s 2025 report says top-performing mid-market companies averaged about 90 days from pilot to full implementation, while large enterprises took nine months or longer. Company size explains part of that gap. My read: shared identity and a tiered release gate attack a real share of the rest, because they remove reviews that otherwise start from zero.

Which workflows should go to production first?

Asked to split a hypothetical $100 of generative AI budget, MIT NANDA’s executives put the largest share into sales and marketing (the report gives the share as about 50% in some places and about 70% in another), yet back-office automation often returns more, with reported savings of $2 million to $10 million a year from replacing outsourced processing. Gartner suggests using agents where decisions are needed, plain automation for routine workflows, and assistants for simple retrieval.

Do we need an AI center of excellence?

You need a platform team more than a committee that picks use cases. CNCF’s platforms paper says platform teams should treat the platform as a product, gather user feedback, and measure how people actually use it.

Is agentic AI ready for enterprise production?

For narrow, well-owned workflows, yes, with care. Deloitte’s 2026 survey found close to three-quarters of companies plan to deploy agentic AI within two years, yet only 21% of those report a mature model for agent governance. Gartner estimated in 2025 that only about 130 of the thousands of agentic AI vendors are real, so test vendor claims on your own data.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect