Enterprise
The prototype is quick. Production is the work.
Updated 13 July 2026. The original version of this essay repeated a conference anecdote: an enterprise agent took two weeks to build and twelve months to reach production. It is a memorable ratio, but I cannot verify it as a published case study. The lesson survives without pretending one story is a general measurement.
A convincing prototype can be quick because it proves a narrow technical possibility. Production has a different job. It must prove that the system uses authorised data, behaves within agreed boundaries, survives failure, can be monitored, and creates enough value to justify ownership after the demonstration ends. The gap is not “code versus bureaucracy.” It is possibility versus accountable operation.
Correct the number before using the lesson
The original essay also used Accenture's “12% AI Achievers” figure as if it described the success rate of agentic projects. It did not. The number came from a 2022 study of roughly 1,200 companies and described a broad AI maturity category. It is useful historical research, not a 2026 failure rate for agents. Accenture's official release defines the measure and its sample.
A more current signal comes from Accenture's 2026 Agentic Imperative report. In its Generating impact research, 28% of surveyed Great Britain executives reported having tools to monitor and audit agent behaviour at scale, while 57% said their organisations were not ready to supervise agents in real time. Those are vendor-reported survey results from one market, not universal laws. They point to a practical bottleneck: operational control is less mature than enthusiasm for building.
The control layer is part of the product
Security, privacy, legal review and change control are not delays attached to the real system. They are requirements of the real system. The mistake is leaving them as manual surprises at the end. A good harness turns repeatable policy into enforceable permissions, logged decisions, automated checks and explicit approval gates. People still own consequential judgement. They no longer have to rediscover the same control for every run.
This is consistent with the voluntary NIST Generative AI Profile, which treats governance, mapping, measurement and management as work across the AI lifecycle. Production readiness is therefore more than a passing test suite. It includes named responsibility, documented risk, evaluation on the intended use and a plan for monitoring and incidents after release.
Work one release decision all the way through
Imagine an agent that drafts supplier-renewal recommendations. A prototype retrieves a contract, compares prices and writes a neat summary. Production introduces the questions the demo avoided. Which contracts may it read? How fresh must pricing data be? What counts as a conflict of interest? Who can approve a recommendation? What happens when a source is missing? How will a reviewer reconstruct the evidence six months later?
Start in shadow mode: run the agent beside the existing process on a fixed set of renewals, with no ability to send or commit anything. Compare its factual accuracy, missed risks, unsupported claims and review time against the human baseline. If the evidence clears thresholds defined in advance, move to advisory use, where a named person accepts or rejects every recommendation. Only a narrow, reversible action should be considered after the system has a reliable record and a tested stop path.
Each rung has an owner and an exit condition. A failed threshold sends the workflow back for diagnosis or stops it. This is the operational meaning of treating completion as a claim. Evidence does not eliminate uncertainty, but it gives the next decision something firmer than confidence in a demo.
Production authority is earned one rung at a time
Each stage adds reach only after predefined evidence and a tested stop path.
- Prototype Prove the narrow technical possibility with synthetic or controlled data.
- Shadow mode Run beside the current process with no ability to send or commit.
- Advisory use A named person accepts or rejects every recommendation.
- Narrow reversible action Grant one bounded action after reliable evidence and rollback.
- Monitored operation Track drift, incidents, value and continuing ownership.
A failed threshold sends the workflow back for diagnosis or stops it.
Use a portfolio, but do not abandon the business case
AI pilots contain uncertainty, so a portfolio approach can be sensible: make several small, bounded bets and stop weak ones early. That does not make a business case obsolete. It changes its shape. State the operational pain, current baseline, risk budget, evaluation threshold, pilot cost and person accountable for the decision. The early case buys learning, not certainty.
The learning loop matters too, but it is not “the only moat.” Useful advantage can also come from distribution, domain expertise, service, trust and access to lawful, high-quality data. Feedback becomes durable when a correction is evaluated, governed and retained, the same mechanism described in renting capability while retaining learning.
The student lesson is to stop treating the prototype as proof of the whole system. The professional lesson is to design ownership, evidence and control while the workflow is still small. That discipline may make the first release look slower. It is what makes later releases safer to repeat, and protects everyone from spending a year scaling a result that was never ready to leave the room.
Sources & caveats. The 12% figure is Accenture's 2022 AI maturity category, not an agent-project success rate. The 2026 monitoring and supervision figures are reported by Accenture. The lifecycle framing comes from NIST's voluntary Generative AI Profile. The supplier example and release ladder are illustrative, and the interpretation is mine.