Agentic Factory: industrializing AI agents in the enterprise

At our latest hackathons, several hundred of our consultants delivered agentic demonstrators to their industrial clients: more than a hundred engineers working in parallel, some 250,000 lines of code produced, under three hours of clocked delivery per demonstrator, and every result scored out of one hundred against a no-AI baseline. I say this without false modesty, but also without illusion: what those sessions prove is a capability, not yet an industry. Between the demonstrator that draws a smile from the client at five in the afternoon and the system that will run in their environment a year from now, there is a wall. This paper is about that wall, and about how to get past it.

The paradox: never so many demos, never so little production

The figures for 2025-2026 sketch a paradox McKinsey eventually put a name to: nearly eight companies in ten use generative AI, and just as many see no significant impact on their bottom line. More brutal still: fewer than 10% of deployed use cases ever make it past the pilot stage^[1]. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, driven by escalating costs, unclear business value and inadequate risk controls^[2]. And the gap is widening rather than closing: BCG identifies 5% of "future-built" companies capturing most of the value, 35% scaling, and 60% getting almost nothing out of it — while agents, already accounting for 17% of AI's total value, are expected to reach 29% by 2028^[3].

What is striking across all these studies is that the gap is explained neither by access to the technology — everyone has the same models — nor by the talent of the teams. McKinsey points to the real dividing line: moving "from use cases to business processes, from siloed AI teams to cross-functional squads, from experimentation to industrialized delivery"^[1]. Google's DORA report puts it in a sentence I quote at every executive committee: AI is an amplifier — it magnifies the strengths of organizations that have them, and the dysfunctions of the rest^[4]. So the question is not "which agent should we build?" but "which factory should we build around it?".

Why the demonstrator dies in demo mode

First, one has to understand why the wall exists, because it is not the product of lazy teams. A demonstrator proves that something is possible; production demands that it be reliable — and those two properties do not progress together. Agent benchmarks measure it: the best systems succeed at a given task, but their success rate collapses when asked to succeed at it eight times in a row^[5], and two years of spectacular capability gains have produced only modest gains in reliability^[6]. The demonstrator lives in the world of "it worked"; production lives in the world of "it works every time, including when everything goes wrong". Getting past the wall means industrializing precisely what separates those two worlds.

The second cause is organizational. The prototype is born of individual enthusiasm — a gifted developer, an inspired afternoon — and everything that made it succeed stays in that person's head. Microsoft, which reports a fifteen-fold growth in active agents across its tools in one year, also measures the flip side: even among the most advanced professionals, those orchestrating multi-step agentic workflows, only 26% have documented, replayable workflows at team level^[7]. Three quarters of organizations' agentic know-how therefore lives where the knowledge of star developers used to live: in people's heads. We know how that ends.

The four floors of the factory

An agentic factory is not a tool you buy; it is an operating model you build, and it rests on four floors.

1. Versioned, replayable workflows

The first floor turns individual successes into organizational processes: every recurring task — documenting a module, migrating a screen, generating tests from requirements, reverse-documenting a legacy system — becomes a written workflow, versioned, replayed identically from one project to the next, and improved with every iteration. This is exactly what the two reference labs recommend, each in its own vocabulary: Anthropic observes that "the most successful implementations weren't using complex frameworks" but "simple, composable patterns", and carefully distinguishes workflows — orchestrated by predefined code — from agents free to direct their own process^[8]; OpenAI advises starting with a single agent, and only evolving to a multi-agent system when necessary^[9]. Sophistication is not a goal: replayability is.

2. Systematic verification gates

The second floor is the one that separates the reliable from the plausible. No deliverable ships without passing its gates: a Definition of Done written as three to five objective criteria before starting, an adversarial review after every high-volume production run, and layered guardrails — because, as OpenAI's guide puts it, a single guardrail is rarely sufficient on its own; it is the combination of multiple guardrails that creates resilient agents — with two explicit triggers for human escalation: exceeding failure thresholds, and high-risk or irreversible actions^[9]. This is not a good intention: it is a tooled process, planned into the workflow just like the production itself.

3. Honest measurement

The third floor steers the whole. Every deliverable is evaluated against a no-AI baseline — that is the score out of one hundred from our hackathons, and it is what turns an impression into data. In production, this measurement becomes an evaluation suite that grows with the system; and there is no need to wait for a perfect one to get started: twenty to fifty simple tasks drawn from real failures make an excellent starting point, advises Anthropic in its guide to agentic evals^[10]. Honesty is part of the measurement: output volume is not a demonstrated speed gain, and saying so strengthens the credibility of everything else.

4. Capitalization that compounds

The fourth floor is the one that turns the factory into a cumulative advantage: at the end of every engagement, the tooling bricks, the method rules and the evaluation cases join the common foundation — so that each project starts higher than the one before. It is the difference between an organization that has done ten agentic projects and an organization that has ten times the experience of one project. Without this floor, the gains remain individual, invisible and volatile; with it, they become an asset — the very one I have argued elsewhere outlives the models that come and go.

The operating model: the POD, and the art of picking your battles

What remains is running these four floors day to day, and that is where format matters. Our unit of production is the POD: a compact team — an experienced lead, tooled-up consultants, and the client themselves in the loop — dedicated to a single, bounded use case, with its real material (code, specifications, data) collected before starting and its Definition of Done validated in writing. The client is not a spectator: the best results come from sessions where they arbitrate live. Microsoft describes the same trajectory at market scale: hybrid human-agent teams, where the organizational factor weighs twice as much as individual aptitude in the impact achieved^[7].

And because a factory is only worth what you decide to produce in it, qualifying the use case is a discipline in its own right. Years of hackathons have taught me to recognize what works: an achievable scope — one brick, not a whole project —, real material available, a result that can be demonstrated visually, a pain quantifiable in human hours, and repeatability across the client's estate. Migrating a legacy module to a modern stack from its specs, generating test cases from requirements, reverse-documenting a historical piece of software: these are cases that get past the wall. "Migrate our entire IT landscape", a vague need with no success criterion, a case dependent on an inaccessible environment: these are the ones that die in demo mode. Even confidentiality gets handled upstream — a well-built synthetic dataset demonstrates the same thing as real data, and unblocks a deal in a week instead of a month.

From demonstrator to decision

One last floor, often neglected, is not technical at all: it is the path from the successful demonstrator to the decision to industrialize. Two rules of honesty first, which paradoxically help sell: a demonstrator is not production-ready as it stands — industrialization is a project in its own right, and saying so up front avoids the disappointments that kill trust; and the demonstrator does not create the need — it accelerates a decision already germinating. Experience confirms it every time: follow-on work materializes with clients where a modernization project already existed, and where the decision meeting was put in calendars before the demonstration, not after. So the factory does not stop at the deliverable: it includes the same-day debrief, the dated readout, and the proposal that turns three hours of proof into a trajectory.

Where to start

If I had to compress this paper into one grid, it would have four maturity levels. Artisanal: individual, undocumented successes — the state of three quarters of the market. Tooled: written, replayable workflows and a prepared environment. Measured: systematic verification gates and evaluation against a baseline. Capitalized: a common foundation that grows richer with every project, and hybrid teams that live off it. Each level takes months to climb, not years — but none can be skipped. And the first step costs almost nothing: take your best agentic success of the year, and write it up as a workflow someone else can replay. The day your best success no longer belongs to a person but to the organization, your factory exists. The rest is growth.