From Pilot to Rollout: Why Only a Quarter of AI Programs Get Past the Pilot

A client-reporting agent drafts a strong first version. A service-request agent routes routine cases correctly. In a pilot, both look like wins.

Then the workflow reaches people who were not in the pilot, inputs that are incomplete or contradictory, cases outside the happy path, and managers with no time to supervise another process. The agent may still be capable. The workflow is no longer reliable.

That is the catch: a small pilot is a controlled exception; rollout is a reliability test across real people, systems, edge cases, and incentives.¹1By "rollout system" we mean owners, routines, and a baseline measure — not more software. Every item in it is organizational; none of it is a procurement decision.

This article is a practical guide to scaling agentic workflows: AI agents that carry out or support a repeatable work process, with people accountable for the outcome. It identifies the specific conditions that break during rollout, shows how to design for them, and ends with a readiness gate to use before expanding access.

The point is not to prove that an agent can produce an impressive result. It is to build a workflow that produces a useful result reliably, safely, and repeatedly—and to show that it creates value.

The scale problem is real

The broader evidence on enterprise AI points to the same problem: organizations can start pilots far more easily than they can turn them into durable value. BCG's 2024 survey of 1,000 executives found that only 26% of companies had moved beyond proofs of concept to tangible value.^1◦ Deloitte's 2025 survey found that more than two-thirds of leaders expected 30% or fewer of their current gen-AI experiments to scale fully within three to six months.²SourceBCG, "Where's the Value in AI?", October 2024By the numbers26%of companies got past proof of concept to tangible value — BCG, 2024SourceDeloitte, "State of Generative AI in the Enterprise", Wave 4, January 2025

The issue is not simply that organizations abandon weak ideas. Gartner expected at least 30% of gen-AI projects to be abandoned after proof of concept by the end of 2025, citing data, risk, cost, and unclear value.³² MIT's 2025 Project NANDA study similarly found that around 95% of enterprise gen-AI pilots had no measurable P&L impact.⁴³SourceGartner press release, 29 July 20242Gartner's figure was a forward-looking prediction, not a measured outcome.SourceMIT Project NANDA, "The GenAI Divide: State of AI in Business 2025" (via Fortune)3This figure is not peer-reviewed and reflects 150 leader interviews, a 350-employee survey, and 300 public deployments. “No measurable P&L impact” does not mean the tools were useless; the same report found individual productivity gains.

Thailand shows why access is an incomplete measure of progress. AWS and Strand Partners found continuous AI use rose to 43% of organizations in 2026, but 74% remained at basic use and only 9% had reached advanced use.^5◦ The rollout question, then, is not “Can people use AI?” It is “Can this workflow keep working when normal work gets messy?”⁴SourceAWS with Strand Partners, "Unlocking Thailand's AI Potential 2026" — vendor-commissioned and self-reported (via Forbes Thailand)By the numbers9%of Thai organizations had reached advanced AI use; 74% remained at basic use — AWS with Strand Partners, 20264Vendor-commissioned and self-reported — AWS sells the infrastructure the study encourages buying. Directional, not audited.

What breaks when an agentic workflow leaves the pilot

A pilot proves that an agent can complete a task under supportive conditions. It does not prove the organization can run that task reliably at scale. The difference appears in four predictable places:

1. Edge cases. The pilot usually uses clean, familiar inputs. Rollout introduces missing information, conflicting policies, unusual requests, and failures in connected systems. Define the workflow boundary before scaling: what the agent handles, what it hands to a person, and what it must stop and flag.
2. Ownership. Pilot teams catch errors because they care and are nearby. At scale, someone must own performance, changes, and failures. Assign a named workflow owner, a backup, and a regular review.
3. Handoffs. An agent's output is rarely the end of the work. People need to know when to accept it, review it, edit it, or escalate it. Make those decisions explicit in the workflow—not implicit in training.
4. Proof of value. A striking demo can hide extra review time, rework, or new risk. Baseline one or two outcomes before rollout: cycle time, error rate, escalation volume, rework, conversion, or customer resolution time.

BCG has measured that second population twice, and the movement between the two readings matters more than either one. In 2025, across more than 10,600 employees in 11 countries, regular gen-AI use among leaders and managers ran above 75% while frontline use sat at 51%⁶. By the 2026 edition — 11,749 workers across 14 markets — frontline regular use had reached 74%, up 23 percentage points in a year^7◦.SourceBCG, "AI at Work 2025"SourceBCG, "AI at Work 2026: Strategy Matters More Than Tools", June 2026 (via BCG press release)By the numbers51% → 74%frontline regular AI use across BCG's 2025 and 2026 AI at Work editions — the access gap closed inside a year

So the frontline caught up on access. What did not follow was the work itself: in that same 2026 survey, 47% now spend more time directing AI than doing the task, and 66% get little or no guidance on what to do with the time they save. The enthusiasts in your pilot are still not a small version of your organization — but the tail you have to reach is no longer people without access. It is people with access and no redesigned job to put it in.

The question is not "did the pilot work?" It is "what has to be true for the 400th person to use this without you in the room?"

Do not scale the agent before you scale the operating model

The instinctive next step after a good pilot is to buy more licenses, connect more tools, or give the agent to more people. That expands exposure, not reliability. First redesign the work around the agent: its inputs, decisions, handoffs, quality checks, and accountable owner.

McKinsey's March 2025 State of AI survey of 1,491 organizations tested around 25 attributes against bottom-line impact. The one with the biggest effect was fundamental workflow redesign — and only about 21% of adopters had redesigned any workflows at all⁸.^◦ The same survey found more than 80% of organizations seeing no tangible enterprise-level EBIT impact from gen AI, and fewer than one in five tracking well-defined KPIs for their gen-AI solutions.⁵SourceMcKinsey, "The State of AI", March 2025By the numbers~21%of adopters had redesigned any workflow, though redesign was the strongest predictor of EBIT impact — McKinsey, 20255McKinsey's EBIT findings are correlations with self-reported impact, not causal measurements. The direction is well supported; the magnitude is not.

MIT's study found the same thing from the other end: the driver of failure was a "learning gap" in workflow integration, not model quality. Notably, purchased or partnered solutions succeeded roughly 67% of the time, while internal builds succeeded about a third as often — which is the opposite of what most engineering-led AI programs assume.

BCG's rule of thumb for where AI value actually comes from is 10-20-70: 10% algorithms, 20% technology and data, 70% people, process and culture⁹.^◦ Whether or not the exact split holds, it points at the same conclusion as the measured studies: you cannot buy your way out of the 70%.SourceBCG's 10-20-70 heuristic — a consulting rule of thumb, not a measured statisticDefinition10-20-70 BCG's heuristic for where AI value originates — 10% algorithms, 20% technology and data, 70% people, process and culture. A consulting rule of thumb from client experience, not a measured statistic.

Microsoft's 2026 Work Trend Index put a measurement next to the heuristic. Across 20,000 knowledge workers in ten markets, organizational factors — culture and manager support — accounted for 67% of reported AI impact, against 32% for individual mindset and behavior: more than twice the weight^10◦. Microsoft's own summary of its findings is blunter than anything we would write: workers are ready, their organizations are not.⁶SourceMicrosoft, 2026 Work Trend Index Annual ReportBy the numbers67% vs 32%how much of reported AI impact comes from organizational factors versus individual mindset — Microsoft, 20266Microsoft sells the Copilot this study encourages buying, and the sample is knowledge workers who already use AI — so it cannot speak for non-adopters. Read it as a vendor study whose direction is corroborated by BCG's independent 2026 survey.

Build the operating system around the agent

1. Name a workflow owner. This is not a central AI team member who promotes adoption. It is the person accountable for whether one specific workflow performs: monitoring quality, resolving failures, approving changes, and reporting value. Give the owner a backup and a recurring review.^◦RelatedWhat changed when adoption got a named owner
2. Define the workflow boundary. Write down the trigger, inputs, expected output, systems the agent may use, and the cases it must escalate. A service-request agent might resolve standard requests, route incomplete requests for clarification, and escalate policy exceptions. A report-writing agent might draft from approved sources but never invent an unsupported recommendation.^◦FigureA rollout holds when a named owner carries it, department by department.
3. Design the human handoff. Specify who reviews what, at which risk threshold, and how quickly. “Human in the loop” is not a process; a named decision, an escalation route, and a response expectation are.
4. Instrument quality and value. Establish a baseline, a target, and a monthly read on real outcomes—not licenses. Track the measure the workflow exists to improve, alongside agent error, rework, and escalation rates.^◦RelatedPutting a defensible number on adoption
5. Keep a visible feedback loop. Give users one clear place to report failures, missing context, and recurring exceptions. The owner should turn those reports into workflow fixes fast enough that people keep reporting.

None of these is solved by a better model, a bigger license pool, or another generic training session.

Where rollouts fail in practice

The operating model is straightforward to describe. The hard part is keeping it alive when the workflow meets real pressure:

The owner role disappears into the day job. In a pilot, ownership can run on enthusiasm. At scale, make workflow performance part of the owner's objectives, appoint a backup, and put a short quality-and-value review on the calendar. If the review does not happen, the workflow has no owner in practice.

Measurement disappears before the workflow does. A monthly review should look at one value measure and two reliability measures. For a report-writing agent, that might be drafting time, editor rework, and unsupported claims. For a triage agent, it might be resolution time, incorrect routing, and escalations. Assign the review to someone whose role is not simply “AI champion.”

People treat plausible output as a reliable answer. Judgment training is part of the workflow, not a general AI-literacy extra. Dell'Acqua and colleagues found that AI-assisted consultants performed far better on tasks inside AI's capability, yet were 19 percentage points less likely to be correct on a task outside it.^11◦ Teach users the boundary: what evidence to check, which outputs need review, and when to escalate.^◦SourceDell'Acqua et al., "Navigating the Jagged Technological Frontier", 2023By the numbers−19pplower likelihood of a correct answer on a task outside AI's capability — Dell'Acqua et al., 2023RelatedMeasuring judgment before you widen access

Routine use gets mistaken for transformation. Basic adoption can be a useful first outcome, but it is not proof that a workflow has changed. Decide upfront what a successful next stage looks like: fewer handoffs, faster resolution, higher-quality output, or a redesigned service. Without that decision, a program can plateau on an early win for years.

A practical first move

Choose one workflow that has already shown promise and run a controlled scale test with one additional team. Before expanding further, appoint its owner, document the workflow boundary and human handoffs, assemble real edge cases, and establish a baseline.^◦ The goal is not more agent activity. It is reliable performance on work that matters.^◦DefinitionScaffolding The supporting system around an agentic workflow—ownership, workflow design, measurement, and feedback—that makes adoption hold.RelatedSee how one firm applied this approach to report writing

Before you scale an agentic workflow

Use this as a go/no-go gate:

• Boundaries: Can we state which cases the agent handles, which it escalates, and which it must not attempt?
• Ownership: Is one named person accountable for quality, changes, and outcomes—with a backup?
• Inputs: Are the agent's sources approved, current, accessible, and clear enough to support reliable output?
• Handoffs: Do people know when to accept, review, edit, or escalate the agent's work?
• Reliability: Have we tested routine, messy, and high-risk cases, and set a quality threshold?
• Value: Do we have a baseline and a monthly measure of the outcome this workflow should improve?

If any answer is no, do not widen access yet. Fix the workflow around the agent first. A pilot becomes a durable capability only when the organization can run it without the pilot team in the room.