← Insights
AI · Adoption
Copilot, agents, and the gap between deployed and adopted
Seventy-nine percent of enterprises have adopted AI agents in some form. Eleven percent are running them in production. The story in between is not a technology problem.
Key takeaways
- 79% of enterprises have adopted AI agents in some form; only 11% run them in production — a 68-point gap described as the largest deployment backlog in enterprise technology history.
- 88% of AI agent projects never reach production at all — and when they fail, infrastructure gaps, governance barriers, and unclear ROI are roughly equally to blame.
- Trust hasn't caught up with deployment: only 29% of developers trust AI-generated output today, down from 40% two years ago.
- The gap has four common, fixable causes — broad rollout before narrow proof, assumed trust, no named owner for the number, and governance arriving after the rollout instead of before.
The Copilot licenses went out to the whole finance team in March, with a rollout email and a lunch-and-learn. By January, the usage report told a familiar story: a spike in week one, a slow bleed through the summer, and by year-end a handful of power users carrying whatever number still showed up in the renewal deck. Nobody uninstalled it. It just stopped being where people went to get work done.
That pattern is closer to the norm than most rollout plans assume. 79% of enterprises have adopted AI agents in some form Digital Applied, 2026 — but only 11% are running them in production Digital Applied, 2026. A 68-point gap between "we have this" and "we use this" has been described as the largest deployment backlog in enterprise technology history.
88%
of AI agent projects never reach production at all. When they fail, the causes split roughly three ways: infrastructure gaps, governance and security barriers, and an inability to measure ROI convincingly enough to justify scaling further.
Deployed is not the same as adopted
The two numbers get treated as if they measure the same thing. They do not. "Adopted" usually means a license was issued, a rollout was announced, a pilot was greenlit. "In production" means people are actually depending on it to do real work, week after week, without someone having to remind them it exists. Nearly every organization can claim the first. Very few can claim the second — and the gap between the two is exactly where the budget for the initiative quietly disappears.
Decisions made
Pilot & rollout plan
Broad license deployment
Launch comms
Adoption gap visible
Months 1–2
Months 2–4
M4
Months 5–18+
Where structural decisions are madeWhere the gap becomes visible
The gap is structural, not temporal. By the time a usage report makes the adoption problem undeniable, the decisions that caused it — who it was rolled out to, what problem it was scoped to solve, who owns the number afterward — were made months earlier. A renewal-time usage review is the wrong point in the timeline to fix a scoping problem.
Why trust hasn't caught up
Scoping explains part of the gap. The rest is trust, and the data on that front hasn't moved in the direction anyone hoped. Only 29% of developers trust the accuracy of AI-generated output today Unico Connect, 2026 — down from 40% two years earlier — and 46% actively distrust it, which is precisely why human review remains standard practice rather than a transitional inconvenience.
The same story shows up closer to home. 44% of enterprise users who stopped using Copilot cite not trusting the AI's answers as the reason they walked away Recon Analytics, 2026. That's not a training problem or a change-management gap — it's a straightforward statement that the tool didn't earn continued use. No adoption plan that skips the trust question closes that gap; it just delays when the gap gets noticed.
The pattern isn't confined to smaller rollouts, either. One widely-discussed analysis of Microsoft's own Build 2026 announcements put the same gap at the largest scale: 64% of provisioned seats for the built-in sidebar Copilot assistant reportedly went unused industry reporting on Microsoft Build 2026. That's one analyst's read, not an official figure, and it deserves the same scrutiny as any other number in this piece — but it points the same direction as everything else here: deployment and adoption aren't the same thing, at any scale.
None of this makes agentic AI the wrong bet — Gartner still expects 40% of enterprise applications to include task-specific agents by the end of this year, and the organizations that get the adoption question right first are positioned to capture a real advantage over the ones still treating deployment as the finish line. But Gartner also expects more than 40% of agentic AI projects to be cancelled by the end of 2027, for reasons that sound less like technology limits and more like the scoping and trust problems above: escalating costs, unclear value, inadequate risk controls.
90%
68%
50%
39%
34%
30%
Rollout week
Month 1
Month 3
Month 6
Month 9
Month 12
Illustrative pattern, not a specific client result — shaped to land in the same range as the 64% unused-seats figure reported for Microsoft's own sidebar Copilot at twelve months. Actual decay rate varies by how narrowly the use case was scoped and whether anyone owned adoption after rollout.
Four reasons the gap opens — and stays open
1. Deployed broadly before anyone proved narrow value
The default rollout model is enterprise-wide licensing on day one — every team, every seat, all at once. It is the easiest thing to announce and the hardest thing to walk back. A tool that hasn't yet proven it saves one team real time gets asked to prove it across twelve teams simultaneously, and the organizations that skip the narrow proof rarely go back and get it later.
2. Trust was assumed, not built
Most rollouts explain what the tool can do. Very few explain why its output can be trusted, or how a user is supposed to verify it before acting on it. Given that 46% of developers actively distrust AI-generated output, an adoption plan that doesn't address trust directly is missing the single largest reason people quietly stop using the tool.
3. No one owns the number after go-live
Usage is almost never tracked as a named responsibility with a review cadence — it's checked, if at all, at renewal time, which is the worst possible moment to discover a problem that started in month two. By the time it's someone's job to look, the number has already decayed past the point where a small intervention would fix it.
4. Governance arrived after the rollout, not before
A meaningful share of agent projects stall specifically on governance and security barriers discovered mid-flight rather than designed in from the start — the same gap this firm's Risk & Governance service exists to close, and covered in more depth there. The point worth making here is narrower: governance debt and adoption debt compound each other. A tool nobody trusts is a tool nobody uses; a tool nobody's watching is a tool whose governance gaps go unnoticed until they're expensive.
Healthy
Weekly active users by role
>70% of licensed users completing at least one AI-assisted task per week, 90 days after rollout.
AI domain
Watch
Output acceptance rate
% of AI-generated output used without material rework. Declining acceptance signals a trust problem, not a training gap.
AI domain
Watch
Escalation-to-human rate
How often an agent hands off rather than completes. Rising rates mean scope was set wider than the tool can reliably cover.
AI domain
Act now
Named owner for the usage number
If no one can name who reviews adoption monthly, the number is already decaying and nobody will notice until renewal.
AI & Governance domain
Signal thresholds are illustrative starting points, not universal benchmarks. What matters is the practice of watching one at all, not the specific number chosen.
What actually closes the gap
None of the four reasons above require a bigger AI budget to fix. They require someone independent of the deployment measuring what's actually happening, early enough to matter — the difference between finding out at renewal that a majority of seats went unused, and finding out at month three that the rollout was scoped too broadly to prove its own value.
Sources
- 1Digital Applied, 2026. Agentic AI Statistics 2026: 150+ Data Points Collection. Used for: 79% adoption vs. 11% production gap; 88% of AI agent projects never reach production, with failure split across infrastructure gaps (41%), governance/security barriers (38%), and ROI measurement failures (33%).
digitalapplied.com/blog/agentic-ai-statistics-2026-definitive-collection-150-data-points
- 2Unico Connect, 2026. AI Statistics 2026, Adoption, ROI & Impact. Used for: 29% of developers trusting AI-generated output today (down from 40% in 2024), 46% actively distrusting it.
unicoconnect.com/blogs/ai-statistics-2026
- 3Recon Analytics, 2026. Used for: 44% of lapsed Copilot users citing distrust of AI answers as their reason for stopping — the same figure already used in Relatio's core positioning materials.
via Relatio internal stat library
- 4Industry reporting on Microsoft Build 2026. Used for: the reported 64% of provisioned sidebar Copilot seats going unused, and this pattern's framing as a driver of Microsoft's agentic pivot announced at Build. Verify against original reporting before publishing — this is a fast-moving, contested narrative and deserves a direct-source check, not just secondary citation.
vaasblock.com/research/microsoft-build-2026-copilot-agentic-pivot-outage-reliability
- 5Gartner, Inc., 2025–2026 (primary press releases). Used for: 40% of enterprise applications expected to include task-specific agents by end of 2026, up from under 5% in 2025; over 40% of agentic AI projects expected to be cancelled by end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Independently corroborated by Reuters and Forbes coverage of the same releases.
gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps..., gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects...