issues resolved per hour
The least experienced agents improved by 34%, suggesting the assistant helped most where patterns were learnable and answers were easy to check.
The Production Gap · A Concepta Research Feature
July 2026
AI has made writing software faster. Shipping it safely to production hasn’t gotten easier. That space — between code that’s written and a system that’s live, reliable, and trusted — is where most enterprise AI value is won or lost. We call it the production gap.
Adoption is no longer the question. Tools are everywhere; agents are entering the conversation. And yet the value keeps failing to arrive at the enterprise level. Most teams measure AI by how fast a person produces a draft, a code change, or an analysis. That’s a useful local signal — but it isn’t enterprise speed. Enterprise speed is the time from need to a change that is approved, deployed, adopted, and producing value. Across every serious study of the last three years, the research points to one consistent pattern.
What the evidence says
Five findings, one pattern
1
AI accelerates bounded, pattern-shaped work. Drafting, summarizing, test scaffolding, routine implementation.
2
It does not uniformly accelerate complex delivery. Mature systems still run on context, judgment, integration, review.
3
The bottleneck moves downstream. When creation gets cheaper, review, validation, and rework become the constraint.
4
Self-reported speed is inflated. People feel faster than the delivery data supports.
5
The winners redesign the workflow, not just the tooling. Value shows up when review, decisions, and release are rebuilt around the new speed.
Part01The state of play
McKinsey 2025 · BCG 2025
Adoption is broad. Scaled value is not.
Enterprise AI has passed the awareness phase. McKinsey reports that 88% of organizations use AI in at least one function, yet BCG finds only 5% consistently generating substantial value. Adoption is not transformation.
use AI in at least one function
Up from 78% a year earlierconsistently generate substantial value
BCG’s “future-built” companiesMcKinsey62%
experiment with agents: 23% scaling in at least one function and 39% experimenting.
McKinsey~33%
have begun scaling AI enterprise-wide; 39% report any EBIT impact, mostly below 5%.
McKinsey6%
qualify as “AI high performers” seeing significant value.
BCG35% / 60%
are scaling and seeing returns / report minimal revenue and cost gains.
Part02The mechanism
Theory of constraints
The production gap, defined
Enterprise speed isn’t set by the fastest step. It’s set by the slowest constraint on the path to production. AI compresses one step — the draft or the build. It does not automatically compress the rest of the path.
AI makes one step faster—not the system.
Request defined
Context supplied
more output reaches the same review gate
New work arrives faster than the organization can check and release it.
review
revision
approval
release
adopt
measure
The system moves only as fast as review clears.If review, approval, and release capacity stay fixed, faster creation becomes waiting work—and the 70% gain is absorbed before production.
Where the speed gets absorbed
Creation expands immediately. Review and release capacity do not—so more output becomes more work waiting.
CreationOutput volume expands
ReviewCapacity stays narrow
ProductionOnly trusted work ships
If AI cuts drafting time by 70% but review time doubles, delivery barely moves. Measure the whole path—including review, rework, coordination, governance, and adoption—then find the new constraint before it becomes a quality, cost, or trust problem.
Part03The upside
Brynjolfsson et al. · Peng et al.
Where AI speed is real
The strongest gains show up when work is bounded, repetitive, pattern-shaped, and easy to verify. Two randomized studies carry the causal weight.
time to complete the task
Developers using Copilot completed the same HTTP-server task in 55.8% less time than the control group.
The shared shape: clear inputs, checkable outputs, low ambiguity.
| Workflow | Why AI helps | What stays human-owned |
|---|---|---|
| Content and research draftsDocuments, synthesis, comparisons | Learnable patterns and text-based inputs | Voice, source quality, claims, recency and citations |
| Support assistSuggested responses and knowledge retrieval | Repeated questions and established policy | Escalation, empathy and exceptions |
| Engineering scaffoldingTests, documentation, routine scoped implementation | Repetition and reusable knowledge | Coverage, edge cases, integration, review and acceptance |
Part04The reversal
METR randomized trial, 2025
Where AI speed is overstated
Speed is least reliable where work is complex, ambiguous, high-risk, or context-heavy. The most important lesson comes from a setting that looks like it should favor AI: 16 experienced open-source maintainers, working on 246 real issues, in codebases they knew deeply.
+24%Expected before the trial
+20%Believed after the trial
−19%Measured completion speed
They expected AI to make them 24% faster. Afterward they believed it had made them 20% faster. Measured, they were 19% slower with AI allowed.
The same tool that speeds up boilerplate can slow down work in a mature codebase. Task type decides the outcome — not the model.
| Workflow | Why speed is unreliable | Better role for AI |
|---|---|---|
| Complex debugging | Root cause is unknown and context-heavy | Hypothesis generation, log analysis |
| Legacy modernization | Deep context, hidden business rules | Code explanation, inventory, test generation |
| Security-sensitive code | Working code can still be unsafe | Secure-coding checks, threat modeling |
| Enterprise integrations | Contracts between systems, sequencing | Dependency mapping, decision records |
| Strategy and architecture | Requires owned trade-offs | Scenario generation, evidence gathering |
Part05The constraint
DORA · Faros AI · LinearB
The bottleneck moves to review
The best-supported finding in the field is not that AI makes organizations faster. It’s that AI moves the constraint from creation to review and control.
+51%
Pull-request size
5×
Median review time
+54%
Bugs per developer
4.6×
Longer wait before review
PR acceptance rate — LinearB, 4,800+ organizations
Human-authored
84.4%
AI-assisted
32.7%
DORA 2024
As AI adoption rose, delivery throughput dropped about 1.5% and delivery stability about 7.2% — even as individual productivity and job satisfaction improved. The cause it points to: bigger batches and weaker testing discipline.
DORA 2025
Throughput’s relationship with AI turned positive — but AI still shows a negative relationship with delivery stability. Without strong testing, mature version control, and fast feedback loops, more change volume means more instability.
Faros 2026
Telemetry from 22,000 developers across 4,000 teams: PR size up ~51%, median review time up ~5×, bugs per developer up ~54%, and the incidents-to-PR ratio more than tripled.
LinearB 2026
Across 8.1M+ pull requests from 4,800+ organizations, AI-assisted PRs waited about 4.6× longer before review and were accepted far less often than human PRs — 32.7% versus 84.4%.
Part06The debt
GitClear · Veracode · Stack Overflow
Quality and security debt
AI can raise output volume before an organization has improved its quality system. The result is delivery debt: the growing backlog of checking, correcting, integrating, and maintaining AI-assisted work.
In 2024, duplicated code overtook reused code for the first time in the dataset’s history.
45%
of AI code samples introduced a vulnerability
Veracode, 100+ models. For cross-site scripting, models failed to defend in 86% of relevant cases.
66%
of developers frustrated by “almost right” answers
Stack Overflow 2025. 46% distrust AI accuracy; 33% trust it. 45% say debugging AI code takes longer.
Part07The perception trap
METR · Stack Overflow
Why teams feel faster before they are faster
AI changes how work feels. Blank pages disappear, search feels conversational, drafts arrive fast. That momentum is real — but felt speed and measured speed are different things. In the METR trial, developers believed AI sped them up by 20% while it was slowing them down by 19%. Stack Overflow’s data shows the same split: high usage, falling trust.
Perception+20%Developers believed they worked faster
Reality−19%Measured task completion was slower
Adoption tells you people will use the system. Only measured workflow data tells you it paid off.
Treat enthusiasm as an adoption signal, not an ROI signal
Part09The scoreboard
Measure the path, not the draft
Metrics that matter
Vanity metrics make AI look successful while the business sees nothing: prompts run, tools purchased, lines generated, drafts produced, self-reported time saved. Measure the path instead.
Request-to-approved-output cycle timeThe actual speed of the workflow
First-pass creation timeWhere AI helps locally
Review timeWhere speed gets absorbed
Rework rateThe hidden quality cost
Defect / incident rateKeeps speed from masking quality loss
Change stabilityWhether faster change breaks production
Adoption by workflowWhether teams actually use the system
Business outcomeThe tie to revenue, cost, risk, or CX
Part10The close
Concepta · governed delivery
Production confidence.
How Concepta helpsWe own the path from delivery risk to reliable release. AI speeds the work; senior technical leaders keep the system controlled, stable, and shippable.
20 yearsOwning critical systems
600+Projects delivered
98%Delivery satisfaction
Start here · 2–3 weeks · fixed scope
Delivery Risk Assessment
We map the request-to-production path, identify the delivery and release risks, and give you a clear ship, hold, or fix-first decision—with a sequenced stabilization plan.
Assess the release path →
When ownership is the gap
Technical delivery stewardship
Senior technical ownership through decisions, stabilization, and release.
When the system is the work
Business-critical delivery
Modernization, custom platforms, workflows, and integrations built to hold after launch.
Selected outcomes
+27%
Truist · fintech
4.7★
AdventHealth · healthcare
$3M+
FEMA · public sector
More new applications after rebuilding the virtual-loan experience.
Touchless care scaled in five months and earned 16,000 reviews.
Savings from a unified emergency-response supply chain.
AI accelerates the work. We own the release.
Build faster without losing control. That is the promise of governed delivery.
Appendix
12 primary sources
Evidence at a glance
| Finding | What the evidence shows | Sources |
|---|---|---|
| Adoption vs. value | AI is widely used but rarely scaled to real impact | McKinsey 2025, BCG 2025 |
| Speed on bounded work | Real gains on support and scoped coding tasks | Brynjolfsson et al. 2023, Peng et al. 2023 |
| Speed on complex work | Can slow experienced devs in mature codebases | METR 2025 |
| Review bottleneck | Constraint shifts to review, stability, and control | DORA 2024/2025, Faros 2026, LinearB 2026 |
| Quality and security | More output can mean more debt and vulnerabilities | GitClear 2025, Veracode 2025/26, Stack Overflow 2025 |
| Perception gap | People feel faster than the data supports | METR 2025, Stack Overflow 2025 |
How to read this evidence
- Not all evidence is equal. Randomized trials support causal claims; telemetry and surveys describe patterns and perception. Where a number moved between a working paper and a published version, we use the published one.
- Software evidence doesn’t transfer perfectly — but the pattern does. When creation speeds up, validation and governance become the constraint.
- Capability keeps moving. Capability gains don’t automatically become delivery gains — the operating model has to absorb them.
- Local speed can be real even when enterprise speed is flat. A team writing faster without shipping faster hasn’t failed. It has found the next bottleneck.
Sources
- McKinsey & Company. The state of AI in 2025: Agents, innovation, and transformation. Nov. 2025.
- Boston Consulting Group. Are You Generating Value from AI? The Widening Gap. Sept. 2025.
- Brynjolfsson, Li & Raymond. Generative AI at Work. QJE, 2025.
- Peng, Kalliamvakou, Cihon & Demirer. The Impact of AI on Developer Productivity. 2023.
- METR. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. 2025.
- DORA / Google Cloud. Accelerate State of DevOps Report 2024 and State of AI-assisted Software Development 2025.
- Faros AI. AI Engineering Report 2026: The Acceleration Whiplash.
- LinearB. 2026 Software Engineering Benchmarks Report.
- GitClear. AI Copilot Code Quality: 2025 Research.
- Veracode. 2025 GenAI Code Security Report and Spring 2026 update.
- Stack Overflow. 2025 Developer Survey: AI.