Ninety percent of engineering leaders report productivity improvements from AI coding agents. The same Gartner research predicts that over 40% of agentic AI projects will be canceled by end of 2027 — with runaway costs as a primary driver. Both statements are true at once. That tension is where most enterprise teams are operating right now, and understanding it is the difference between the organizations achieving 17x ROI and those burning through annual budgets in four months.
The productivity numbers — and what they’re hiding
Gartner’s 2026 Market Guide for Enterprise AI Coding Agents puts the net average productivity gain at 19.3% across organizations that have adopted these tools. Ninety percent of engineering leaders say they see improvement. That’s a headline worth celebrating — until you look at what “net” is absorbing.
IBM’s “Bob” AI coding agent, deployed to 80,000 IBM employees, delivered a 45% average productivity gain — with specific teams reporting 69–70% time savings. A Fortune 100 retailer documented 17x ROI from GitHub Copilot Enterprise: 450,000 hours saved, $33.75M in measurable savings against a $1.9M licensing cost. These results are real — and they’re what structured, governance-first rollouts look like.
The counter-signal: a METR study found experienced developers were 19% slower with AI assistance despite perceiving themselves 20% faster. A Faros AI study of 20,000 developers showed rising output accompanied by rising bugs and rewrites. According to Uvik Software’s 2026 AI Coding stats, code churn rose from 3.1% in 2020 to 5.7% in 2024; code duplication jumped from 8.3% to 12.3% of changed lines. The productivity dashboard looks green while the quality debt accumulates below the surface.
The budget trap hiding inside consumption-based pricing
The pricing shift from per-seat licenses to token consumption is where most enterprises get caught. TechCrunch reported in June 2026 the cases that have become cautionary tales across the industry: Uber onboarded 5,000 engineers in December 2025 and burned through its entire 2026 AI coding budget by April — a roughly 3x overrun. Priceline saw its Cursor renewal come back 4–5x more expensive than the prior term. One unnamed enterprise accumulated a $500M Anthropic bill in a single month from agentic workflows running without cost caps.
The data at scale confirms these aren’t outliers. 85% of companies miss AI cost forecasts by more than 10%, and nearly 25% underestimate costs by 50% or more, according to Keyhole Software’s 2026 AI cost survey. Per-developer token consumption rose approximately 18.6x in nine months. The root cause is structural: finance teams applying static, license-based procurement models to dynamic, consumption-based costs. The AI governance layer that prevents this isn’t optional infrastructure — it’s what the successful cases have in common.
Trust is eroding as adoption scales
Adoption is accelerating. GitHub Copilot is deployed at roughly 90% of Fortune 100 companies; Cursor hit $2B ARR in February 2026. But the same 2026 AI coding stats show that only 29% of developers trust AI coding outputs for accuracy — down from 40% in 2024. Sixty-six percent report frustration with code that appears correct but contains subtle errors. AI-coauthored pull requests show 2.74x more security vulnerabilities than human-only PRs.
This isn’t a model capability problem. It’s a process gap. Teams that treat AI-generated code with the same rigor they’d apply to a third-party library — mandatory review, security scanning, provenance tracking — see fundamentally different outcomes than teams that use it as a text autocomplete with higher stakes.
What separates the 17x ROI teams from the rest
High-performing organizations reach pilot-to-production in roughly 90 days; struggling ones take nine months or more — and only 5% of task-specific GenAI pilots reach successful implementation at scale. The difference isn’t the model selected. It’s the governance layer built around it.
The patterns that consistently appear in successful deployments:
- Governance before scale: IBM’s Bob succeeded because the rollout addressed model selection policies, identity systems, cost management, and observability before expanding to all 80,000 employees — not after. EY achieved 4x coding productivity by connecting AI agents to internal engineering standards from day one.
- Cost observability as a product: Teams using AI gateways for cost governance report 40–60% reductions in inference costs, according to multiple cost governance reports. Treating token spend like cloud infrastructure spend — with dashboards, budgets per team, and automated circuit breakers — prevents the Uber scenario from repeating.
- Redefining developer roles, not eliminating them: The coding rate limiter today isn’t code generation — it’s requirements clarity, system integration, and post-production maintenance. Teams that redeploy developer time toward these activities get compounding returns; teams that just measure lines of code generated see churn go up.
The bottom line heading into the second half of 2026
The AI coding agent market tripled in 16 months to reach $9.8–11B annualized as of April 2026. IDC projects a tenfold increase in AI agent usage among G2000 companies by 2027. The technology is not going away, and the productivity gains are real for teams that earn them. But Gartner’s prediction that 40%+ of agentic projects will be canceled by 2027 is a forecast, not a destiny — it describes what happens to teams that scale adoption without scaling governance. The question is not whether to use AI coding agents. It’s whether your organization is building the infrastructure to make that bet pay off.