80% see ROI, 40% will be cancelled: reading the 2026 agent adoption numbers

Two datasets landed within a few weeks of each other this summer and they cannot both be the whole story. Anthropic's 2026 State of AI Agents report says 80% of companies already see measurable ROI from agents and 81% are scaling multi-step agentic workflows. Gartner, in the same news cycle around July 21, predicts that more than 40% of agentic AI projects will be cancelled by 2027 because of cost, unclear ROI and weak governance. Both are true, and the reason is the most useful thing I have learned about agent adoption this year.

The numbers, side by side

Put the optimistic and pessimistic surveys in one list and a pattern shows up.

• 80% report measurable ROI from agents (Anthropic); 81% are scaling multi-step workflows.
• Only 29% of organisations see significant ROI from generative AI, and 23% from agents (Writer, Enterprise AI 2026).
• 41% of agent deployments pay back within 12 months, 18% within six; median time-to-value is 5.1 months.
• Engineers report saving about 9.4 hours per week, yet org-level value lags behind.
• 72% of companies have at least one AI workload in production, but 56% of CEOs report zero measurable benefit and 42% abandoned most of their AI initiatives last year.
• Teams that actually measure report 2.5-3.5x ROI on AI coding tools, with the top quartile at 4-6x.
• AI now writes roughly 61% of enterprise code.

The spread between 80% and 23% is not a sampling error. It is the difference between the population that instruments its agents and the population that ships them and hopes.

Why the split is about measurement, not models

Look at the 9.4 hours per week. That number is real; engineers feel it. But hours saved by an individual are not a line on a P&L unless someone attributes them to a shipped feature, a closed ticket, or a headcount decision. The 56% of CEOs seeing nothing are not lying and neither are their engineers. The measurement layer between them does not exist.

The 61% figure cuts the same way. If AI writes most of your code and nobody tracks what was reviewed, tested or reverted, you have generated volume, not value. Gartner's 40% is a forecast about exactly those teams.

The gap between 80% seeing ROI and 40% getting cancelled is not a model gap. It is the distance between teams that measure and teams that hope.

The one enterprise example in this window with a hard, auditable number is ServiceNow: AI annual contract value passed $1 billion, and more than 40 customers now resolve 80-85% of L1 IT requests autonomously with Now Assist. That is a metric a CFO understands, and it is why it turned into revenue.

What to measure before you scale

If you run agents in production or plan to, the checklist is short and unglamorous.

• Pick one business metric per agent (tickets closed autonomously, PR cycle time, cost per resolved case) and baseline it before rollout, not after.
• Log every agent run with model, prompt version, tokens and outcome. Without this you cannot compute the 2.5-3.5x, you can only assert it. A prompt versioning setup is enough for a small team.
• Track cost per outcome, not cost per token. Price times volume is what the invoice shows, and routing by task is where most of the cheap wins live.
• Set a kill criterion up front: if the metric has not moved by month five (the median time-to-value is 5.1 months), the project is a candidate for Gartner's 40%.
• Separate individual productivity (the 9.4 hours) from organisational value in your reporting. Conflating them is how you end up in the 56%. Claude Code admin usage tabs and similar dashboards are raw material for this, not the analysis.

The honest gap

Every figure above is a survey, and surveys of AI adoption have obvious selection effects: Anthropic is polling its own customers, Writer is polling enterprises with a reason to answer, and Gartner is forecasting. None of them share a definition of measurable ROI. I would not build a business case on any single number here. I would build it on the pattern, which is consistent across all of them: organisations that instrument their agents report returns, and organisations that do not report nothing. That is a finding you can act on without trusting any one vendor's arithmetic.