Engineers Use AI Agents Daily at 80%. Enterprise Production Deployment Is Stuck at 9%
Temporal's 2026 survey found 80.8% of engineers use AI agents daily, but Ness Digital Engineering found only 9-14% of companies have agentic AI in production. Here's why both numbers are true at once.
Contents5
Two reports, two numbers that seem to contradict each other
Two studies landed within days of each other in August 2026, and read side by side they expose the agentic AI production gap in plain numbers. Temporal's State of Development 2026 survey found that 80.8% of engineers now use AI agents daily, up from 47.3% a year earlier, a 70.8% relative jump. Separately, Ness Digital Engineering reported that 99% of companies plan to deploy agentic AI, but only 9-14% have actually gotten it into production.
Both numbers are real. They're not measuring the same thing, and the gap between them is the most useful story in either report.
What Temporal's 554 engineers actually said
Temporal surveyed 554 engineers and technical leads between April 29 and May 25, 2026, weighted toward the US (67%) with the rest in the UK and EMEA, mostly mid-career people with six to fifteen years of experience. The headline numbers:
| Metric | Value |
|---|---|
| Use AI agents daily or more | 80.8% (up from 47.3% a year ago) |
| Trust agent output at least somewhat | 85.5% |
| Completely trust agent output | 24.7% |
| Report issues with agent output daily | 41.1% |
| Cite security concerns as the top blocker to full autonomy | 39.5% |
| Go from AI-generated prototype to production-ready code within hours | 51.3% |
| Reach production-ready within days | 77.1% |
That's a genuinely high-trust, high-frequency picture. It's also, importantly, a survey about individual engineers using coding agents inside their own workflow, mostly IDE and CLI tools like the kind covered in [how to cut your AI agent's API bill](https://questloops.com/blog/how-to-cut-your-ai-agent-s-api-bill-token-compression-smart-routing-and-when-a-gateway-pays-for-itself). One engineer approving one agent's pull request is a low-stakes, reversible decision. Nobody outside the team needs to sign off on it.
What Ness Digital Engineering's report is actually measuring
The Ness report describes something structurally different: the chasm between an approved agentic AI pilot and an agent running unsupervised in production, executing multi-step business processes with real customers or real money attached. Ninety-nine percent of companies surveyed plan to deploy this kind of system. Only 9-14% have.
Ness names three specific reasons projects stall in what it calls "Death Valley":
- Trust in probabilistic systems is thin. A wrong output from a coding agent gets caught in code review; a wrong output from an unsupervised customer-facing agent is a live incident.
- Pilots often produce no visible change in daily operations. If nobody can point to a workflow that got measurably better, nobody champions the project past proof-of-concept.
- Infrastructure usually isn't ready. Domain data, legacy systems, and API access frequently aren't in a state that supports an agent acting without a human in the loop.
Financial services splits along a predictable line in the report: traditional institutions lean on agentic AI for back-office and internal IT work, where mistakes are contained, while fintechs are more willing to expose it directly to customers.
The agentic AI production gap is the real story
Line the two reports up and a pattern falls out that neither report states directly: agent adoption is moving fast at the individual-task layer and barely moving at the autonomous-system layer, and those are different products with different failure costs.
A coding agent that gets something wrong costs one engineer a few minutes in review. An unsupervised agent that gets something wrong in production can cost a support queue, a compliance filing, or a customer relationship. Temporal's own data hints at this: 39.5% of engineers who trust their agents daily still name security as the reason they won't hand over full autonomy. High trust in a supervised tool doesn't transfer to trust in an unsupervised one, and it shouldn't.
What this means if you're planning agent rollout
If your team is stuck between a successful pilot and a stalled production rollout, the two reports together suggest a specific fix, not "try harder": keep a human in the loop for anything that isn't fully reversible, and measure the pilot against a visible, specific operational metric before asking anyone to sponsor its move to production. "It works in the demo" is not the bar Ness's report says companies are failing to clear. Visible, measured impact on a real daily process is.
Engineers trust and use AI agents constantly now, at a scale that would have sounded aggressive a year ago. Enterprises still don't trust agents running unsupervised with real consequences attached, and that gap isn't closing because people are slow. It closes, or doesn't, one specific, measurable, reversible deployment at a time.