Evidence note 02 · AI agents
Which industries have entered the agent era?
In 0 industries, most newly documented AI deployments are already agentic. Energy & Utilities shows the fastest statistically significant rise.
Agentic systems account for 36% of the 1,263 enterprise AI deployments documented in the last 12 months, up from 31% across the 1,747 deployments of the 2023-era window. The average hides the real story: now publish more agentic deployments than classic AI, while the statistically significant surges belong to Energy & Utilities, Consumer & Food, Manufacturing.
My view is that the agent era will arrive workflow by workflow, long before an entire industry looks transformed. Manufacturing can lead in factory scheduling while maintenance remains anchored in classic machine learning. That unevenness is exactly what serious adoption looks like.
Majority-agent industries
0 of 12
Fastest significant riser
+31.1 pp
Energy & Utilities: 18% → 49% of new deployments
Corpus agent share
36%
456 of 1,263 deployments in the last 12 months, up from 31%
The finding
Four kinds of industries, one dividing line
Each row compares the agent share of an industry's newly published deployments in two windows. The digital-native sectors started high and stayed high. The surprise is who is crossing the 50% line now: Energy & Utilities jumped from 18% to 49%, with Consumer & Food, Manufacturing on the same trajectory.
The agent divide
Agent share of newly published deployments, by industry
Rows past the orange line publish more agentic deployments than classic AI. Solid connectors mark statistically significant shifts.
Automotive
Energy & Utilities
Professional Services
Manufacturing
Tech & Comms
Logistics
Consumer & Food
Insurance
Healthcare
Public Sector
Retail
Finance
| Industry | 2023 – Aug 2025 | Since Aug 2025 | Change | Significant |
|---|---|---|---|---|
| Automotive | 40.5% of 42 | 50% of 34 | 9.5 pp | no |
| Energy & Utilities | 17.5% of 57 | 48.6% of 35 | 31.1 pp | yes |
| Professional Services | 44.4% of 108 | 47.4% of 95 | 3 pp | no |
| Manufacturing | 30.5% of 236 | 44.4% of 81 | 13.9 pp | yes |
| Tech & Comms | 43.6% of 140 | 41.2% of 182 | -2.4 pp | no |
| Logistics | 25.5% of 51 | 40% of 40 | 14.5 pp | no |
| Consumer & Food | 14.8% of 61 | 37.9% of 66 | 23.1 pp | yes |
| Insurance | 28.9% of 135 | 34.5% of 55 | 5.6 pp | no |
| Healthcare | 31.1% of 212 | 33.9% of 192 | 2.8 pp | no |
| Public Sector | 37.6% of 93 | 33.9% of 62 | -3.7 pp | no |
| Retail | 33.3% of 114 | 30.9% of 81 | -2.4 pp | no |
| Finance | 28% of 168 | 28.2% of 170 | 0.2 pp | no |
Interpretation
How to read the divide
- 1
The second wave is the physical economy.
Every statistically significant riser operates physical assets or physical flows: Energy & Utilities (18% → 49%), Consumer & Food (15% → 38%), Manufacturing (31% → 44%). Software and consulting firms adopted agents first, but the growth is now on factory floors and in supply chains.
- 2
Regulated back-offices move late — but deep.
Finance, Retail, Healthcare sit well below the corpus average, and none of their shifts clear the significance bar. Where these sectors do deploy agents, they go deep: 48% of Finance's recent agent deployments orchestrate multiple agents. Late adoption can still mean ambitious deployment.
- 3
Industry averages hide the real divide.
At sub-industry resolution the same sector holds both extremes: Healthcare's Digital Health runs 35% agentic while Diagnostics and Imaging sits at 11%; Manufacturing's Factory Operations runs 32% agentic while Maintenance and Reliability sits at 10%; Finance's Retail Banking runs 36% agentic while Financial Crime and Fraud sits at 14%. The laggard subsegments are the vision- and sensor-heavy workloads where classic machine learning still fits the job.
Conclusion and outlook
The next divide will be inside companies
The market will soon care less about which industry crossed 50% first and more about which operating teams can govern agents at scale. Watch the subsegments: once the low-agent parts of a sector begin to rise, agentic systems have moved beyond isolated workflow wins and into the operating model.
Method
Controls and limits
Windows are built from each case's original publication date: the trailing 12 months versus 2023 up to the start of that window. An industry is compared only with at least 30 published cases in both windows — Education (28 recent), Pharma (28 recent), Agriculture (18 recent), Real Estate (16 recent), Legal (8 recent) fall below the bar and are excluded. Shift markers use a two-proportion z-test at 95% confidence. Sub-industry shares pool all years and require 40 cases.
These figures measure the visibility of agents in published deployment evidence, not production adoption: vendors choose what to publish, and the mix of sources feeding the corpus shifts over time. Treat the cross-industry ordering and the significant shifts as the finding; single percentage points are noise.