Evidence note 01 · AI agents
What changes when AI starts acting?
Agent cases report automation about twice as often as other AI cases. Multi-agent deployments carry the strongest speed and throughput signals.
Across 2,427 documented AI deployments published since 2024, measurable-outcome coverage stays close: 54.9% for other AI, 55.3% for single-agent systems, and 51.4% for multi-agent systems. After adjustment for provider and industry mix, automation appears in 13.9% of single-agent outcomes and 8.2% of other AI outcomes. Multi-agent evidence reaches 65.1% for time and speed and 18.7% for productivity and throughput.
My interpretation is that agents earn their keep when a workflow has enough handoffs to make coordination itself valuable. The strongest signal sits in automation and throughput because those are the outcomes produced when software can carry work across steps, instead of stopping after an answer.
Automation evidence
1.7×
13.9% single-agent vs 8.2% other AI
Multi-agent speed signal
+9.5 pp
65.1% of outcome-bearing multi-agent cases
Multi-agent throughput signal
+6.5 pp
18.7% report productivity or throughput
The finding
Agent evidence concentrates on automation, speed, and throughput
Automation and deflection appear in 8.2% of standardized non-agent outcomes. The share rises to 13.9% for single-agent cases and 14.7% for multi-agent cases. Multi-agent deployments register the largest shares for time and speed (65.1%) and productivity or throughput (18.7%).
Outcome fingerprint
What each deployment type reports
Share of outcome-bearing cases mentioning each category. Categories can overlap.
Automation & deflection
Time & speed
Productivity & throughput
Cost savings
Quality & accuracy
Adoption & scale
| Outcome | Other AI | Single-agent | Multi-agent |
|---|---|---|---|
| Automation & deflection | 8.2% | 13.9% | 14.7% |
| Time & speed | 55.6% | 57.2% | 65.1% |
| Productivity & throughput | 12.2% | 14.1% | 18.7% |
| Cost savings | 18.8% | 12.9% | 14.3% |
| Quality & accuracy | 16.4% | 9.4% | 11.6% |
| Adoption & scale | 4.6% | 9.8% | 5.9% |
Interpretation
How to read the fingerprint
- 1
Automation is the clearest agent signal.
Automation and deflection are roughly twice as prevalent in agent outcome evidence as in other AI evidence.
- 2
Multi-agent evidence centers on speed and throughput.
Time and speed appear in 65.1% of multi-agent outcomes; productivity and throughput appear in 18.7%. These systems often coordinate several steps in one workflow.
- 3
Treat the chart as an evidence profile.
The chart tracks categories reported in public case studies. Causal impact, effect size, and success rates require other evidence.
Conclusion and outlook
Where this goes next
The next question is whether these systems can keep the speed signal as controls, audits and exception handling catch up. I expect the useful frontier to move from impressive single tasks toward dependable end-to-end work. Teams that measure handoff time, intervention rates and recovery from errors will learn more than teams counting agent launches.
Method
Controls and limits
The chart covers outcome-bearing cases published since 2024. Multi-agent classification takes precedence over single-agent, and Copilot-only cases are excluded from the adjusted chart because common support is still too thin. The displayed rates are standardized across 12 provider-by-industry strata, representing 521 comparable outcome-bearing cases. Every included stratum has at least 5 cases in each displayed stage.
Public case studies over-represent successful deployments, and providers differ in what they publish. Standardization addresses provider and industry composition. Selection bias remains, and the figures come from public claims without independent audits. Categories can overlap because one deployment may report several kinds of value.