The telemetry became the report
A vendor published an industry index built from its own product data. Read the denominator before the numbers — and then read the one number that survives the discount.
Salesforce has published an Agentic Enterprise Index drawn from anonymised Agentforce data: deployment up threefold, 57% of organisations running multi-step agent workflows, 16% with cross-functional agents, 81% planning to expand.
The denominator
This is telemetry from customers of one agent platform. The 57% figure describes companies that already bought an agent product and then used it — a population selected precisely for the behaviour being measured.
That is not a gotcha, it is a reading instruction. The number is real and it does not mean what a survey of enterprises generally would mean.
Why it still beats the alternative
Platform telemetry has one decisive advantage over self-reported surveys: it counts what was deployed rather than what someone said they intended to deploy. The gap between those two has been the defining measurement failure of enterprise AI for three years running.
Read the direction and discount the level.
The number that survives
Sixteen per cent. Cross-functional agents spanning multiple teams is the genuinely hard part — it requires permissions, data access, and organisational agreement that single-team automation does not need. That only one in six has reached it, inside a population already committed and already paying, is the realistic picture.
Everything above that line is easier than it sounds. Everything below it is where the value was supposed to be.
The integrators agree with the pessimistic reading
Cognizant has stood up a dedicated EMEA agentic-AI unit with three tiers — strategy, rapid prototyping, and multi-agent production squads. Splitting those into separately priced services is what you do after watching enough engagements die at the seam between prototype and production.
The tiering is a diagnosis dressed as a price list. It says the hard part is not building an agent; it is the transition from one that demos to one that runs, and that transition is expensive enough to sell separately.
What the region adds
EMEA is also where the AI Act now has demonstrated enforcement. "Build, deploy and run" means something heavier there than in North America — running an agent in production carries transparency obligations somebody has to own, and integrators are positioning to be that somebody.
Consultancies do not staff units for demand that does not exist. This is not evidence agents work. It is evidence that enough enterprises are committed to finding out that the finding-out is billable.
Wazzup — Enterprise Agent Deployments More Than Double — Salesforce 2026 Index Insights → · Arcade.dev — State of AI Agents 2026: 5 Enterprise Trends → · AI Agent Store — AI Agents News — Week of August 15, 2026 →