Salesforce's Agentic Enterprise Index Says 5 Agents Became 13. Who Was Counted?

Salesforce's second Agentic Enterprise Index says the average organisation in its dataset ran 13 activated agents in April 2026, up from 5 in February 2025. That is a 2.6x rise in 14 months, and it is real platform telemetry rather than a survey. It is also a count taken only from companies that kept agents live every single month, which changes what the number can tell you.

This post is written for the founder or director who is deciding how much agent capacity to plan for next year and keeps meeting vendor benchmarks in board decks. That reader needs to know which number can carry a plan and which cannot.

The short answer. The Salesforce Agentic Enterprise Index measures how many agents a committed Salesforce agent customer runs, not how many enterprises use agents. Its cohort only includes businesses live every month for 14 months. McKinsey and KPMG survey executives instead, so their adoption shares cannot be compared with Salesforce's agents-per-organisation count.

What the Salesforce Agentic Enterprise Index actually measures

The index is Salesforce's periodic read of usage on its own agent products. The second edition was published on 7 August 2026, under the headline "Agent Deployments More Than Double Year over Year". It covers the 14 months from February 2025 to April 2026.

Its central number is a level, not a share. The average number of activated agents per organisation rose from 5 to 13. Alongside it, Salesforce reports that agent creation time fell 53% to an average of 1.9 days, and that escalations from agents to humans are "holding steady at 32%".

Whose data it is

The index uses Salesforce's own usage data. The landing page says Salesforce "analyzed and aggregated usage data of a cohort of businesses", powered by Agentforce and other Salesforce products. The newsroom story calls it "aggregated AI usage data from the Agentforce platform".

That makes it observational telemetry, which is a genuine strength. Nobody had to remember or estimate anything. It also means every business in the dataset is a Salesforce customer running Salesforce agents. Companies building agents on other platforms, or in-house, are outside it by design.

A separate survey layer sits on top. Salesforce says additional data came from 4,689 survey responses collected in May 2026 across seven countries. The shopper trust figures on the page come from that layer, not from telemetry, and the page does not always say which number comes from where.

The 5 to 13 arithmetic

Thirteen divided by five is 2.6. Salesforce's body text calls this "nearly 3x", and its headline calls it "more than double year over year". Both can be reconciled once you notice the window is 14 months, not 12.

Salesforce also reports a 7% compound monthly growth rate in agent production. Compounding 7% over 14 months gives 1.07 to the 14th power, which is about 2.58. That reproduces the 2.6x total almost exactly. Scaling the same rate to 12 months gives roughly 2.3x, which is what "more than double year over year" should mean.

So the three phrasings are consistent with one another. In my view the headline is the most careful of them. "Nearly 3x" and the third-party "nearly tripled" both round 2.6 upward and drop the extra two months.

The survivorship filter inside the index methodology

One sentence in the methodology decides how the whole index should be read. "To qualify for inclusion in the dataset, businesses needed to have activated agents in production every month across the analysis period."

Every month, for 14 months

That rule means the cohort is fixed at its start. A business had to be live with agents by February 2025, which was early in Agentforce's commercial life, and stay live through April 2026. Anyone who started later is excluded. Anyone who paused for a month is excluded too.

This is a sensible design for one question. If you want to know how usage deepens inside committed customers, a constant cohort stops new arrivals from dragging the average down. It is the wrong design for the question most readers bring to the headline, which is how fast enterprises are adopting agents.

What the filter removes

Three groups are missing by construction. Late adopters are missing, because they were not live in February 2025. Churned or paused customers are missing, because they broke the every-month rule. Non-Salesforce agent users are missing, because the data comes from Salesforce products.

Each of those groups would probably run fewer agents than the survivors. So the 13 is best read as the ceiling of what a persistent early adopter does, not the middle of the market. When this blog tallied the ledger of cancelled agentic AI projects, the cancellations were exactly the kind of record a survivor-only cohort cannot contain.

I do not read this as a trick. Salesforce printed the rule in plain text. The problem is downstream, when "agent deployments more than double" travels without the sentence that defines whose deployments.

What counts as an agent, and what the index leaves undefined

A methodology note earns trust by defining its units. The Salesforce index defines one unit carefully and leaves the others open. The table below lists each term the headline depends on and what the published pages say about it.

Based on the second-edition landing page and the 7 August 2026 newsroom story, retrieved 4 November 2026. "Not defined" means not defined on those public pages.

The undefined "agent" is the most consequential gap. If a business splits one service bot into three topic-specific agents, its count triples while its workload does not change. Salesforce's own skills metric points at this: unique skills per agent rose from 2 to 6 across 2025, so agents were also getting bigger.

Counting things that are loosely defined is the same failure this blog flagged in the case against the AI use-case count. A rising tally of agents is good news only if each agent is roughly the same size each time you count.

How the first and second editions of the index differ

The first index was published on 17 September 2025 and covered roughly the first half of 2025. Its headline was 119% growth in agent creation among first-mover companies between January and June. Reading the two editions together shows what changed in the method as well as the numbers.

The extrapolation sentence disappeared

The first edition's methodology said that "several factors are applied to extrapolate macroeconomic figures for the broader market". The second edition's methodology note does not contain that sentence. It keeps the same inclusion rule and the same disclaimer that results "are not indicative of Salesforce performance".

That edit points in the right direction. A survivor cohort is a poor base for market-wide figures, and the second edition no longer claims to produce them. It would still help if Salesforce said so explicitly, because older market-sized claims from the first edition are still being quoted.

Escalations stopped falling

The escalation rate is the most useful operational number in either edition. In the first, escalations from agents to humans rose from 22% in Q1 2025 to 32% in Q2 2025. In the second, they are "holding steady at 32%".

Read plainly, about one in three agent conversations still ends with a human. That matches the second edition's claim that agents solved 7 in 10 service chats without help. Agent counts nearly tripled over the window, and the hand-off rate did not improve. For a director, that is the number to plan headcount around, not the 13.

McKinsey and KPMG ask a different question about enterprise AI agents adoption

The two independent surveys most often set beside vendor data are McKinsey's annual State of AI and KPMG's US AI Quarterly Pulse. Both ask executives about their organisation. Neither counts agents. Both have methodology quirks of their own, and they deserve the same reading the Salesforce index got.

McKinsey counts respondents, not companies

McKinsey's 2025 survey, fielded from 25 June to 29 July 2025 across 1,993 participants, found that 23% of respondents said their organisation was scaling an agentic AI system somewhere. Another 39% said they had begun experimenting. No more than 10% reported scaling agents in any single function.

The 2026 edition, fielded from 4 May to 8 June 2026 with 1,719 participants, split the result by size. At organisations with $1 billion or more in revenue, the share scaling agents in at least one function rose from 27% to 40%, while smaller organisations stayed at 22%. That size gap is examined in the breakdown of McKinsey's large-firm agent gap.

The unit is the respondent. McKinsey surveys people at all levels, so two respondents from one company count twice. "Scaling in one function" also treats one production agent and fifty production agents the same.

KPMG's series changed its own wording

KPMG's pulse surveys US leaders at companies with $1 billion or more in revenue. Through 2025 the sample was 130 respondents per quarter. The deployed share went from 11% in Q1 2025 to 33% in Q2 and 42% in Q3. Then, in the Q4 2025 release, it fell to 26%.

KPMG explained the drop by saying leaders had adopted "more sophisticated definitions of what constitutes a true agent". In other words, respondents changed their minds about what the word meant. That is the clearest public evidence I have found that agent adoption figures move with definitions as much as with behaviour.

By Q3 2026 the question had changed too. 62% were "building, deploying or developing" agents, up from 53%, on a larger sample of 314. Building and developing are not deploying. The series hub page uses a third phrasing, "building, testing, or deploying", for the same figure.

Putting the three AI agent deployment statistics side by side

Placed in one table, the three sources stop looking like disagreeing measurements of one quantity. They are measurements of three different quantities, taken on three different populations, with three different filters.

McKinsey figures are from FM Magazine's report of the 2026 survey and McKinsey's 2025 summary; the full McKinsey document did not load during research. All rows retrieved 4 November 2026.

The practical reading follows from the matrix. Use the surveys for "how many firms like mine are doing this". Use the Salesforce index for "how deep does it go once a firm commits to Salesforce agents". Never use one to answer the other's question.

The same separation explains why vendor revenue disclosures and adoption surveys diverge. This blog's ledger of Agentforce ARR disclosures tracks what customers pay, which is a third quantity again.

Where this argument is weakest

This post has spent 2,000 words on the index's filters. It should concede where the index is better than its critics, this post included.

Telemetry beats recall. Surveys ask an executive to describe a system they may not run. KPMG's 42% to 26% swing shows how far recall and definitions can move a number in 90 days. Salesforce's figure does not depend on anyone's memory. On the narrow question it answers, it is likely the most accurate number of the three.

The surveys have filters too. KPMG samples only US firms above $1 billion in revenue. McKinsey weights and reports respondents rather than companies, so firms with many respondents carry more weight. Screening an adoption survey to large companies lifts its figures in much the same way a survivor cohort does, just less visibly.

The cohort may be larger than critics assume. Salesforce does not publish the count, and I have not been able to find it. If the cohort runs to thousands of businesses, the average of 13 is a stable figure for that group. My criticism is about scope, not precision, and I cannot rule out that the figure is very precise.

The direction agrees across all three. Telemetry rises, McKinsey's large-firm share rises, and KPMG's broad measure rises. When three methods with different biases move the same way, the direction is probably real. What is unresolved is the level, and the level is what a budget needs.

How to use a vendor index in your own agent planning

A vendor index is a benchmark for customers who look like its cohort. If you are a Salesforce customer already running agents, 13 agents and a 32% escalation rate are a fair reference point for where a committed peer sits after 14 months. If you are not, the number describes someone else.

The useful habit is to translate every headline into the question it answers before it enters a plan. "Agents per surviving customer" goes in a capacity estimate. "Share of executives scaling" goes in a competitive read. Neither goes in a revenue forecast. This is the same discipline argued for in the review of how AI ROI surveys are built.

I would also weight the escalation figure above the agent count in any staffing model. An agent count can grow by splitting work into smaller agents. A hand-off rate stuck at 32% means a third of conversations still need a person, and that is the line that shows up in your payroll.

When a vendor sends you an index, ask for four things in writing: the cohort size, the inclusion rule, the definition of an agent, and the churn rate of customers who left the cohort. A vendor that answers all four has measured something you can use. Comparing that answer with the vendor's peers is covered in the checklist for reading AI agent adoption surveys.

Frequently asked questions

Where to start this week

Find the last agent statistic that appeared in a board deck or budget paper at your company. Write next to it the unit it counts, the population it came from, and the filter that decided who was included. If any of those three cannot be filled in from the source, remove the number from the deck.

If you run Salesforce agents yourself, pull your own escalation rate for the last full month and set it beside Salesforce's 32%. That single comparison tells you more about your agent programme than the cohort average of 13 ever will.

Related reading