Licensed to multiply
Agents can do more than you think.
That’s a message that is lifted from a new study from Salesforce – the Agentic Enterprise Index, which sets out as its goal to analyze aggregated AI usage data from the Agentforce platform to uncover how businesses are deploying, using, and getting value out of AI agents.
It’s also one that I suspect we will be hearing repeated a lot of times in various forms in the coming weeks as the firm gathers its forces for the annual Dreamforce jamboree, an event pitched by Joe Inzerillio, President Enterprise and AI Technology as the company’s “Superbowl”.
Two years ago CEO Marc Benioff tore up his prepared presentation for the event’s keynote to pivot the entire company around agentic AI. Last year was about a progress report and further proving his thesis around the technology was correct. This year? Watch this space.
But while the first year was powered by a near evangelical fervor around the new tech’s transformative potential and the second by early impressive use cases among selected customers, this year’s may well be a bit trickier. While the percentage growth rates for Agentforce adoption have been impressive, in terms of actual user numbers, the overwhelming majority of the Salesforce customer base remains novices – and Wall Street is restive.
At the same time there’s a new air of realism, pragmatism and – whisper it softly – skepticism about AI out there in the wild. At this point, we have to nod to the wretched MIT study from last summer that found that alarmingly high numbers of AI investments had not delivered – or were not perceived to have delivered, which might be an important distinction – the value that was expected.
Several similar studies have followed since and, with the added complication of tokenomics shock bills coming in to startle Finance officers everywhere, there are increasingly large numbers of stories of organizations not so much abandoning their AI journeys, but certainly hitting the brake pedal to slow down.
That’s not a situation that is helpful for Salesforce. Later this month, prior to Dreamforce, the firm will issue its latest update on Agentforce usage adoption as part of its quarterly finance reports. A lot of attention will be paid to those numbers.
In the meantime, a message that says agentic potential has not yet been tapped into fully is clearly highly useful.
So what does it say?
Some context first. What is the Salesforce Agentic Enterprise Index based on? It looks at trends between February 2025 and April 2026, the latter being roughly around the time when the tokenomics crisis kicked in. The Index looks at “the activity and engagement of real businesses leveraging the power of AI agents to drive ROI.”
Unlike most Salesforce studies, the demographic breakdown of those polled isn’t as publicly clear-cut as it usually is. To qualify for inclusion in the dataset, businesses needed to have activated agents in production every month across the analysis period. Additional data was sourced from Salesforce’s global proprietary research studies, totalling 4,689 responses in May 2026. Survey audiences include the USA, UK, France, Canada, Australia, Spain, and Italy.
All that said, the study, unsurprisingly, is overwhelmingly ‘glass half full – and rising’. It finds that organizations increased activated agents by nearly 3x by the end of this fiscal year and reduced the average creation time by 53%.
Meanwhile agents are becoming highly versatile, with their skill sets expanding by up to 350% to enable them to handle increasingly complex tasks during peak demand. While most routine work would only require an agent to be able to perform a couple of simple, automated tasks, on average agents today can act on up to six skills, says the Index, up from two when the study period began.
That said, it’s worth noting that across the entire study base, the number of unique skills per agent hit its high of 6.4 in November last year – just after the 2025 Dreamforce and at the start of the Holiday season in the US – but has since fallen to 3.7. Why? The Index report doesn’t comment on this, but is this perhaps indicative of that new sense of ‘slow down, where’s the value in what we’ve got?’ mentality kicking in?
(That Holiday season timing for the high is also surely attributable to the Retail sector’s enthusiasm for agentic AI? Among Retail & Consumer Goods firms polled, the average number of activated agents went from five in April 2025 to 11 a year later!)
More and more and more
So what are we to make of all this? Over to Inzerillo who declares confidently:
There’s just definitively more agents in production now. I think last year was really the year about doing a lot of talking about agents [and] people were definitely experimenting, but now we’re seeing real traction in production situations, and we’re seeing a lot more – over three times the number of agents that we saw five quarters ago.
That widespread concern around value derivation is genuine, but Inzerillo argues it’s not the only factor to be considered:
Obviously a big part of it is delivering value. I think that’s a really, really important thing…But I also think it cuts into the situation [whereby] it’s actually just getting easier to build. It’s easier to deploy…We’re definitely getting to the point where you can see this real reduction in the time it takes to essentially initiate an agent and start to get it to the point where you can get it ready for production, and so now less than two days is [how long] we see the average person is able to get their their agent in a situation where they can start to really make it fit for purpose.
And the nature of the agents being produced is changing for the better and the more practical, he suggests:
What we saw five quarters ago was a lot of informational agents; now we’re seeing a lot of actions that agents are taking, being able to do things on behalf of the customer, which I think is really interesting.
Start with Service!
The Customer Service use case remains the most common one industry-wide. This makes perfect sense to Inzerillo:
I did my agentic transformation a couple of years ago when I was at SiriusXM [where he was Chief Product and Technology Officer]. We did a service transformation. When you think about service, most companies are outsourcing this to call centers and things like that. It’s generally not the most skilled employee at the company that are doing it, so there’s a lot of stuff that’s on the road.
This is the perfect starting point, he argues:
To me, every single company should have an agent there because they’re going to give you more consistent answers less expensively, with a lot more capabilities as you start to add more and more complexity. So that is by far and away I think the absolute best, service use cases are absolutely the best ROI to start with.
He adds:
I can’t emphasize enough for the external customers how important the Service use case is because it’s just it’s exactly where to start. You’ll see the ROI on it, and it’s also a great place to experiment where you can. Most people have a decent amount of traffic going to the Support side, so you can take five percent and run an experiment to try to introduce some new features and functionality and see how that works in production, which I definitely recommend. The multi-variate testing in production is the way you continue to make these agents better and better and better.
Salesforce’s own ‘Customer Zero’ policy here of ‘try it yourself before you ship it to punters’ has thrown up its own lessons on this front., he adds, People are becoming more aware of what agents could do for them:
You start to ask for more and more things. Five quarters ago, it was a lot of Q&A, like, ‘How do I do this?’ or “Where can I go find that?’. Now what you’re starting to see people do is very action-oriented. So instead of asking our Employee Agent, ’How do I file a form to take out my vacation?;, they’re telling the Employee Agent, ‘Hey, I’m taking a vacation. You need to enter this form for me, and here’s the details’. That bias towards action is really what we’re seeing the evolution go to
He goes on:
The thing about this that’s also really exciting is that you’re now starting to see that seven out of 10 interactions are handled by the agents. But what you can see here is that’s been very consistent over the last five quarters, even though the volume has really just exponentially started to lift. I think that’s a testament to the volume is there because people are getting good results out of it by and large.
He points to the Retail sector here:
You’re talking about 4x improvement of the sales growth when a shopper agent is present, and so shopper agents are are affording people an easier time buying, and they buy more. I buy more, buy more often, and I think these are both really interesting things.
My take
An interesting snap shot to set the scene for a deeper exploration of the bigger picture as Dreamforce approaches, and one to which we will undoubtedly be returning on the ground at the event. These research findings are interesting, certainly, but let’s road test them against customer sentiment among the assembled ‘Ohana’.
The other question I have is around the Agentic Work Unit (AWU) metric and how well this is resonating among the Salesforce base at a time when value derivation is being measured in basic ‘it cost me how much?’ cash terms.
The AWU, which measures a single discrete unit of work completed by an agent, whether it’s a decision made, a problem reasoned through, or an action taken, was devised by former CMO Patrick Stokes and getting buy in around the concept as more than just a proprietary metric from Salesforce is an ask.
As of April of this year, Agentforce agents hit 734 million AWUs, according to Salesforce, growing about 15% month-over-month over month.
Again, I’m intrigued to see how far AWUs are becoming factored into end user thinking about ROI on AI investment?
Lots more to come from Dreamforce next month.