It’s been quite the year for Salesforce’s flagship AI tool, Agentforce. Introduced just over a year ago, Marc Benioff dubbed 2025 the “year of Agentforce”, with promises to change the agentic enterprise as we know it. But as many of our regular readers might know, the Salesforce ecosystem hasn’t exactly warmed to agentic AI in the same way that the mothership has.
Salesforce’s Q4 2025 results showed that adoption was fairly slow out of the gate. 5,000 Agentforce deals were secured in that time, with just 3,000 of those deals paid for. We wanted to know what the real roadblocks were to enterprise adoption at that time, and found that many of the same issues came up – unclear pricing, messy orgs, and weak enablement. These problems seemed to have continued across the year, and ahead of their Q3 earnings call, the CRM giant needed to show some signs of improvement to keep investors on board.
And surprisingly, the company was able to do just that. Described as “momentum drivers”, Agentforce and Data 360 (formerly Data Cloud) hit $1.4B in Annual Recurring Revenue (ARR), with 18,500 Agentforce deals closed in 2025 (9,500 paid), alongside 3.2 trillion tokens processed. Pretty impressive numbers, given the skepticism from a lot of businesses we’ve seen this year.
Yet, it still begs the question: does this mean real, genuine adoption in the field, or just better packaging and financial storytelling? To answer this question, we reached out to two industry heads with real Agentforce implementation experience to see if those same roadblocks have been successfully mitigated, or if those issues raised in April are still impacting the Agentforce experience.
Has Agentforce Actually Matured?
Salesforce’s latest earnings call certainly suggests that Agentforce is showing momentum on paper, but numbers like these don’t really tell you how an AI product or agent actually behaves inside a real Salesforce org – and this was a big criticism back in April. Many we spoke to earlier this year described the early iterations of Agentforce as unpredictable, fragile, and not yet ready for prime time.
Eight months on, and it’s fair to say that the picture has changed; not completely, but noticeably.
Both implementation experts we spoke to for this article agreed that Agentforce has indeed made substantial strides in basic stability, documentation, and support readiness – especially compared to the messy, buggy experience of early 2025.
Joseph Monroe, Senior Salesforce Consultant at Blue Gator, has worked with both enterprise and nonprofit clients this year to help with implementing Agentforce, and didn’t hesitate when asked if he thought Agentforce had grown since its release.
“As far as growing up or advancing, I definitely think it has,” Joseph explained. “Back in May, there was a lot less information. The Agentblazer trails, as well as information and guides, were thinner, I guess you’d say. Now, everything has advanced – from best practices, how to work with agents, things that you should consider, plus all of the guides that are available now. I think that has been a huge leap forward.”
The general consensus concerning the early Agentforce experience was arguably characterized by missing documentation and unclear guidance. But over the past six months, it feels as though Salesforce’s internal enablement has definitely caught up. Finding best practice guides and learning resources is now much easier to find and apply.
However, product maturity isn’t just about documentation. Whether the systems are predictable is also vital, and this is where Agentforce has made the most important progress.
Salesforce MVP Zachary Banks, Director of System Architecture at Link Logistics, has also worked extensively this year implementing Agentforce, and recalled to me how rough those early months were for Agentforce.
“Compared to March, Salesforce has definitely made strides with market readiness for Agentforce,” Zachary explained. “Months ago, it was very buggy. It would lose context. It would have different issues when you try to promote it across different sandboxes. It would have different issues across different form factors (mobile vs. desktop). So, it has definitely made a lot of strides.”
Losing conversational context became an instant blocker for customers. If an agent can’t reliably remember what the user is asking, it probably can’t be trusted in production. The same was true of the inconsistent behavior between desktop and mobile that Zachary mentions, and broken deployments when moving from one environment to another.
While Zachary is clear that Salesforce still has some work to do, the changes/improvements that we’ve seen to end the year are influential enough to change the tone of the conversation.
Zachary said: “Yes, it’s in a much better position, but there’s still a lack of documentation. They’ve really leaned into their partners on having to go and implement different product areas that they sell, but they really lack the finer-tune areas for how to configure a certain topic, such as deploying Agentforce across sandboxes and having it remain having context. I don’t think you could find an article anywhere on that issue.”
So, in other words, Agentforce might no longer be “very buggy”, as Zachary put it, but it still isn’t frictionless.
Meanwhile, perhaps the biggest technical breakthrough Joseph highlighted was around instruction writing – arguably the weakest part of the early product.
The first-generation instruction model struggled with even basic structured login, as Joseph explained.
“The instruction model didn’t really handle structured ‘if-then-else’ logic – it was very very common in a lot of our use cases. So we typically had to offload those onto Flows or Apex to handle the very structured type of logic. It was kind of an art and a science to write really good instructions.”
Salesforce has gone a long way to answer this with the release of Agent Script – which, according to Joseph, is one of the most meaningful functional improvements since launch.
“Agent Script – which is like the new version of how you write instructions – handles things a lot better from what I’ve been able to test within the beta. It’s easier to write up an agent than free-text instructions. That was a really big hurdle at the beginning to write really good instructions that gave consistent results. So from what I’m seeing, that’s going to be vastly improved from where it was.”
So all in all, Agentforce has definitely matured. Not enough to declare the ecosystem fully ready, not enough to eliminate frustration, and certainly not enough to match Salesforce’s current marketing velocity. But enough that seasoned Salesforce experts are no longer calling it “immature” or “not ready for production”.
It’s a more stable, more predictable, and more supportable product than it was in spring 2025.
Pricing, Accessibility, and Buyer Confidence
For all of Agentforce’s technical evolution, its adoption story has always hinged on something fairly basic: can customers afford it, and do they trust it?
In April, the answer for many was a hard no. The original $2-per-conversation pricing model wasn’t only unpopular, it was prohibitive. It immediately priced out nonprofits, SMBs, and any organization that did not have a large enough AI budget. Even enterprise customers were uneasy, with many describing it as unpredictable, hard to forecast, and disconnected from the real cost of serving AI actions.
As Founder of Elements.cloud Ian Gotts aptly told Salesforce Ben in April: “They don’t know how to price them yet, and right now, nobody wants to hand out a blank cheque – they want caps and predictable ROI. So the $2 price? It’s what you pay to learn of a narrowly scoped agent.”
Fast forward to now, and it feels as though the landscape has changed in a meaningful way.
Pricing Isn’t the Barrier It Once Was
When Salesforce shifted Agentforce toward a flex-credit model, it blew open an entire segment of customers that had otherwise been locked out previously. Joseph described this change as a huge moment – flipping the opportunities smaller customers now have with Agentforce.
“Back when Agentforce conversations were $2 a pop, for smaller nonprofits, that was a complete non-starter for them,” Joseph said. “That was just far too expensive for what they could get out of it at that point in time. Now, with the change to flex credits, that’s been opening up a lot more conversations with our clients.
“Now that could be as low as like 10 cents a conversation depending upon what actions are being taken, but also the smaller initial investment that it takes. We’ve been having a lot more conversations with smaller organizations that are interested and now it’s kind of democratized to where they feel comfortable jumping in.”
This shift in particular cannot be understated. For smaller businesses or nonprofits, a 20x reduction in per-action cost is the difference between AI being for companies with money or something that’s accessible to the entire ecosystem.
But pricing changes alone aren’t the reason for adoption becoming more democratized. Buyers have also become far more educated over time, and far more cautious. As Ian implied earlier, customers can’t rely on Salesforce’s narrative and need more proof before purchase.
Joseph elaborated on this when applying his real-life experiences with customers, saying: “[Customers] wanted to kind of wait it out and see like ‘what are people doing?’ ‘Can you give us solid examples of agents that you have in production?’ I think that’s what a lot of people were waiting on, is showing agents in production. I don’t want to see a sales demo – show us what’s actually happening in an org.
“And once we were to the point where we could actually show that and show an agent actually operating, what the ROI on that was, what it took to get there, that conversation got a lot better.”
But Pricing Is Still Confusing, Especially for Enterprises
Despite some recognized improvements, the pricing story is far from simple. For many customers, especially those who purchased early Foundations bundles, there is still uncertainty about what they’re actually consuming and paying for.
Zachary offered a blunt assessment of the situation, saying: “I definitely think it’s still a bit of a gray box. I do see Salesforce trying to improve it, but here’s the thing – a lot of customers got Agentforce when Foundations was announced for free at Dreamforce. Then Agentforce was layered into that SKU to activate and many never really realized the utilization cost for the credits.
“And then recently Salesforce announced, ‘Oh well, if you buy the Agentforce plus SKU, you don’t have to worry about that.’ It’s like a user license based. But I don’t know how many customers have done that switch.
“So, it’s still a very frequent question when we’re planning on targeting a subcapability area. If we don’t have that unlimited user license, what does it cost? And that takes a lot of cycles with different sales teams at Salesforce to figure out how much it costs.
“I don’t think it’s crazy expensive, but I think it’s just uncertainty.”
So this is a paradox Salesforce now may find itself in – Agentforce is no longer prohibitively expensive, but still may be poorly understood. It’s clear from both perspectives that adoption today really isn’t driven by Salesforce’s marketing, but by social proof such as real deployments and real ROI.
Pricing may no longer be the biggest barrier for Salesforce, but buyer confidence is still fairly fragile. Progress is there, but the trust gap from April hasn’t closed completely.
Was the “Perfect Org” Narrative One Big Myth?
If there was one universal conclusion from our original article, it was that messy orgs were practically the kryptonite of Agentforce.
Technical debt, legacy processes, and years of admin patchwork all created the belief that organizations need to completely “clean their house” before they even consider implementing agentic AI.
But this may not be the full truth. Eight months on from the original conversation, it feels as though the narrative has changed dramatically. When speaking to both Joseph and Zachary, there was an agreement on one core truth: waiting for the perfect org isn’t just unrealistic, but unnecessary.
When I asked Joseph about Agentforce needing flawless architecture to perform, he said: “I’ve been in the ecosystem around six years as a consultant and dealt with probably 30 or 40 orgs. I’ve never seen one have perfect data. I don’t think I will ever see it. That’s kind of the reality. In a perfect world, would we like your data to be perfect? Sure. But that’s not the reality and it’s not realistic for most organizations.”
This alone marks a major mindset shift from the first half of the year. The new, more reassuring reality is that you really don’t need a flawless org to get started, but more of a bounded domain you can control, with data consistent enough for an agent to reason over.
Starting Small and Contained
Throughout this year, many have treated technical debt as a big blocker. But Joseph stressed that teams are now leveraging Agentforce proofs of concept to surface architectural issues rather than avoid them.
“What we found when we’re running a proof of concept is that starting in a controllable section with a controllable use case with clear ROI was the best catalyst for success.
“Because it’s not sexy, it’s not going to land you on a stage. It’s not going to be incredibly exciting necessarily for a proof of concept. But what it is going to do is, number one, flesh out that this product works and show it can provide you actual value. And number two, it actually highlights areas of concern within the data or within the architecture. Through our implementation we identified a lot of gaps because of the way that the agent was handling some of that data.
“So we were able to fix those things, and it was very clearly highlighted what needed to be fixed, what could wait a little bit and then what sort of road map should we have for the future.”
As Joseph explained, it seems as if we’re now seeing more small-scale agents that are functioning as diagnostic tools, exposing data inconsistencies, broken flows, and the like. These are gaps that would otherwise remain invisible in day-to-day operations.
Joseph captured this approach with an interesting metaphor: “Everyone wants to hit the home run, but they’re scared of hitting a double. A double is still a great outcome! And that’s where people should start.
“Start small, start scalable, and really track your progress along the way.”
Zachary echoed this sentiment from an architectural point of view, reinforcing that the smartest Agentforce implementations are the ones with the narrowest scope.
“I’d recommend starting with a niche area so that the amount of data needed isn’t going to break the bank proverbially on project cost, cleanup cost or any of that.
“Once you get good adoption, and you get your feet wet with that area, then you could expand to other capability areas now that you have, like an agent with working contacts that has gotten buy-in with how it’s answering questions in its non-deterministic manner. And then you could expand it a lot easier.”
Successful Agentforce adoption is no longer about achieving architectural perfection. More focus on choosing the right starting point and building momentum with small wins is the most practical approach going forward, and is likely to slowly remove one of the biggest psychological barriers concerning Agentforce adoption earlier this year.
Final Thoughts
So in the end, Agentforce’s story in 2025 isn’t one of overnight success, but rather a narrative of steady and measurable progress made across the board. The product is indeed maturing, pricing is opening doors rather than closing them, and organizations are working out how to navigate the realities of agentic AI rather than the marketed successes or promises.
The roadblocks identified in April haven’t vanished, but evolved massively, as has the ecosystem’s understanding of what “adoption” really means and looks like.
If Salesforce can continue narrowing this gap between marketing and lived experience, 2026 could be a breakthrough year when the entire ecosystem can operationalize it.
