If you ask a Salesforce professional if they’re using AI, the answer will almost certainly be yes. Our data reveals that a significant proportion of admins, developers, and architects use these tools on a daily basis – with very few at all never making use of them.
But look a little closer, and another trend emerges. While Salesforce professionals are, on the whole, adopting AI, trusting what it produces is another question entirely. The most recent SF Ben data from admins, developers, and architects reveals that trust is the number one barrier to using AI.
1. Adoption and Trust
AI adoption is high across all three roles mentioned above, broken down like this:
- Architects: 63.6% daily/regularly; 33.1% occasionally/ad hoc; 3.2% not at all.
- Developers: 49.3% daily/regularly; 39% occasionally/ad hoc; 11.7% not at all.
- Admins: 43.4% daily/regularly; 41.3% occasionally/ad hoc; 15.3% not at all.
But the disconnect is evident when we look at the figures for the biggest barriers to using AI. These are broken down like so:
- Architects: 19.3% trust; 19% skills and knowledge; 17.6% accuracy; 17.3% company buy-in; 16.1% cost; 10.8% capacity and resources.
- Developers: 21.8% trust; 21.3% cost; 19.8% skills and knowledge; 16.8% capacity and resources; 14.7% company buy-in.
- Admins: 29.3% trust; 22.8% skills and knowledge; 16.9% capacity and resources; 16.1% cost; 8.6% company buy-in.
So, adoption is high, but trust is consistently the biggest barrier. While it’s the most popular reason among all three personas, it’s not a runaway winner with any of them – excepting perhaps for the admin, where it has a seven-point lead over second place. Why might this be?
Hallucinations are still a “baked-in” problem with generative AI models, and even if the chances of them taking place are minor, Salesforce professionals still need to be able to trust that the work their AI is doing is actually sane.
At Dreamforce ‘25, I spoke to Shibani Ahuja, SVP, Enterprise IT Strategy, Salesforce, about the problem with hallucinations and trust.
She told us that phenomena which might initially appear to be hallucinations could simply be the natural result of conflicting or chaotic data. Shibani told us that, when Salesforce initially rolled out Agentforce on help.salesforce.com, they believed that the agent was hallucinating because responses were inconsistent – forcing them to shut it down.
But, after analyzing what actually happened, the data was the problem – because there were conflicting knowledge articles the agent was relying on. Shibani told SF Ben: “This problem has persisted for as long as we’ve had our website, but we didn’t identify it until that agent was put on, and that surfaced something.
“So then we turned that agent inwards to identify the anomalies, which is a brilliant use case. Now you’ve got this guardian agent that is overseeing and helping you clean your data, or do QA, QC. That’s like another value or another learning of failing and just trying it.”
In any case, Salesforce is often saying how trust is their #1 value. It may be somewhat auspicious, then, that it also appears to be the biggest barrier to AI adoption.
2. Skills and Knowledge
As many as 19% of Architects selected ‘Skills and Knowledge’ on the question of barriers to AI adoption – the second most popular option, just shy of being tied with ‘trust’.
Salesforce Architects are among the most technically fluent in the ecosystem, so this is a notable finding on its own. Architects need to develop an entirely new layer of competency on top of their existing platform expertise – that being, how to effectively prompt AI, evaluate its output, and fold it into governance and review processes, while judging where an AI recommendation should be valued, and where it should not.
Part of the challenge is that traditional Salesforce learning paths were not built with the challenges of today. Trailhead is great for teaching the fundamentals of the platform in a controlled environment, with clean data and requirements that don’t radically shift mid-project. Enterprise AI work is not quite so tidy.
For admins, ‘skills and knowledge’ came second at 22.8% – seven points behind trust (29.3%).
For developers, it was third (19.8%), behind ‘trust’ (21.8%) and ‘cost’ (21.3%).
The skills gap is not yet fully closed – even among the most proficient Salesforce professionals. What’s interesting is that the skills gap scales inversely with technical seniority. Admins report it as a bigger barrier than developers, who report it more than architects.
This makes sense intuitively. Admins are often less technical, so an AI layer which assumes prompt literacy, data governance judgment, or agent-building know-how hits them hardest.
3. Accuracy
‘Accuracy’ got 17.6% of responses in our Architect Survey. It’s worth reading this alongside ‘trust’, rather than in isolation, because they do have some overlap.
Trust and accuracy together account for 36.9% of all responses – more than a third of all responses. For Architects, whose decisions affect security models, data architecture, integrations, and enterprise-wide systems, being dependable is far more important than simply being useful.
An AI tool can blow people’s minds in a demo, but then fail to meet the standards required for actual use in production. If an AI is not accurate, then it’s no good. That’s the only sane policy.
For admins, accuracy concerns might show up as ‘trust’ by proxy. Admins are typically the ones fielding end-user complaints when an agent gives a wrong answer. Since they reported the highest ‘trust’ score of all three roles, we might conclude that this answer is doubling up in some way as ‘accuracy’ too. The line between the AI ‘lying’ or being ‘unreliable’ might not be as sharp for them.
Last year, CEO Marc Benioff said that Salesforce’s agents were operating at around 93% accuracy – a figure which, in enterprise environments, arguably falls quite short of what is needed.
While 93% might look like success to the outsider, this is arguably a long way from the level an agent should be performing at in many cases. When measuring business systems, Six Sigma is a critical framework. It is a quality control framework originally developed by Motorola in the 1980s to minimize errors in manufacturing – but the principles have been applied across a range of different industries, from hospitals to software.
This framework argues that there should be no more than 3.4 defects per million opportunities (PMO), which is around 99.99966% accuracy – statistically, almost perfect.
If you were to apply this framework to Salesforce’s 93% figure, it does not seem quite as impressive. Does that mean that of the one million Agentforce support cases, 70,000 of them were done wrong? In any case, it is a far cry from the high standard of Six Sigma.
4. Company Buy-In and Cost
Company buy-in was selected by 17.3% of respondents in our SF Ben Architect Survey 2026, on the question of what the biggest barrier to using AI is. This is just 0.3% behind accuracy – close enough that it deserves equal weight in any adoption strategy.
This is a familiar shape for enterprise technology. Individual practitioners can see the potential value of a tool long before their business is structurally ready to support it. Without leadership sponsorship, clear governance frameworks, procurement pathways, and alignment from other stakeholders, even an architect who (hypothetically) trusts AI and knows how to use it may simply be unable to deploy it.
This sits outside the architect’s control, making it harder to solve through training alone.
Last year I spoke to Salesforce Application Manager and Agentforce Lead at Arjo UK, Will Turner, about the human side of implementing AI – particularly Agentforce.
He told us that it helps to get buy-in at every level, not just senior leadership. They need to work through their concerns, but user buy-in matters just as much, and shouldn’t be skipped.
Security fears should also be taken seriously, even far-fetched “Skynet situations” in early training sessions. At this point in AI’s journey, hopefully most people understand that ChatGPT isn’t the Terminator, but there is still real value in walking users through the actual boundaries and controls in place.
It also helps to engage stakeholders early, listen to every concern raised, and make sure each one gets a clear answer, because transparency is what ultimately builds trust, and company buy-in.
As many as 16.1% of respondents to our SF Ben Architect Survey 2026 selected ‘cost’, when asked what the biggest barrier to using AI is.
Given how often cost is treated as the default explanation for slow technology adoption, this is quite a useful finding for vendors and decision-makers to absorb. Trust, knowledge, accuracy, and organizational support are all bigger barriers. Cost is a factor, but people are willing to pay for systems they trust. Salesforce’s history of becoming the world’s number one CRM is proof enough of that.
Aside from having to factor in the specific cost of an AI solution, there’s also a degree of unpredictability with AI prices.
This is something Salesforce has consistently addressed, starting out Agentforce at a “$2 per conversation” model, then to flex credits, and more recently to a pay-per-resolution model for some agent offerings.
Salesforce currently offers consumption-based pricing with Flex Credits or Conversations, or per-user licensing. Using Flex Credits, you pay per action, at a cost of $500 per 100,000 credits.
Nobody wants to write a blank check to their AI provider, and consumption-based pricing can sometimes feel like this. Caps and predictable ROI are reassuring.
5. Capacity and Resources
Our Architect Survey saw 10.8% of respondents selecting ‘capacity and resource’ when asked what the biggest barrier to using AI is. On the surface, a resource shortage might seem like the most difficult problem to solve, since it means genuinely lacking time, people, or bandwidth. Arguably, having this as the lowest of the options is good news.
It arguably means that architects are more likely saying they need proof of AI working, rather than lacking resources to adopt it. That’s a solvable problem. If you can build trust and educate stakeholders about the realities of the risks associated with AI, then genuine AI adoption might be within reach. Those interventions seem eminently more achievable than a wholesale increase in headcount or budget.
Developers selected ‘capacity and resources’ 16.8% of the time, with admins selecting it 16.9% – a remarkable similarity. This suggests that whatever is driving this barrier is not necessarily role-specific, but a more structural, org-wide constraint which affects Salesforce teams regardless of depth.
Final Thoughts
The clearest takeaway from the data seems to be the lack of a single dominant barrier. Trust, skills, accuracy, company buy-in, and cost all sit within a few percentage points of each other. Any organization or vendor that concentrates its efforts exclusively on just one aspect will likely only solve part of the problem.
Still, trust, accuracy, and skills account for 55.9% of all responses. The challenge for vendors like Salesforce is proving their capabilities are reliable, explainable, and governable enough to earn a permanent place in enterprise architecture.
