The AI challenge facing enterprise customers is how to turn fast-moving capability into trusted, organization-wide value without being overwhelmed by uncertainty around models, data, regulation, sovereignty, cost and the future of work.
That was the message from Mick Costigan, VP of Salesforce Futures, speaking with CX Today at Salesforce Agentforce World Tour London. Costigan leads a team inside Salesforce’s strategy organisation that helps the company’s leaders and customers “anticipate, imagine, and shape the future”.
As Costigan put it, this means tracking what is changing outside Salesforce across technology, customer behaviour, geopolitics, and the wider business environment and turning those signals into useful stories about what could come next.
“Most of the stories about AI come from science fiction, unfortunately, and they’re not particularly useful,” Costigan said.
“We’re trying to get people beyond the suspicion that Skynet is about to take over, and we’re all going to be human slaves in a robot future, and think about things that are actually going to be meaningful for how we need to make decisions.”
It’s a task that has become more urgent as customers try to understand what Salesforce calls the “agentic enterprise,” an organization where AI agents can reason, act, access tools and support or automate work across business functions.
The Three Challenges Customers Face
Costigan framed the problem around three basic questions customers are asking about AI, with agents having emerged as the “killer app.”
“How good are agent capabilities going to get, number one. Number two, how do I bring them into my organization, and what does it mean in terms of changes for how the organization works? And then point three: what do humans do?”
Those three questions sit at the heart of enterprise AI adoption.
The first is about capability. Customers can see models improving quickly and agents becoming more powerful, but it is unclear how far that will go, how quickly, or which types of models will matter most. The second is about implementation. Even if AI agents become highly capable, enterprises still need to connect them to real data, tools, processes, systems of record, permissions, governance frameworks and user interfaces. The third is about people. If AI can take on more work, companies must rethink jobs, skills, workflows, accountability and the human role in decision-making.
Costigan said customers are often caught between the promise of frontier AI and the reality of their existing organizations. The answer is to pursue both near-term ROI from practical AI use cases and more ambitious innovation around how the business itself might be restructured where frontier models may play a larger role. “My advice is to kind of do both things simultaneously.”
The Gap Between AI Capability and AI in Use
Costigan made a clear distinction between the pace of model development and the pace at which enterprises can actually use AI effectively.
Model capability may be improving at what feels like “the beginning of an exponential rate”. But enterprise adoption is slower because it depends on far more than model intelligence, Costigan said.
“Getting data right, getting access to tools right, getting the interface right, so that people can work with it, generating productivity gains that aren’t just at the level of an individual, but are actually across the organization, is more challenging.”
Enterprise agents need to be able to operate inside complex business environments with trusted context, access to the right data and guardrails that ensure their outputs are auditable.
Salesforce is increasingly focused on what it calls the “agentic harness” to make AI viable for the enterprise.
“When we started building Agentforce, one of the first things we said was models are great, but you need this layer around that model,” Costigan said. Salesforce has worked with frontier model labs on requirements such as zero data retention, which is an enterprise requirement.
“It’s way beyond just trust,” Costigan added. “It’s access to data, everything around context engineering… How do we bring better access to data? That challenge is huge. Access to tools; how does all of this integrate?”
Why Salesforce Believes the Harness Matters
To explain the importance of the harness, Costigan compared today’s AI models with the earliest cars, recalling seeing an 1886 Benz motor wagon at a car museum comprising an engine, “three bicycle wheels and a garden bench, basically” as an analogy for large language models.
“That’s like an LLM in 2021, where the engine is the whole thing, and there’s these flimsy things around it.” The evolution of cars, Costigan pointed out, was not just about engines. “We added a whole bunch of other ingredients… to make it more beautiful… but also safer, more aerodynamic, more efficient and engines less likely to blow up.”
In the same way, AI models may be the engine, but enterprises need brakes, steering, dashboards, safety systems, interfaces and rules of the road. This is key because models can still behave unreliably if they are not properly constrained or connected to trusted information, Costigan noted.
