Why Enterprise AI adoption is stuck on trust, not technology
Salt's AI Breakfast in Dubai brought founders, CTOs and operators building enterprise AI into one room. Here's what they agreed on, and where they're still stuck.
Five tables. A room full of builders, CTOs, and operators. Everyone thought they were talking about a different problem. Engineering workflow, enterprise agents, ROI measurement.
They all landed on the same three words: ops, trust, data.
In July, Salt hosted an AI Breakfast in Dubai. An invite-only roundtable bringing together founders, CTOs and operators building enterprise AI. Five tables, one morning, no agenda beyond an honest conversation about what’s actually working. Here’s what came out of it.
Nobody at this event was stuck on model capability. Everybody was stuck on the plumbing underneath it, and on whether anyone believed the output enough to act on it without checking.
If there’s a single sentence that captured the room, a fintech-heavy table wrote it down almost as an aside:
“Ops is the lever, and trust is the denominator”
That could serve as the title of the whole event.
The SDLC didn’t evolve. It inverted, in five months.
The most energised topic in the room and the timeline, is the shocking part. Multiple tables, independently, described the same shift on the same clock: February to now.
Before: write the product roadmap, review it, write the code. Now: write the requirements, test, push the code. The requirements-writing step used to be a formality before engineering. Now it’s the bottleneck, and it’s being done by people who never used to write specs at all. At one company, Project Managers prototype directly in Lovable, then hand off to Claude for design. No engineer touches the sandbox until it’s time to harden it.
That’s the superpower. But here’s the problem it created. In the room’s own words: better speccing bloats scope, and autonomy causes breakage. The fix people landed on was more upfront research to narrow scope, not more polish after the fact.
The walls everyone hits aren’t technical.
Three separate tables, three separate domains, same conclusion. In Fintech, one team cut ticket volume by 65-70% by automating customer verification checks and reconciliation. But getting from a working prototype into production is blocked by regulation and institutional trust, not model quality.
Payments themselves are quietly working. One team’s automated validation agents now handle roughly 92% of payment approvals with a human layer on top. The irony of the room? Payment validation is one of the more mature, trusted agentic use cases, while payment processing itself is where clients get nervous. Trust isn’t uniform across a value chain; it’s earned function by function.
Nobody is measuring AI’s value the way finance wants it measured.
Traditional finance has three lenses: revenue up, budget down, headcount down. AI’s actual wins-better customer experience, preserved institutional knowledge, faster onboarding-don’t map cleanly to any of them. That mismatch creates real resistance: nobody wants to sponsor a project whose success will be judged by metrics that don’t reflect what it actually did.
The case study that cut through this most clearly wasn’t a tech company. It was an events business making a 700+ SKU chair catalogue instantly searchable and turning vendor selection into an objective comparison instead of a relationship-driven pick. None of that shows up as “cost cut.” It shows as service quality. Exactly the kind of improvement traditional finance frameworks aren’t built to value.
The data underneath the agents is the actual problem.
Nearly verbatim from the room: “Not the agents, the data underneath them.” Spread across Notion, Slack, SharePoint, Jira, with nobody updating documentation once a project moves. You end up asking an agent to decide on data that’s already wrong.
The best answer in the room came down to building the “second brain” as an index, not a store, pointing the AI layer at existing sources of truth rather than copying everything into one dumped repository.
AI didn’t remove cognitive fatigue. It moved it.
A study came up that gave pause: Radiologists using AI for cancer detection saw a 30% performance drop after three months. The counter-argument that won more of the room? AI absorbs the easy cases, so the humans left in the loop are disproportionately handling the hard ones.
“You can outsource your thinking, but you should not outsource your understanding.”
The skills gap is real, immediate, and nobody is training for it.
Two sides of the same coin came up here. Juniors are now producing work, in one case genuinely better architecture than a senior engineer had proposed, without being able to explain why it works. Output without understanding.
Zoom out further, and the room’s honest read was uneasy: institutions haven’t defined what the future jobs even are, and governments aren’t training toward AI-adjacent roles. Multiple teams have already responded on their own, pairing company-wide token budgets with regular prompt-engineering coaching, on the logic that a cap without training just produces frustration, not better habits.
One line to end on:
“The world hasn’t actually changed, it’s just happening faster. Same human problems with AI amplifiers.”
Every hard problem discussed- trust, data quality, org structure, who gets credit, who gets blamed when the model is wrong-predates AI by decades. What’s changed is the clock speed. Five months to invert an entire development lifecycle is not normal, and nobody in that room sounded like they’d fully caught up to their own timeline yet.
If your organisation is trying to work out where AI talent actually needs to sit, Salt’s Essential AI Talent Report maps the roles, skills and gaps shaping 2026.
