Buying your way past the AI trust problem

On July 2, Microsoft announced it was spending $2.5 billion to embed roughly 6,000 engineers, industry specialists, and technical consultants directly inside client companies. The unit is called Frontier Company, and its job is to sit inside enterprise clients and physically make AI systems work, rather than sell a tool and walk away.

Two days earlier, Amazon had made the identical bet at a smaller scale: $1 billion for a forward-deployed engineering unit that embeds teams of five or six engineers inside customer environments for roughly 45-day sprints. OpenAI and Anthropic had already made versions of the same move earlier this year. Four of the largest AI companies in the world, all reaching the same conclusion within months of each other: the technology isn’t the bottleneck. Getting organizations to actually use it is.

The gap everyone’s spending money to close

The number driving all of this is uncomfortable. IBM’s 2026 CEO Study found that 85% of employees have AI capability available to them, but only 25% use it regularly. CEOs now rank employee adoption as their top AI concern, ahead of cost, security, and accuracy. Forrester’s survey of 1,500 AI decision-makers tells a similar story: three years into the generative AI era, most enterprises still can’t convert adoption into measurable transformation.

Microsoft and Amazon’s answer is headcount. Send in engineers, embed them in the org, close the gap by force of expertise. It’s a reasonable instinct. It’s also treating a culture problem like a staffing problem.

What forty-five days can’t fix

Amazon’s own stated goal for its forward-deployed engineers is to leave clients “self-sufficient” after the engagement ends. That’s a tell. If self-sufficiency were the natural outcome of embedding experts for six or seven weeks, IBM wouldn’t need a CEO study to explain why adoption keeps failing.

Part of the reason a fixed-length engagement struggles here is that AI isn’t like the software rollouts organizations are used to. Installing a word processor is a bounded problem: there’s a finite set of features, a manual, a right way to use it. AI is open-ended by comparison. There’s no manual that tells a given employee exactly how it should reshape their specific job, which means the real work of figuring out how to use it well can’t be handed to a visiting team and declared finished when the engagement ends.

The harder question these programs sidestep is what happens after the engineers leave. Does the rest of the organization keep following the practices that were built for them, or does it revert the moment the external team packs up? A 45-day sprint can produce a working system. It can’t produce an organization that trusts the system enough to keep using it the way it was designed, and trust is not something that gets installed by a visiting team on a deadline. It has to be built by the people who are going to live with the tool every day.

Adoption was never a deployment problem

IBM’s own research points in a different direction than IBM’s own client-facing pitch. The CEO study’s prescribed plays are about embedding AI into how teams already collaborate and transfer knowledge, integrating it into real workflows instead of bolting it on top, choosing tools by specific use case rather than a single enterprise-wide rollout, and orchestrating when human judgment should override the AI rather than defaulting to automation. None of that is about adding people. It’s about changing habits.

A related study puts harder numbers on the same problem. Harvard Business Review Analytic Services surveyed 325 HBR-audience respondents in a report sponsored by Hyland, and found that 94% call well-connected data, processes, and applications critical to AI success, but only 27% say those elements are actually connected at their organization. Just 12% say AI is embedded directly in the flow of work, most of it still runs through separate, standalone tools, and only 45% say their AI projects are delivering the outcomes they expected. The gap between what leaders know matters and what their organizations have actually built is the real story here, and it’s not a gap $2.5 billion in embedded headcount closes on its own.

The moving target problem

Part of what makes this adoption cycle different from past technology rollouts is that the target doesn’t hold still. It isn’t just the tools that keep changing, it’s the roles built around them. Employees are being asked to redefine what their job even means while the capability underneath them shifts month to month, which makes “are we learning fast enough” a genuinely unanswerable question most of the time. Defining value gets harder for the same reason. It’s not just a C-suite measurement problem, it’s an individual one: an employee needs to know what value they personally get from a tool before they’ll trust it enough to change how they work.

That’s also a communication problem as much as a technical one. The organizations making progress aren’t the ones with the most sophisticated rollout plan, they’re the ones translating that plan into simple, individually relevant stories: what this actually changes for you, specifically, rather than an enterprise-wide value proposition nobody recognizes themselves in. Forrester’s finding that customer-led adoption outperforms internally-focused efficiency plays runs on the same logic. Customers still want a human in the loop for plenty of interactions, and forcing AI on them doesn’t build trust any more than forcing it on employees does.

What actually closes the gap

The piece most of these programs skip is knowledge governance. It’s easy to assume an AI system can simply crawl a company’s documents and understand how the business runs. In practice, there’s enormous nuance in how organizations actually operate that never makes it into a wiki page, and an AI system trained on incomplete or ungoverned knowledge will confidently produce wrong answers, which erodes exactly the trust adoption depends on. Governance cuts the other way too: without deliberate control over what the AI can access, sensitive information ends up reachable by people, or systems, that were never supposed to have it. Closing the capability gap requires upskilling employees to manage both sides of that governance deliberately, accuracy and access, not assuming the AI will sort out either one on its own.

None of this is something an outside team can hand off on its way out the door. It requires employees who are given room to redefine their own roles on their own terms, rather than have new workflows spoon-fed to them, and it requires transparency about how the AI actually works and what it’s drawing from, because transparency is what earns buy-in instead of demanding it. The organizations closing the adoption gap won’t be the ones with the most embedded engineers. They’ll be the ones that built the internal trust no consulting engagement can install.