Ask most executives who own AI adoption in their company and they’ll point to the CIO. Ask who is actually adopting it, and the answer looks very different.
The people putting AI to work right now are revenue operations managers wiring lead routing across a CRM and a billing system. Support operations leads automating refund flows. Finance analysts reconciling data between platforms that were never designed to talk to each other. Gartner calls them business technologists, employees who report outside of IT but build technology or analytics capabilities for the business, and by Gartner’s measure they already make up 41% of the workforce. The same research found that 74% of technology purchases are now funded at least in part by business units outside IT.
This isn’t shadow IT rebranded. It’s a structural shift in where technical work happens, and it is about to determine which companies get real value from AI and which ones stay stuck in pilot purgatory. I’d argue it’s also created a gap wide enough to need its own category of software, one built specifically for the business technologist rather than the engineer.
The operator knows the workflow. The backlog owns the operator.
Here’s the pattern I see inside almost every mid-market and enterprise company. The people closest to a broken process can describe, in precise detail, the automation that would fix it: “When a cancellation request comes in, pull the customer’s spend and recent NPS, and draft a save offer if the numbers justify one.” That’s a complete functional spec, spoken in plain language, by the person who owns the outcome.
What they can’t do is build it, because the workflow touches four systems, one of them has no pre-built connector, and the engineering queue is booked through next quarter. So the work either dies in the backlog or gets quoted out to a systems integrator at a price that makes everyone quietly drop the idea. The process stays manual, and the operator goes back to copy-pasting between tabs.
For a decade, our answer to this gap was low-code tools and drag-and-drop workflow builders. They helped, but they all share the same ceiling: they only reach as far as their connector libraries, and they tend to break the moment a step in the process needs actual judgment rather than a fixed rule. The moment a workflow touches a legacy ERP, a niche vertical tool, or an internal system with an unusual API, the business technologist is back in the engineering queue.
What actually changed
Large language models can now write production-grade integration code on demand. That sentence sounds incremental. It isn’t. It removes the exact constraint that kept business technologists dependent on scarce engineering time: the connector that doesn’t exist yet.
When an operator can describe a workflow in natural language and have working code generated in minutes, the question stops being can the model build it and becomes should this run be trusted: is it deterministic, is the code reviewable, does a human approve the steps that matter? That reframing is the whole game for enterprise AI adoption, and it’s exactly where today’s three most common tools fall short.
- Legacy automation is reliable right up until a process needs judgment, then it breaks.
- General-purpose AI chat tools are remarkably good at helping a person build something, but weren’t designed to be trusted running a live write to a production system on their own, unsupervised, at scale.
- “We’ll just build it ourselves” starts as a clever internal project and, a year later, is a graveyard of half-maintained agents that only one person on the team still understands.
None of the three actually answers the trust question. The capability question is largely answered; the governance question is now the differentiator, and it’s what opens the door to a fourth option: agentic workflow orchestration, software that learns a process like AI, then runs it like code.
And notice who this puts in the driver’s seat. Not the team with the deepest ML bench, it’s the team whose operators best understand their own workflows. Gartner’s analysis found that organizations that successfully empower business technologists are 2.6 times more likely to accelerate their digital business outcomes. AI compounds that advantage, because the highest-value automations are exactly the messy, cross-system ones that only the process owner fully understands.
We see this daily at Brainfish. It’s why we built Ballet, an agentic workflow orchestration platform that generates custom API integrations on the fly and runs them as deterministic, reviewable code. Workflows that used to wait weeks for a forward-deployed engineer now ship the same day, built by the operator who owns the process. But whatever tools companies choose, the direction of travel is the same: the build is moving to the business.
What this means for how companies adopt AI
If you accept that the center of gravity is shifting, three implications follow.
Stop running AI adoption as an IT procurement exercise. The companies getting results aren’t the ones with the most rigorous vendor evaluation matrix. They’re the ones that identified their business technologists, the 41% already doing this work, and gave them sanctioned tools, real system access, and a clear lane. Adoption strategies built purely top-down keep producing impressive demos and empty production environments.
Make IT the governor, not the gatekeeper. The instinct to centralize control is understandable and wrong. In Gartner’s research, 80% of business technologists said they found more value partnering with IT than working around it. The winning posture is guardrails over gates: reviewable code rather than black-box agents, a full audit trail and source citations on every run, and human approval checkpoints on anything financial or destructive. Say yes to the workflow; be uncompromising about the conditions.
Judge AI tools by their trust surface, not their demo. Anyone can make an agent look brilliant for ninety seconds. The questions that matter are duller: Can someone inspect what it will do before it runs? Does it behave identically on run 400 as on run 4? Who signs off when it touches money? This is the real dividing line between a chat agent you supervise by hand and an orchestration platform you can actually hand a process to. Tools built for business technologists succeed or fail on whether the security team, and the operator, can trust them on a Tuesday afternoon with real customer data.
The rise of the business technologist was already reshaping enterprise software budgets before generative AI arrived. AI didn’t create the trend; it removed its last constraint. The companies that treat those 41% as their AI adoption strategy, rather than a compliance risk to be contained, and that give them a category of tooling actually built for how they work, will be the ones whose backlogs finally start shrinking.
How Ballet can help. Ballet was built with all three of these in mind: a clear lane instead of a procurement queue, guardrails IT can actually rely on, and a trust surface that holds up to scrutiny. It lands directly with the operator in hours to weeks, not months. Every workflow ships as reviewable code with a full audit trail, source citations, and human approval checkpoints on anything financial or destructive. And it’s inspectable before it runs, consistent from the first execution to the four-hundredth, with a clear sign-off step before anything touches money.
