The research is in: only 5% of enterprises see substantial AI returns. For mid-market operators in regulated industries, that number should be a wake-up call — not about AI itself, but about how you're deploying it.
A meta-review of 16 research reports — IBM, McKinsey, Deloitte among them — landed on a number that should give every executive pause: only about 5% of companies achieve substantial AI ROI. Another 35% see partial returns. That leaves the vast majority somewhere between "we ran a pilot" and "we have a very expensive proof of concept."
For mid-market companies in regulated industries, this isn't a technology problem. It's a deployment problem. And there's a specific flavor of it that we see constantly.
Here's the pattern: a company runs a focused AI pilot — maybe it's automated document review in a legal workflow, or flagging anomalies in a manufacturing quality log. It works. Stakeholders are impressed. Someone puts together a deck with projected savings.
Then it hits the real environment.
Compliance needs a formal risk assessment. IT needs to review the API integrations. Change control wants documentation that doesn't exist yet. The vendor's SLA doesn't meet your data residency requirements. The process the AI was trained on turns out to be the *clean* version — not how the work actually gets done.
Six months later, the pilot is still a pilot. Or it got quietly shelved.
This is not a failure of AI. It's a failure to treat AI deployment the way regulated industries have to treat any operational change: as something that requires infrastructure, not just inspiration.
The companies that see real returns aren't necessarily using better models or more sophisticated tools. They're doing something more mundane: they start with a workflow problem, not an AI solution.
Specifically, they do three things differently.
They scope to a single, measurable process first. Not "improve operational efficiency." Something like: reduce the time to complete a supplier audit from 14 days to 5. That specificity is what makes ROI measurable — and defensible when the CFO asks.
They build compliance into the workflow architecture, not as a review step at the end. In medical device, financial services, manufacturing — any regulated space — trying to retrofit governance after you've built the workflow is how you end up in a six-month compliance review loop. The 5% treat audit trails, access controls, and model validation as design requirements, not afterthoughts.
They instrument the workflow before and after. You can't prove ROI on a process you weren't measuring to begin with. This sounds obvious. It almost never happens. Companies deploy AI into processes with no baseline data, then wonder why they can't demonstrate value to skeptical stakeholders a year later.
All of this is harder in regulated industries — and the gap between the 5% and everyone else is wider here. Here's why.
In an unregulated context, a broken AI workflow is a productivity problem. In medical device or financial services, it's a compliance event. That asymmetry makes organizations cautious in ways that paradoxically make them more likely to fail. They scope pilots so narrowly that the workflow never touches anything consequential. Then they can't demonstrate ROI because the use case was never connected to an actual business outcome.
The answer isn't to take more risk. It's to build workflows that are designed to survive regulatory scrutiny from day one — which actually makes them faster to scale, not slower.
When we start working with a mid-market operator, we run every proposed AI use case through four questions before we recommend moving forward:
1. Is there a measurable baseline? If you don't know how long the process takes today, or what the error rate is, stop. Instrument first.
2. Does a compliance-ready architecture exist? Not "can we add compliance later" — does it exist now, in the design.
3. Who owns the workflow outcome? AI projects that report to IT die in procurement. They need an operational owner whose job performance is tied to the result.
4. What does failure look like, and is it recoverable? In regulated industries, some failure modes are unacceptable. Know them before you build.
None of this is glamorous. But it's why the 5% are the 5%.
The companies getting real AI returns aren't the ones who moved fastest. They're the ones who built workflows that could actually survive the environment they operate in.
Dealing with a similar challenge?
We work with mid-market companies in regulated industries to build AI workflows that actually hold up.
Let's TalkSean Cummings
Founder of Laminar Flow Analytics. Specializes in AI workflow automation for regulated industries — medical device, financial services, and complex logistics operations.