Everyone has seen the demo. The model reads the contract, flags the risk, drafts the report — flawlessly, on stage, against clean data. And almost everyone has seen what happens next: six months later, the pilot is still a pilot, procurement has a hundred questions, and the workflow it was meant to replace is still running. The gap between a demo and a business is where most enterprise AI dies.
Why regulated industries are different
In consumer software you can ship, observe, and iterate. In banking, payroll, insurance, or payments, three walls stand between a working model and a working product:
- Compliance review. Every automated decision needs an explanation, an owner, and an audit trail. "The model said so" is not a sentence a regulator accepts.
- Legacy data. The demo ran on clean data. Production data lives in a 1998 core system, three acquisitions deep, with fields repurposed twice. Models don't fail in production because they're weak; they fail because the data isn't what anyone said it was.
- Real users. Analysts, accountants, and payroll ops have seen tools come and go. If the AI adds a tab instead of removing work, they will politely ignore it to death.
The playbook that survives contact
Across eight product lines in regulated markets, the pattern that works has stayed remarkably stable:
1. Engineer compliance in, not on
Evaluation harnesses, decision lineage, and model versioning from the first sprint — not retrofitted for the audit. When compliance is a property of the architecture, review cycles shrink from quarters to weeks, and the burden that scares off generic AI vendors becomes your moat.
2. Deploy engineers forward
The single highest-leverage decision we make at Ideallio: engineers work inside the customer's organisation, against the real systems. Discovering that the data is dirty in week two costs a sprint; discovering it in month nine costs the company.
3. Put humans in the loop where stakes are high
Confidence thresholds decide what flows straight through and what a person reviews. This isn't a concession — it's what makes high automation rates safe, and it generates the labelled data that improves the system.
4. Move a number the business already watches
False positives down 40%. Books closed in hours. Manual document handling down 99%. Adoption follows when the metric is one the CFO already tracks — and stalls when the pitch is "transformation".
In regulated markets, the idea is rarely the moat. The deployment is.
The opportunity hiding in the difficulty
Here is the strategic consequence most miss: because deployment in regulated industries is hard, the companies that master it face a clear field. The compliance burden that filters out generic AI products is a feature of the market, not a bug. It's why we deliberately build where the walls are highest — financial crime, payroll, accounting, payments — and why each model generation widens the gap for products already wired into the workflow.
The models will keep getting better on their own. The deployment won't. That's the work. It's also the conversation we most like having.