Whichever of the three ways you run your AI, getting it right takes the same four steps. Most organisations start at step three:

1. The redesign. Rebuild the decision workflow around the judgement your experts already apply, then add what AI now makes possible. I work with the person who knows the workflow best, including where it fails. Skip this, and you automate a workaround nobody needs.
2. The spine. From the redesign comes the boundary: which decisions are safe to hand to AI, and which always need a person, however confident the system is.
3. The guardrail. The infrastructure that enforces the boundary on each live case. This is hard engineering, and some of the best minds in AI are working on it.
4. The automation. The work that runs inside the boundary: fast, repeatable, and safe because someone decided it was. This is what most AI vendors build and sell today.

Almost everywhere, the first two are missing. That is my focus.

How a decision flows once the boundary is set

Phase 1: The Decision Spine Diagnostic

This is where we start: Two weeks. One workflow. Fully mapped.

We begin by focusing on one high stakes process. That is where AI or software helps make decisions about people, such as screening applicants, flagging performance or handling claims. We look at how it actually runs, not how it’s documented. Most organisations have others, but one done properly shows you where the gaps are and gives you the pattern for the rest. You leave knowing:

• Every decision in that process, and which ones AI makes on its own
• Which should stay with AI and which need a person
• Who has the authority to say no at each step, and whether they really can
• What record you’ll need to prove each decision was made properly

You’ll find where your boundary is missing before a regulator, a lawyer or a headline does.

Phase 2: Implementation

A longer engagement, typically three to six months, to put the boundary in place: first using workflow built out in Phase 1, then to other high stakes workflows.

I work with your team on who holds authority at each decision, what they check, how it fits the way the work actually runs, and the record that shows it held. Where AI decides alone, that means who owns the boundary, and how you’ll know if AI starts drifting outside it.

The guardrail and the automation are built by your engineering team or technology vendors. I give them the boundary they need to enforce, and stay involved while they build it.

I don’t simply hand over a diagram and leave.

If a wrong call in your system has real consequences, and you can’t say, by name and in writing, what it’s authorised to decide alone, that’s where most of my conversations start.

Free 30-minute conversation. Contact me: mattsheehan@spatialnext.io

Cigna kept the human in: a doctor signed every denial, in an average of 1.2 seconds. Robodebt took the human out: Australia’s welfare system raised 470,000 wrong debts because a gap in the figures was enough to send a notice. Opposite designs, same failure.

Nobody had decided what the software could do alone, what needed a person, or what that person was there to do.. Read more: [Cigna] · [Robodebt]

The same gap shows up well beyond hiring and claims. In Aurora, Colorado, a family was held at gunpoint when only one of three checks on a licence plate matched; nobody had decided a partial match wasn’t enough to act on. In Baltimore County, a Doritos bag was mistaken for a gun and a 16-year-old was handcuffed. A signal became an action, and nobody had decided in advance whether the signal was enough.

That’s the gap my practice exists to close. Not after the harm is done. Before it.

The work starts with the Decision Spine Diagnostic: two weeks, one workflow, fully mapped. Then we set the boundary together.

Let’s have a free 30-minute conversation: mattsheehan@spatialnext.io

A WordPress.com website