Closing the trust gap

McKinsey has recently published a series of reports, on the state of AI, on the three horizons of AI transformation and on the state of AI trust in 2026. Together they highlight a problem that comes with every new wave of technology, the latest of which is AI. How do we trust it?

AI is complicated by several separate topics that get conflated, and it can only be trusted once you apply the context of the problem it is helping to solve.

1. The people and organisations pushing for AI adoption have varying messages and reputations.
2. The technology itself varies widely in cost and quality, and is often sold on a “caveat emptor” basis.
3. Legislation cannot keep pace with the human, cultural and business change that AI is driving.
4. The black-box nature of AI sparks aversion (distrust) and curiosity (worship) in equal measure.

I have used AI over the last 15 years at various stages of its maturity. It is far more accessible now than it has ever been, but it has one of the worst experiences for setting user expectations that I have ever encountered. What I expect turns out to be paper thin, and it takes a lot of work to get it to where I need it to be. None of that work comes pre-baked.

Who holds the licence?

When I get in my car and drive it, the car company is accountable for making sure the features work as described when I bought it. I am accountable for what happens on the journey and beyond, and I hold a licence that says I can be trusted with the car.

If I get in a self-driving car, who is accountable if something goes wrong, and who holds the licence? Is it the sensor manufacturer, the model trainer or the chassis assembler? That question is the jeopardy in introducing AI into familiar processes without revisiting accountability, because accountability is the foundation of trust.

Every time I drive, I subconsciously ask myself a string of questions. Do I trust the manufacturer? Do I trust my mechanic? Do I trust the county or country that maintains the roads? Do I trust the modifications I have made to my car? Have I paid my tax? Am I allowed to drive, and am I safe to?

When I am running a business, I ask who is accountable when a decision turns out to be a mistake. Is it the product, the assumptions about the market, the person I hired, the technology they are using, or the data that underpins the whole process? A good governance process drills down to the precise part of the process that is at fault and allows a fix that I can trust to last. The explanation is then given to each persona in their own language, so they can adopt the changes required. In short, the process is deterministic.

An AI driving licence

Let me flip this. How do we get AI to a point where the process it sits in, or replaces, is deterministic when I use it in my business?

Most of the time, as Jake Hall will cover in his Dark Factory series, it comes down to embedding every expectation of the mapped-out process (the manual, in other words) into the model, as a rebuttal witness to the part that makes the decisions. That is, in effect, the AI driving licence. It lets me trust that the model understands the rules of its environment and can be trusted within those guardrails.

The UK’s Highway Code has over 300 rules governing drivers. How many rules are in your AI rule set? More importantly, how many would you need it to follow before you trusted it with your future?