WHAT THE RECORD DOESN'T TEACH

2 days ago
6 min read

Image of tools with text overlay Who you gonna call?

In 'Learning Loops & Intelligence Capital' we set out a hypothesis: the organisations that get the most from AI will be the ones that build extensive learning loops.

The learning loop is a system which every successful organisation already operates. At least in an analogue, human sense. It embeds people's current expertise into collective information, knowledge and intelligence. It then informs future judgements, procedures and behaviours.

Which means, run continuously, it sharpens the organisation every time work is done. New information and knowledge is captured, allowing for better and better decisions to be taken.

It also leads to differentiation. Because each organisation's learning experience is unique, the learning loop delivers unique DNA no competitor can clone.

The ability to accelerate learning new loops will, we believe, be one of the most critical abilities of any organisation taking full advantage of digital intelligence.

This analysis takes this idea and asks what that means in an advanced manufacturing context.

Documented to the letter

In manufacturing, everything is written down. Add decades of detailed regulation and you'll end up with one of the best-documented industries in the world. Medtech, for instance, is replete with design history files, batch records, compliance reports and audit trails. Together they tell you what was done, when it was done, against which specification and on whose authority.

And yet – this is far from a complete blueprint. Documents do not capture everything.

Outside of the corpus remain considerations and judgements that only happen inside human heads – the reason why a result is borderline, the thought behind a pause to look again before a sign-off. Or the logic of the call that looks wrong to most – but is 100% right because the responsible person has seen a dozen cases like it before.

The record, the document, that which is written down, holds the outcome of that moment of judgement – the decision, the timestamp, the criteria applied. It can prove that the decision was controlled.

What it cannot do, however, is teach the next person how to make a similar decision again.

The lesson industries are paying for

Over three years, Ford hired around 350 veteran engineers – it calls them 'gray beards' – to mentor younger staff and rework AI tools that had been built without engineering context.

Charles Poon, Ford's vice-president of vehicle hardware engineering, admitted the company 'didn't pay as much attention as we should have' to the experience of its most knowledgeable people. And last month it came top of the mainstream brands in J.D. Power's 2026 Initial Quality Study, its first time since 2010, with roughly $1bn in warranty and material cost savings expected this year.

Few are doubting the causal link with the gray beards. So much of what they knew – and have now reintroduced – wasn't captured in Ford's documents.

Which is also why no one learns to drive entirely from the rules of the road, as helpful as they are. Drivers learn by driving, supervised, until the judgement is their own. And to oversimplify but make the point – a work instruction is the plant's rules of the road; running the line is the driving.

Documents are necessary. But they're never sufficient.

Because there's always a distance between work-as-prescribed and work-as-done: expert judgement measures the distance between them. And the measure of that distance shapes a site's productivity.

Which is why every plant has the people it depends on without quite meaning to. The ones colleagues go to first when a process starts misbehaving. The reviewer who ties this week's odd result to one from years back.

But people retire, they change role, they move to suppliers and competitors. Organisations expect around half their workforce to retire or move on within five years, and fewer than one in ten say they consistently capture what those leavers know.

Two kinds of capital

Most generative AI in regulated manufacturing hasn't been helpful here either. It's often pointed straight at the documented layer – procedures, records, specifications – because that's the layer easiest to govern. But, as we've already established, wholly insufficient.

Treating intelligence as a form of capital helps us reorganise the picture. As 'Learning Loops & Intelligence Capital' explains, if you organise the business correctly there are two kinds of intelligence capital: human capital – the judgement your people hold, which the enterprise rents and goes home every evening. And agentic capital – expert AI which carries the evaluations, memory and decision traces that make them controlled enough to trust in regulated work, which the firm can own. Together they are the sum of an organisation's intelligence capital.

The shift underneath is an accounting function: the judgement an enterprise used to rent, it can begin to keep. Add a learning loop so both human and agentic intelligence capital improves and the owned asset compounds. The engine is a five-stage cycle: elicit, structure, deploy, capture and evaluate.

How judgement becomes an asset

A novice meets uncertainty everywhere. But a 20-year veteran meets far less of it, because most of what was once uncertain has hardened into conditions they read almost without thinking – what passes QA, what gets escalated, what's worth checking first.

A career on the production line is one long act of compression: uncertainty worked down into heuristics, heuristics worn smooth into routine. Pharma even has a name for the compression that never makes it onto paper – the golden batch, the run seasoned operators can reproduce by feel but struggle to specify. And the distance between what they can do and what they can write down is the distance between expertise and an asset.

So how to run that compression at the level of the firm, not invisibly inside individual heads?

The first step is to capture consequential decisions as they're made: the context, the signal, the options weighed, the call, the person accountable, the evidence, the reasoning, the outcome – and any correction that came later.

It's one reason Brightbeam open-sourced CHAP: a shared protocol that gives humans and agents a common language for the work, and signs each decision into an audit trail – who proposed what, who was accountable, what was overruled. And why.

Using CHAP, every reviewed call, every logged override, every escalation becomes training signal. Which means the asset improves with use instead of decaying with turnover.

One reviewer's hard-won judgement on a difficult Tuesday afternoon becomes the site's standard, then the next site's, at a marginal cost that falls each time.

The expert hasn't been replaced, and hasn't been overruled. Their reasoning has simply stopped leaving the building.

The sovereign learning loop

There's another strategic reason to own the asset rather than rent it: one part of the system is guaranteed to change. The model. Capability that demanded a frontier model 18 months ago now runs on hardware a site can own – and the frontier keeps moving.

Build your advantage on a particular model, and it commoditises underneath you, on someone else's release schedule.

But build it on the asset and the relationship inverts: every new model that arrives makes the same asset more valuable, because the record of how your organisation decides travels, intact, from one model to the next.

A competitor can license the models you can, and hire people from the same pool. But what they can't buy is the compounded record of how your organisation, specifically, has resolved its hard decisions over years.

That record isn't for sale, because it isn't a product – it's a by-product, emitted by your own people doing their own work, and kept.

Own the record of how you decide

Regulation taught medtech and biopharma to keep an impeccable record of what was done.

The next advantage belongs to the firms that also keep a governed record of the nuances of how they decide – and put it to work.

That's what Brightbeam means by integrating digital intelligence: connecting the model to the humans, their intelligence and judgements, the controls and the audit trail of the organisation using it. Do this well and the two intelligences compound instead of running in parallel.

Whether the obligation is written as FDA quality-system regulation or the EU AI Act, it points the same way: on decisions of record, digital intelligence can become an expert part of the loop. Even as expert humans stay inside the loop and accountable for each decision taken.

Sources