When judgement becomes infrastructure

I’ve been developing and testing a marketing system with a client that brings together LinkedIn company engagement, intent data, CRM and commercial context, qualification and routing logic, and AI-assisted research and interpretation.

Using it in a live environment makes the interesting questions appear quite quickly.

The machinery can surface potentially interesting companies faster. It can connect activity that would otherwise sit across several tools. It can research organisations, compare signals and help work out what might deserve attention.

That is genuinely useful.

But quite early in the work, another problem became more interesting.

Before any of those potentially interesting things becomes an action, the system needs to know what they mean commercially.

An account might look relevant from the outside. But is it already a customer? Is there already a live deal? Is somebody already working it? Is this fresh demand, or activity around a commercial relationship that already exists?

An early example made this concrete. CRM data identified an account in the system as an existing customer, so it belonged in the customer state rather than being treated as a new opportunity.

That sounds obvious when a human says it.

The important bit was making the system know it too.

More signals don’t remove the need to interpret them

This is where some of the conversation about marketing automation gets the emphasis wrong.

The interesting question isn’t simply how much information we can collect, or how many actions an automated system can trigger. It is what we decide that information means.

Engagement is not automatically intent.

Intent is not automatically a buying window.

Fit is not automatically priority.

A relevant company is not automatically a relevant opportunity.

And context is not always another signal to add to the pile. Sometimes it changes what the signal means.

An organisation becoming more engaged with your brand may increase your interest in it. Intent data may give you another reason to pay attention.

But CRM status can change the classification entirely. An existing live deal can change the route. An active sales conversation can make an otherwise sensible automated new-business action completely inappropriate.

Different facts have different authority over different decisions.

The challenge is therefore not just to connect more data. It is to decide which information should influence interest, which should affect probability, which changes classification and which should override what the behavioural data appears to suggest.

Those are marketing and commercial judgements.

Once the system starts acting on them, they become part of its architecture.

The rule matters as much as the machinery

A human marketer looking at an account can absorb quite a lot of context without describing every step.

“That company is already a customer.”

“There’s a deal open.”

“We spoke to them last week.”

“They fit the profile, but there’s no reason to do anything yet.”

Some of that is knowledge. Some is experience. Some is judgement.

Automation forces us to be more explicit.

If customer status should change how an account is treated, that needs to become part of the logic. If an active opportunity should suppress a new-business action, the system needs to know that. If a combination of signals deserves more research but not sales outreach, that distinction has to exist somewhere too.

A signal requires judgement. Useful judgement can be codified. Once codified, it can become part of the infrastructure through which future decisions are made.

Good automation turns judgement into infrastructure. Bad automation does too.

The upside is that a sound decision does not have to be rediscovered every time the same conditions occur.

The downside is that a bad assumption can become wonderfully efficient.

If we decide that every sufficiently engaged account should go to sales, automation can apply that rule quickly and consistently. The workflow can run perfectly. The routing can be flawless. The research can be excellent.

The underlying decision can still be wrong.

This problem predates generative AI by quite some distance.

In her 1983 paper Ironies of Automation, Lisanne Bainbridge argued that automating industrial processes could change or even expand the problems left to human operators rather than simply removing them. Her subject was process control, not marketing, but the underlying tension is familiar: automation changes where the difficult decisions sit.

Generative AI has made that problem relevant to more forms of knowledge work.

It didn’t invent it.

So why not automate the judgement as well?

There is an obvious objection.

If we can describe the judgement clearly enough to put it into the system, why keep asking a human to make the same decision?

Quite often, we shouldn’t.

Humans are inconsistent. Experience can become habit. Senior marketers are not magically immune to weak assumptions or bad pattern recognition. Manual approval is not intrinsically valuable just because a person is involved.

If a decision is sufficiently well understood, making it repeatable is one of the benefits of automation.

Ajay Agrawal, Joshua Gans and Avi Goldfarb make a useful distinction between prediction and judgement. Prediction estimates uncertain states or outcomes; judgement, in their formulation, determines the value attached to possible outcomes and actions. Their work also shows why better prediction can sometimes make more of the decision itself automatable.

That is more useful than the reassuring version in which machines predict while humans retain permanent control of all the important decisions.

Some judgement absolutely should become automation.

If the objective is clear, the relevant inputs are observable, the decision is repeatable, errors can be detected, feedback arrives quickly enough and the cost of getting it wrong is acceptable, continuing to make the same decision manually may add little except friction.

The problem isn’t automation making decisions. The problem is automating a decision before we understand it well enough.

And understanding it means more than being able to write down a rule.

We need some reason to believe the rule reflects commercial reality. We need to know what might invalidate it. We need to recognise where apparently similar situations are materially different. And we need to learn from what happens after the system acts.

Otherwise we haven’t codified good judgement.

We have just made an assumption executable.

The marketing role moves upstream

This changes where senior marketing judgement is most useful.

If systems can increasingly collect information, connect signals, research accounts and execute routine responses, there is little value in a senior marketer manually inspecting every case simply to preserve a human step.

The harder work moves upstream.

What is the system actually trying to optimise?

Which evidence matters for that decision?

Which facts have authority over others?

When should commercial context override behavioural signals?

What should be excluded?

Where does ambiguity justify stopping or escalating an action?

What happens afterwards, and how should that change the logic next time?

Those are not really AI questions.

They are marketing questions made more consequential by automation.

The marketer’s contribution shifts from making every individual decision towards improving the decision logic that many actions depend upon.

That should also make judgement more accountable, not less.

A rule can be challenged. An exclusion can be tested. A classification can turn out to be wrong. Sales can tell us that accounts the system considered interesting were not useful at all.

The aim is not to take an experienced marketer’s instincts and hard-code them forever.

It is to make the reasoning explicit enough to use, inspect and improve.

There will still be ambiguous cases. There will still be incomplete commercial context. There will still be decisions where the consequence of getting something wrong justifies human review.

But “keep a human in the loop” is too crude a principle.

The machinery is going to get better at finding signals, researching them and acting on them.

That makes the quality of the decision underneath the action more important, not less.

Better automation shouldn’t simply create more actions. It should make better decisions easier to repeat.