Trust

Automation & AI

6 min read

16 Sep 2026

What Makes an AI Recommendation Trustworthy?

AI is increasingly making recommendations about where to spend budget, what to change and what to do next. But how do you know when an AI recommendation is worth acting on?

In paid media, AI recommendations are becoming part of everyday account management. Platforms suggest changes, AI agents flag opportunities and third-party tools advise marketers on where to focus their attention.

The obvious question is whether those recommendations are correct.

The more useful question is whether you can understand why the system made them.

A trustworthy AI recommendation should give you enough information to examine the evidence, understand the level of confidence, consider the commercial context and challenge the conclusion before you act.

What makes an AI recommendation trustworthy?

A recommendation can sound convincing without actually giving you much reason to trust it.

A neat explanation such as “Shift more budget into Campaign A because it has the strongest ROAS” sounds reasonable. But it leaves some important questions unanswered.

What data drove the recommendation? Over what period? How reliable is the conversion data? Does the campaign generate valuable customers, or simply cheap conversions? What assumptions has the system made?

Without that context, the explanation isn’t really helping you evaluate the recommendation. It is simply making the recommendation easier to accept.

This matters as AI takes on more responsibility in paid media. An AI system doesn’t become trustworthy simply because it can explain its conclusions. It becomes more trustworthy when you can examine the basis for those conclusions.

Four things every AI recommendation should show

There are four things we should expect from an AI recommendation before acting on it.

1. The evidence

What data actually drove the recommendation?

If an AI agent recommends moving budget between campaigns, you should be able to see the performance data behind that recommendation rather than being presented with the conclusion alone.

2. The confidence

How certain is the system?

A useful AI agent should be able to acknowledge uncertainty when the evidence is weak, incomplete or contradictory. “I’m not very confident” can be far more useful than a definitive answer based on limited data.

3. The commercial context

Does the recommendation understand what a good outcome actually means for the business?

A campaign with the highest ROAS isn’t necessarily the campaign you should put more money into. Margin, customer value, stock, capacity and the quality of the customers being acquired can all change what a good recommendation looks like.

4. The reasoning

Can you trace the recommendation back to the evidence that produced it?

You don’t necessarily need to see every technical step an AI system has taken. But you should be able to understand the logic well enough to challenge it.

When an AI recommendation gives you all four, you have something you can interrogate.

When it gives you none, you’re being asked to take the answer on trust.

Why AI confidence doesn’t mean accuracy

One of the biggest problems with AI recommendations is that confidence and accuracy aren’t the same thing.

An AI agent can be extremely confident in a recommendation based on poor, incomplete or misleading data. In some cases, that can make a recommendation more dangerous because the confidence discourages further investigation.

Imagine an agent recommends moving a significant amount of budget into a campaign because it has stellar ROAS and the system is highly confident in the recommendation.

It sounds compelling.

But the team checks the underlying data and discovers that the recommendation is based entirely on platform-reported conversions that haven’t been reconciled against actual sales.

The confidence was genuine. The evidence wasn’t good enough.

This is why confidence should be treated as another piece of information, not as proof that an AI recommendation is correct.

AI recommendations need commercial context

This becomes particularly important in paid media because advertising platforms don’t necessarily have a complete understanding of the business behind the account.

A platform can see clicks, conversions, revenue and other performance signals. It doesn’t automatically know that a particular product has a much higher margin, that a customer has unusually high lifetime value, that stock is running low or that the sales team can’t handle another 500 leads this month.

Those factors can completely change what the right recommendation looks like.

An AI system that understands your commercial context can make a different recommendation from one that is simply optimising the metrics available inside an advertising platform.

That’s why transparency isn’t just about showing the numbers. It’s about showing the numbers that actually matter to the business.

AI agents still need human oversight

It’s tempting to think of trust as something you establish with an AI system once.

You test it. It gets things right. You become comfortable with it. Eventually, you start approving its recommendations without looking too closely.

That’s where oversight starts to disappear.

An AI agent can be useful without being infallible. In fact, the ability to check its recommendations is part of what makes it useful.

Trust shouldn’t mean accepting everything an AI system says. It should mean having enough visibility into its recommendations to know when to accept them, when to challenge them and when to investigate further.

That’s the approach we’ve taken with Propel Max. Recommendations are designed to be grounded in the underlying evidence, confidence and commercial context, so the person using the system can understand what is driving the recommendation rather than simply being presented with an answer.

The goal isn’t to remove human judgement. It’s to give that judgement better information to work with.

Trustworthy AI should show its working

The AI recommendations worth acting on aren’t necessarily the ones delivered with the greatest confidence.

They’re the ones you can take apart and check.

If you can’t see why a recommendation has been made, you can’t properly evaluate it. You’re being asked to trust the system rather than assess its work.

As AI takes on more analysis and decision-making in paid media, that distinction will become increasingly important.

The question isn’t simply whether an AI agent can make a recommendation.

It’s whether it can show enough of its working for you to decide what to do with it.