I was part of the team that took programmatic in-house at Foxtel and IAG but as I've said, it really just hasn’t been that good. The tools just aren’t there yet. And the problem is…it goes beyond media.

The most common uses for AI in marketing are the optimisation of media buying and the personalisation of product or content recommendations.

These media buying and personalisation use cases are sound – when they’re solving a clear problem and when it’s your problem.

They are sound because they have well defined problems, they utilise structured data and have clear outcomes.

Fruity AI deployments – beyond media

Of course, there are much fruitier deployments of AI in marketing: To generate brand insights; replace humans in ad testing; to generate copy or creative; to replace people monitoring social media and to have conversations with customers.

These are fruity because the problems are not well defined. They’re using unstructured data which is much harder to create viable training sets from and the outcome – what good looks like – is much less clear. They are much less likely to be sound.

The media optimisation case seems sound. But it’s not...yet. The tools aren’t there yet. 

When I’ve been in positions where a company’s media was run externally, I was a pain-in-the-arse client.

It didn’t make sense that if someone searches for your brand, clicks your ad, then buys your product that we say that ad “drove” the sale. Nor did it make sense when you placed an ad in front of someone who tried to buy your product 10 minutes ago, then bought your product five minutes later.

When we pressured an agency to deliver more sales every month, and they then shifted budget into branded search and retargeting because they knew that would make the numbers go up, I was the annoying shit who pointed out that was what was happening. That it was just touching sales...but the overall sales numbers weren’t moving with it.

As marketers, the AI actually never served us...

Last click attribution won’t die and must - really

At the crux of all these problems is the stubbornly persistent industry standard method of last-interaction attribution. This does not even attempt to differentiate sales caused by the advertising from those that would have happened anyway.

And all those AI algorithms...the ones that power your ad buying…they’re still optimising to this goal.

I think it’s safe to say your goal as a performance marketer is extra sales. Not more ads to achieve the same sales - that would be dumb right?

But that’s literally what is happening. In media optimisation cases, the AI is not serving the marketer. It’s goal is to touch more sales on the way through, which makes a case for the value of the media buy but doesn’t stand up to scrutiny as a case for having delivered more sales.

Facebook researchers, armed with the data from 15 robust experimental measurement cases involving billions of impressions, found that there was no touch-based attribution method that could estimate the incremental impact of advertising as measured by a treatment/control experiment – the scientific gold standard. We could fix this. We could use AI that optimises to incremental sales. But unfortunately I’ve not had the opportunity. In almost all cases that I’ve run, these experiments have been duds. No sales lifts.

I fully expect that with different products, and armed with the right tools, we’d see different results. Genuine sales lifts for cheaper, more frequently purchased products where sales respond more directly to advertising.

Personalisation is good but it’s not – yet

The personalisation case is much less fraught. But also much less common. There are businesses that have a clear problem. Large retailers have thousands or millions of products and limited opportunity to put these in front of a consumer. Media platforms and publishers have thousands of pieces of content, and again, very limited opportunity to put these in front of a consumer.

The AI has a clear task. It serves the right content for an individual at a point in time to produce a clear and desirable outcome. More views, more sales, more ad impressions, more engagement.

So is it the right thing to do, for a brand that has a small number of products, to create enough content that they too have that problem? I’ve seen plenty of cases where a personalised marketing project has been stumped by the realisation, well after commencement, that they suddenly need to generate 10x the content.

Generally speaking, if you need to create the problem so that you can deploy the solution...you don’t need it.

That’s probably why we’re starting to get sick of it.

Real AI solutions for real marketing problems...

There certainly are AI-appropriate problems, with clear goals, where machines can do a better job than people.

And it's certainly true that advances in image recognition, speech recognition and speech synthesis mean that there are a much wider range of things that machines can do just as well as, if not better than people.

There are lots of these AI-appropriate problems. But identifying them requires a bit of understanding. A bit of knowledge of what AI can do, and its maturity in different domains. It requires a bit of investment in scientific AI capabilities. Smart people who actually understand machine learning, and natural language processing, conversational UX and image recognition. The kind of people who seem to have a knack for bluntly telling you when a use case is dumb.

Performance media optimisation could absolutely be one of these problems if we threw out standard attribution, although the brands for which it is appropriate would be much fewer than those that use it today – and reported returns would be lower. But the relationship between reported sales and actual sales would be unquestionable.

Where do we start?

Within the wider scope of marketing - the good old ‘Four Ps’, for instance - there are loads of appropriate problems. The optimisation of offer allocation or product configuration or product bundling. Or routing the right customer to the right contact centre agent and using virtual agents to handle components of calls, or even entire calls to take handling time out of contact centres. This is useful stuff. 

But these are hard problems where you will generally need to build an AI solution.

But when you build it, just make sure to build it to serve you.