Why omnichannel still breaks, and what AI does and doesn't change

Omnichannel was an ownership problem that got sold as a platform problem. AI is about to walk into the same trap, except this time the broken data shows up in the answer where everyone can see it.

Over the last few years I have sat in meetings where ‘AI’ gets approved, on both sides of the table, and the script rarely changes. Every AI budget signed off this year is sitting on the same unjoined customer data that defeated the marketing gods only a few years ago. When the pilot underwhelms everyone, nobody will see that the data was never joined - all they see is that the AI was not good enough, and they go and buy a different one or wait till the technology becomes more ‘Mature’.

The promise had a price tag

The single view of the customer was never some fuzzy aspiration that drifted in from a conference keynote. It was a line item with a business case attached, and somebody’s bonus rode on it. (I know mine did)

Between 2012 and 2014, four vendors went shopping. Salesforce bought ExactTarget in 2013 for about $2.5 billion. Oracle announced its purchase of Eloqua in December 2012, then picked up Responsys the following year. Adobe bought Neolane in 2013, and IBM bought Silverpop in 2014.

None of them built a unified customer record. They stitched acquisitions onto their existing products, put the new logos on one side of the slide and their own on the other, and sold it as omnichannel.

And it was a good slide. I sat through those demos. One customer, one timeline, every touchpoint in order, and you could feel the room relax, because somebody had finally solved it. So we thought…

Then came the implementation, and it turned up four customers. Sometimes seven.

Nobody owned the join

This is the part I can speak to without looking anything up. In 25 years I have never seen a true single pane of glass over one customer. Not in the organisations I worked inside, and not in the ones I only advised.

Plenty had the platform. Several had a programme named after it, with a launch event to match.

The reason is boring and it is not technical. Identity resolution is the hard part of the job and it belongs on nobody’s scorecard. Two teams own half a customer each, both of them hit their numbers, and the gap in the middle belongs to neither.

Worse, the join sits between two budgets. Anything that sits between two budgets gets a steering committee instead of an owner, and you cannot hold a steering committee to a match rate.

[illustrative] A big-four bank, about 2019, three years into a marketing cloud rollout. The email team and the app team each reported a single customer view. Neither could reconcile their active-customer counts, the gap was around 15%, and both dashboards were green.

Then in 2016 the industry did what it always does, which is relabel the problem. David Raab founded the CDP Institute, the customer data platform became a category, and the category described what a customer data warehouse plus identity resolution had always described. Now with a procurement path. (The procurement path was the innovation.)

What AI changes, and what it exposes

Two things are genuinely different this time, and I want to be fair about the first one because it is real. Fuzzy matching, entity resolution, reading unstructured text, working out that fourteen spellings of a job title are all the same job title. Language models are good at this. Rules never were.

I have written those rules. They were awful.

The second thing is less comfortable. AI takes away the place bad data used to hide. A dashboard will bury a broken join inside an average and nobody notices for two years. An AI states the wrong thing in one fluent, confident sentence, somebody forwards it to a director, and the director acts on it.

It is worth separating two things that get sold as one. AI that helps build the join is infrastructure work, and you can measure whether it worked. AI that sits on top and answers questions is a consumer of the join, so it inherits every flaw underneath it. Most of what is being pitched right now is the second kind wearing the first kind’s business case.

The raw material also got worse while the ambition got bigger, which nobody seems keen to mention. Apple shipped App Tracking Transparency with iOS 14.5 in April 2021 and mobile attribution broke. Mail Privacy Protection landed with iOS 15 that September, and the email open rate stopped meaning anything at all.

The platforms already ran this play

If you want to know how this goes, look at what the ad platforms already did to you. Google launched Performance Max globally in November 2021. Meta followed with Advantage+ in 2022. Both took the targeting decisions away from the buyer, and both called it optimisation.

That is the shape of it. AI gets handed the decision, not the data. It will make confident calls inside your broken joins, and it will report success against whatever it can actually see.

Where I am wrong

Two places this argument is weak, and I would rather say them myself than have them said to me.

First, “AI will not fix your data” is too strong, and I have caught myself saying it. This is the first technology that can do a real share of the join. Dismissing that is the same lazy pattern-matching I have spent this whole piece accusing vendors of.

The honest version is narrower and less fun. AI can help build the join. It cannot own it. The tool was never the blocker.

Second, and this one nags at me. The single pane of glass may always have been the wrong goal. Nobody ever needed one unified customer record, they needed a good-enough join for one specific decision. It is possible the single view survived only because it demos beautifully, and we spent a decade chasing a screenshot.

That changed how I ask the question. I stopped asking whether an organisation has a single customer view. Now I ask which decision it cannot make today, and what fixing the join would be worth.

What to do on Monday

Do not buy the platform.

Pick one decision that matters. Name the person who owns the join for that decision, a person with a name, not a forum. Make match rate their number, reviewed in the same place the rest of their numbers get reviewed.

One decision. One owner. One metric. If that fails, no platform was ever going to save you.