Most personalisation projects are scoped as a capability problem. Which platform, which model, which segmentation approach, how quickly it can be stood up.

Framed that way, the projects usually succeed and the outcomes usually disappoint. The capability arrives and the experience does not noticeably improve.

The reason is visible in three numbers published across 2026, and they only make sense read in sequence.

The demand side is not in question

MarTech reported in April 2026, citing Adobe's 2025 AI and Digital Trends report, that 71% of consumers want personalised offers and information, and 78% expect a seamless experience across channels.

Those are not soft preference figures. Seven in ten actively want the thing, and nearly eight in ten expect the mechanics to work without them having to explain themselves twice.

For events this translates directly. An attendee who registered for one edition expects the next invitation to reflect that. An exhibitor who took a stand in a specific hall expects the renewal conversation to start from that fact rather than from a blank form.

The cross-channel figure is the harder of the two to satisfy in this industry. A single event touches a person through a registration site, a confirmation email, an app, a badge, a session room and a follow-up sequence, and those are frequently six different systems procured at different times by different people. Seamless is a high bar when the seams are contractual.

The capability side is also not in question

MarTech reported in July 2026 that 90% of customer experience organisations are piloting or deploying AI, drawing on the Five9 2026 Business Leaders CX Report.

The same research found those organisations split roughly evenly between end-to-end platforms, hybrid environments and best-of-breed combinations, with no single architecture pulling ahead.

That even split is more informative than the 90%. When nearly every organisation is deploying something and no dominant architecture has emerged, it usually means the differences between approaches are not what determines the outcome. Something else is.

The foundation is where it breaks

The third number explains the first two.

Adobe's 2026 AI and Digital Trends report, cited in the same April MarTech piece, found that fewer than half of organisations say their data foundation is adequate to support AI at scale. And in a 2025 MarketingOps study, cited by HubSpot in August 2026, only 16% of RevOps professionals said they trust the accuracy of their data.

Sixteen percent. Not the data of a vendor or a partner. Their own.

Read the sequence now. Buyers want personalisation. Organisations have deployed AI to produce it. The people closest to the underlying records do not believe those records are correct.

Personalisation built on data nobody trusts does not fail loudly. It produces confident, specific, wrong output at scale, which is considerably worse than producing nothing.

What untrusted data looks like at an event

Event data has a structural problem that most CRM data does not. It arrives in bursts, from several systems that were never designed to agree with each other, around a date that cannot move.

Registration holds one version of a person. The badge scan at the door holds another. The session check-in holds a third. The exhibitor's lead capture app holds a fourth, and often that one never comes back to the organiser at all. The same human being appears four times, with four spellings, three job titles and two email addresses.

Nobody in that chain is doing anything wrong. The systems are doing what they were bought to do. But the merged picture at the end is exactly the kind of record a RevOps professional would decline to vouch for, which is what the 16% is measuring.

Point a personalisation engine at that and it will address someone by a job title they left two years ago, or invite a returning attendee as though they had never come.

The failure is asymmetric, which is what makes it dangerous. Generic communication that lands slightly wrong is forgettable. Personalised communication that lands wrong is memorable, because the whole point of it was to demonstrate that you knew who you were talking to. Getting it wrong proves the opposite more loudly than saying nothing would have.

The trust number is a judgement, not a metric

It is worth being precise about what the 16% figure actually captures, because it is easy to misread.

It is not an accuracy measurement. Nobody audited those databases and found 84% of them faulty. It is a survey of confidence, which means it records what practitioners believe about data they work with daily.

That makes it softer as evidence and arguably more useful as a signal. These are the people best placed to know where the gaps are, and their answer is that they would not stake a decision on it. A team that will not trust its own records for a report will also not trust them for an automated message going to ten thousand people, which is why so much personalisation stays in pilot.

Why this rarely gets funded

Data foundation work has the worst possible shape for getting approved.

It has no launch, no demo, and no moment where anyone can see it. It is expensive, it takes months, and when it succeeds the visible result is that things which were already supposed to work now work. Nobody gets promoted for reconciling a registration table against a badge export.

Buying a personalisation platform, by contrast, has a vendor who will help build the business case, a timeline with a date on it, and a demo that shows the outcome. The incentives point consistently at the layer that cannot deliver the outcome without the layer nobody funds.

There is a budget-cycle version of this too. Foundation work spans quarters and platform purchases fit inside one, so the work that would make the purchase effective is structurally harder to approve than the purchase itself. The same asymmetry shows up in how AI budgets get allocated against readiness, covered in the gap between AI spend and the ability to scale it.

This is the same pattern described in the gap between adopting AI and running it. The purchase is legible and the wiring is not, so the purchase happens and the wiring does not.

The decision test is more useful than a data audit

There is a cheaper diagnostic than auditing the whole estate, and it produces a more honest answer.

Take a decision your team made recently on the basis of event data. A segment you targeted, a session you cut, a package you repriced. Then trace it back and ask which system the number came from, when it was last reconciled, and who would have caught it if it had been wrong.

Most teams cannot complete that trace for more than one or two decisions. That is the same finding as the 16%, arrived at locally rather than through a survey, and it identifies which specific records matter rather than declaring the whole estate suspect.

It also tends to reveal that the trustworthy subset is smaller and more useful than expected. This is the underlying argument in why most event analytics do not change decisions: the problem is rarely a shortage of data.

Running that trace has a second benefit. It surfaces who the human owner of each number is, and in most organisations several important figures turn out to have no owner at all. A number nobody owns is a number nobody corrects, which is how a small reconciliation error survives long enough to become the basis of a segment.

Narrow personalisation beats broad personalisation

The practical conclusion runs against how most teams sequence this work.

If only part of your data is trustworthy, personalise only on that part. One reliable field, used well, outperforms a dozen fields used speculatively. Knowing definitively which hall an exhibitor was in supports a better renewal conversation than a rich profile assembled from four systems that disagree.

That means the first question in a personalisation project is not what a platform can do. It is which facts about a person your team would defend in a meeting. Usually there are three or four. Build on those, and expand only as the foundation earns it.

There is a compounding argument for working this way too. Every field you personalise on becomes a field somebody has to keep correct, forever, across every system that touches it. A narrow set is not just safer at launch, it is cheaper to maintain in year three, which is when most personalisation programmes quietly stop being updated.

The alternative is what the three numbers describe: strong demand, deployed capability, and a foundation that quietly makes the output wrong.

None of this is an argument against personalisation, and it is not an argument for delaying it until the data is perfect, because it never will be. It is an argument for matching the ambition of the output to the confidence in the input, and for being honest internally about where that line currently sits.

If you are scoping personalisation for a show and want to work out which of your data is actually load-bearing, book a call.