Written by: Nicole Ogonowska, IT Growth Manager, Digital Colliers

Here's the test I keep coming back to. Ask your stack a question that touches four systems at once. Something like: which customers in the top revenue decile opened a support ticket last week, then downgraded, and were also flagged by the fraud model? If the answer takes a week and three people, your tooling isn't working. It doesn't matter how much you've spent on it.

Custom platforms tend to fail this test in a specific way. Each system works fine on its own. The CRM is fine. The billing system is fine. The product analytics warehouse is fine. The support tool is fine. The failure lives in the seams. And the seams are where the real questions live.

Why cross-system questions are the actual benchmark

Single-system questions are easy. Your finance tool tells you MRR. Your product tool tells you WAU. Your support tool tells you ticket volume. None of these are the questions that change what you do on Monday morning.

The questions that change behaviour span systems. Which onboarding cohort produced the customers who are still active at month 12? Which feature usage pattern predicts a renewal conversation going badly? Which support tickets from paying customers are still open past SLA and involve a bug we already shipped a fix for?

If you can't answer these in an afternoon, your stack is a filing cabinet, not a decision system. And this is where AI has been quietly overpromising. MIT's Project NANDA found that 95% of enterprise GenAI pilots deliver zero measurable P&L impact. S&P Global reported that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before. The pilots that die are almost always the ones that tried to answer cross-system questions without fixing the seams first.

The cost of not fixing this

Operators underestimate what the status quo actually costs. It's not just the analyst hours. It's the decisions you don't make because the question is too expensive to ask.

The pricing change you don't run because segmenting the affected cohort takes two weeks.
The churn intervention you don't ship because you can't tell which accounts are actually at risk this month.
The feature you keep investing in because nobody has pulled the usage-by-revenue-tier view in six quarters.

Add the compliance clock on top. EU AI Act transparency obligations start on 2 August 2026. High-risk obligations follow on 2 December 2027, with fines up to 15M euros or 3% of global turnover for the serious stuff. If you can't answer cross-system questions today, you also can't answer the auditor's version of them in eighteen months.

And the tempting shortcut, having AI coding assistants glue everything together fast, has its own tax. One empirical study found that more than 15% of commits from every AI coding assistant introduce at least one issue, ranging from 17.4% for Copilot up to 29.1% for Gemini. Unresolved technical debt from AI-generated code climbed from a few hundred surviving issues in early 2025 to over 100,000 by February 2026. Speed without structure is just future debt with a modern logo.

What the working setups have in common

The operators who can answer cross-system questions on a Tuesday afternoon tend to share a few things. None of them are exotic.

One canonical customer ID that survives across CRM, billing, product, and support. Not aliased. Not fuzzy-matched at query time. Actually the same key.
A modelled layer between raw system exports and the questions people ask. Call it a warehouse, call it a semantic layer, call it whatever. The point is that the join logic lives in one place, not in eleven dashboards.
A short list of questions the business actually asks, written down. When a new system enters the stack, the test is whether those questions still work. If they break, the system doesn't ship until they're fixed.
AI used narrowly, on top of a clean data layer, for specific tasks with a measurable answer. Not as a universal translator across broken plumbing.

The honest diagnostic

Pick three questions your CEO asked in the last quarter that required data from more than one system. Time how long the answer took, and how many people touched it. Then ask whether you'd bet your next board meeting on the answer being right.

That's your benchmark. Everything else is decoration. The 88% of AI proofs-of-concept that never reach production, per IDC and Lenovo, mostly died because they were trying to answer a cross-system question on top of a stack that couldn't even agree on who a customer was. Fix the seams first. The clever stuff works after that, not before.

Why cross-system questions are the actual benchmark

The cost of not fixing this

What the working setups have in common

The honest diagnostic

Sources

Key-person risk is an AI use case

The bottleneck is the interface, not the data

The take-home test is dead. What replaced it.