The reason your CFO does not believe your marketing numbers is rarely that the numbers are wrong. It is that you have three of them.
Industry research through the first half of 2026 keeps landing on the same point: fragmented marketing technology, where planning, budgeting and measurement sit in disconnected systems, is now the thing blocking AI adoption and weakening the ability of marketing leaders to defend spend in front of a board.
That framing is right, but it stops one step too early. The fragmentation is not a tooling problem. It is a definitions problem that tooling made visible.
Key takeaways
- Three systems producing three different numbers is a definition failure, not a software failure.
- AI adoption stalls because models trained on inconsistent data produce confidently inconsistent answers.
- Fix the order: definitions, then joins, then automation. Reversing it wastes the most money.
- You almost never need to replace tools. You need one agreed key that lets them talk.
The three-number problem
Here is what we find in most mid-size accounts.
The ad platforms report conversions using their own attribution and their own lookback windows. Analytics reports sessions and goal completions using a different model. The CRM reports opportunities, using a definition of “qualified” that sales changed at some point without telling marketing.
Each system is internally consistent. None of them agree. And when the CFO asks a simple question, marketing answers with the number that flatters the channel being discussed, which everyone in the room senses even if nobody says it.
This is why budgets get cut. Not because the marketing did not work, but because nobody could produce a number that survived a follow-up question.
TRACKING AND REPORTING
One number your CFO will actually accept
We join spend, activity and outcome data into a single reporting layer, with definitions written down and agreed by both sales and marketing.
Why this blocks AI specifically
Plenty of teams have tried to point a model at their marketing data and ask it to find insight. The results disappoint, and the usual conclusion is that the model is not good enough.
The model is fine. The data is contradictory.
If your lead volume differs by thirty percent depending on which system you query, a language model will not resolve that for you. It will pick one, sound confident, and produce an answer built on the version it happened to read. That is worse than no answer, because it carries authority.
Every automation you build on top of inconsistent definitions inherits the inconsistency and adds speed to it.
A model trained on three contradictory versions of the truth will give you a confident fourth one. MyDigipal
Fix it in this order
The sequence matters more than the tooling. We run it in three phases, and phase one is the one everyone wants to skip.
Phase one: agree the definitions. What is a lead. What makes it qualified. When does an opportunity start. Who decides. Write it in one document, get sales to sign it, and date it. This costs nothing and takes a week of uncomfortable meetings. Teams that skip it spend the next six months building beautiful dashboards that nobody trusts.
Phase two: create one join key. Every system needs to carry an identifier that lets a record be followed from first touch to closed deal. Usually that is a CRM identifier pushed back into analytics and forward into offline conversion uploads. Without it, cross-channel attribution stays theoretical no matter which model you pick. We covered the modelling side in multi-touch attribution beyond last click, but the join comes first.
Phase three: automate the reporting. Only now. Once the definitions hold and the key exists, the reporting layer is straightforward work and the AI layer on top actually behaves, because it is reading one version of events.
The definitions document, concretely
Phase one sounds vague until you see what the output looks like. It is one page, and it answers five questions.
What is a lead. A form submission is not a lead if the email is disposable and the company field says “test”. Write the filter.
What makes it qualified. This is where sales and marketing genuinely disagree, and the disagreement is the point. Marketing usually means “matches our target profile”. Sales usually means “answered the phone”. Pick one, or name both and use different words for each.
When does an opportunity start. The moment a salesperson creates it, or the moment a discovery call happens. These are different dates and they produce different conversion rates.
Which touch gets the credit. Not the model, the rule. Do you credit first touch, last touch, or do you report both. Deciding this in advance stops the quarterly argument about which channel “really” produced the deal.
Who arbitrates. One name. When a definition needs to change, that person decides and dates the change, so that a shift in numbers can be explained by a documented decision rather than a mystery.
Get those five answered, signed by sales, and dated. Most of the value arrives before you touch a single system.
The objection: we already have a CDP
We hear this often, and it is worth addressing directly, because a customer data platform is exactly the tool that should solve this.
It solves the plumbing. It does not solve the definitions. A CDP will happily unify three contradictory versions of a qualified lead into one profile, and the profile will inherit the contradiction. You will have spent a licence to centralise the disagreement rather than resolve it.
The test is simple. Ask two people in different teams what your CDP counts as a qualified lead. If the answers differ, the platform is not your problem.
What good looks like
You know it is working when three things become true.
The number in the board pack matches the number in the CRM, and both match the number the agency reports. Nobody prepares a separate version for a separate audience.
A question like “what did paid social contribute to closed revenue last quarter” gets answered in a few minutes rather than a few days, because the join already exists.
And a channel that looks unprofitable on last-click can be defended with evidence rather than conviction, because its assisting role is visible in the same dataset. That matters, since the channel that closes is rarely the channel that started the conversation.
The uncomfortable arithmetic
Most teams spend more per year on martech licences than they would spend once on joining what they already own.
We have seen stacks with a customer data platform, two analytics tools, a business intelligence layer and a reporting add-on, none of them sharing a definition of a qualified lead. The licences renew. The confusion persists. Nobody wants to be the person who says the expensive stack is not the problem.
Start with the definitions document. If you cannot write down what a qualified lead is in one paragraph that sales agrees with, no amount of software will save the reporting.
When you are ready to join the data properly, that is the work we do first on every account, before touching a single campaign.
Sources: AI and marketing news roundup, June 2026 - June 2026 marketing news, trends and insights - Marketing technology and AI news, 10 June 2026
