// AI MARKETING OPERATING MODEL
Five Systems, One Customer, Zero Agreement
A customer buys. Marketing has a conversion. Sales has a closed deal. Finance has an invoice. Ask what brought that customer in, and the agreement ends.
One commercial relationship can produce several legitimate records. Trouble starts when an AI workflow treats those records as interchangeable.
You can connect the CRM, analytics, ad accounts, sales tools, and accounting system to the same model. It will still need an answer to a question the connectors cannot settle: which evidence controls this decision?
Before buying another AI tool, I would settle that. Otherwise, the team gets a faster way to argue about the same customer.
Why do marketing systems disagree about the same customer?
Marketing systems count different events, identities, dates, and attribution windows. Ad platforms assign advertising credit, analytics observes trackable activity, CRM records commercial stages, sales logs interactions, and finance verifies financial outcomes. Define evidence precedence before automating decisions.
An AI marketing data layer should preserve those differences while making the records usable together. It needs shared definitions, reliable identifiers, timestamps, source references, and rules for resolving conflicts. Connecting five APIs does not supply any of those automatically.
This is the third piece in my AI Marketing Operating Model series. The first established who owns the decision. The second covered the test before the A/B test. This one deals with the evidence that both depend on.
CRM: a commercial record with definitions someone must own
A CRM can tell you which account owns a deal, who is responsible for it, and what stage it has reached. It cannot make the stage definition true. If a rep marks a deal won before procurement finishes, the record reflects the update. Whether the deal meets the company's booking policy is a separate check.
As reviewed October 1, 2026, HubSpot's lifecycle-stage documentation describes stages for contacts and companies and supports customized processes. That flexibility is useful. It also means a stage label needs a local definition before it becomes evidence for an automated action. HubSpot's lifecycle-stage documentation explains the product behavior; your company has to define the commercial meaning.
Start with the object being counted. A contact is not an account. An account can have several contacts and several opportunities. A new opportunity at an existing customer is not necessarily a new customer. If the model receives only a count and a label, it may never see that distinction.
I would want the CRM extract to retain the account and opportunity identifiers, stage history, effective date, last update, and owner. A current-stage snapshot cannot reconstruct what the team knew at the time of a budget decision. A late update can make last month's pipeline look different without any new customer behavior.
For a sales-accepted opportunity decision, the agreed CRM definition should usually take precedence over a form submission in analytics. For collected cash, it should not. The record's authority depends on the question.
Analytics: the journey you can observe
Analytics is useful for understanding what happens on the site or app: visits, events, referrals, paths, and friction. Its coverage depends on implementation and the signals available. A buyer moving from a work laptop to a phone may be stitched into one journey, appear as separate users, or leave a gap.
GA4's reporting-identity documentation, checked October 1, 2026, describes how identity settings affect the journey shown in reports. User-ID can help connect activity when it is implemented appropriately. It does not grant visibility into every buyer or device. Google's reporting-identity documentation is the starting point for checking what your property actually uses.
A CRM vs GA4 discrepancy can begin with nothing more dramatic than counting different events. The CRM might count an accepted opportunity while GA4 counts a form submission. A repeat submission might create another event without creating another customer. Neither total is a useful substitute for the other.
Compare the event definition, identity, date range, time zone, filters, attribution settings, and processing delay before blaming the connector. Preserve those settings with the extract. A number detached from its configuration is difficult to reconcile later.
AI referrals add another limitation. A visit from an AI answer can lose referral information or arrive after the buyer returns through another route. I cover that separately in the AI attribution gap and AI referral conversion by engine. Missing source information should remain missing. A model should not fill the gap with a plausible origin story.
Ad platforms: conversion credit within a defined view
Platform-reported conversions answer a question defined by the platform's settings and available signals. Which event counts? How long after an interaction can it receive credit? Are views included? Is the number assigned to the interaction date or the conversion date? Those details can change the apparent result.
Google Ads' discrepancy guidance, reviewed October 1, 2026, identifies attribution and reporting differences as causes of mismatched totals. Before comparing reports, use equivalent columns and definitions. Google Ads' discrepancy guidance provides the product-specific checks.
Two platforms can claim credit for the same purchase. Adding their attributed totals does not deduplicate the customer. It also does not establish which purchase would have happened without advertising. Attribution allocates credit under a method. Incrementality asks what changed because the treatment changed.
That does not make an ad platform useless or dishonest. Its reporting can be valuable for delivery diagnostics and optimization inside the account. The mistake is promoting an account-level reporting convention into the final authority for the whole business.
When an AI recommends moving spend, require it to identify the conversion definition and source behind the recommendation. If its rationale says only that one platform has a better return on ad spend, the team still needs to know whether revenue is actual, estimated, gross, net, or duplicated elsewhere.
Sales activity: what happened between the lead and the deal
Sales activity can explain a discrepancy that no attribution setting will fix. Was the meeting held? Did the buyer attend? Was the account eligible? Did the rep accept the handoff? Was a diagnostic request treated as a demo request? Those events change how a campaign should be evaluated.
Activity counts alone do not establish quality. A sent email is not a reply. A booked meeting is not an attended meeting. A completed call is not an accepted opportunity. Before aggregating those events, decide which one the test or workflow is supposed to improve.
The data layer should keep the link between the activity and the relevant contact, account, and opportunity where that link exists. It should also preserve whether the record came from a system event, a rep's entry, or a model's classification. A summary of a transcript should not silently become an official sales stage.
AI can help classify objections or flag missing handoff details. If a person corrects the classification, save the correction with the evidence that justified it. Test future workflow versions against those reviewed examples. That gives the team a learning loop without pretending every correction retrains the underlying model.
For pipeline claims, follow the event through to the accepted commercial outcome. My Proving AI Pipeline to Your Board article covers that chain. A stronger handoff is worth measuring. It should not be described as revenue before the later events occur.
Finance: bookings, recognized revenue, and cash are different
Finance does more than count cash. It can distinguish signed commitments, invoices, recognized revenue, collections, refunds, credits, and margin. The right financial measure depends on the decision. A deposit is cash received, but it is not automatically revenue earned. A booked deal can be real while payment remains outstanding.
That distinction matters when marketing feeds value back into bidding or reports customer acquisition economics. Decide which financial event is eligible, whether refunds adjust it, which currency applies, and when the value becomes final. A gross order total cannot settle a contribution-margin question.
Do not fix discrepancies by declaring that Finance wins every argument. Finance can confirm a financial outcome; an accounting record generally cannot establish which ad caused it. Equally, a platform's attributed revenue cannot override the financial records when the question is how much money was collected.
For AI workflows, provide the relevant financial status and its effective date rather than a single field called revenue. Where definitions need interpretation, have Finance approve them. Code can reconcile the records against that policy; AI can help explain the exceptions. It should not invent the policy while writing the recommendation.
Evidence precedence: decide who wins for each question
Evidence precedence is the order of authority used to resolve a disagreement for a specific decision. It is not a permanent ranking of software vendors. The source that controls cash collection does not automatically control website behavior or causal advertising impact.
I would write the following rules before authorizing budget changes:
- Did money arrive? Use reconciled payment records, including refunds and payment reversals.
- Did Sales accept an opportunity? Use the agreed opportunity-stage definition and its supporting CRM record.
- Did the buyer complete a tracked site action? Use the validated event definition and tracking record, with its known coverage limits.
- What credit does an ad account assign? Use that platform's report and retain its attribution configuration.
- Did marketing cause additional business? Use a suitable causal study with a defined outcome and uncertainty. None of the preceding records settles that question alone.
Then name the person who resolves exceptions. If a duplicate account, missing identifier, or late sales update could change the decision, the workflow needs a review path. The answer can be unresolved. Forcing a single clean answer is how uncertainty gets hidden.
Keep the reconciliation small enough to inspect. For the decision being made, retain the original source reference, event definition, identity match, event time, extraction time, reporting window, and correction history. You do not need to centralize every field from every application to settle one disputed metric.
Start read-only. Compare the records and show where they agree, where definitions explain the difference, and where evidence is missing. Only consider automated actions after the team can inspect those disagreements reliably. Permissions should follow the action's consequences, not the number of connectors the vendor sells.
Why platform self-reporting cannot settle incremental impact
A better conversion count can mean better measurement. It does not necessarily mean more customers. If a tagging change recovers previously missed purchases, reported conversions can rise even when the underlying sales volume has not changed.
On May 5, 2026, Google announced Meridian GeoX for geographic incrementality testing and said testing would begin later that year. The same announcement introduced Meridian Studio, described as an enterprise platform on Google Cloud for managing marketing-mix models. Those were product announcements, not evidence of results at your company. Google's May announcement also reported an average 14% conversion lift for Google tag gateway advertisers, citing Google internal data for the Finance category globally, comparing July through December 2024 with January through June 2025. That is a reported conversion figure, not independently established incremental sales.
The status changed after May. Google's official Meridian announcement dated September 4, 2026 described GeoX's general availability. It would be stale to call it merely an announced future tool today. Availability still does not validate a particular experiment, outcome definition, or budget recommendation. Google's September release announcement documents that change.
As reviewed October 1, 2026, the official Meridian GeoX repository describes cross-publisher geographic testing, study design, analysis, and calibration of marketing-mix models. It is a causal-measurement library, not a customer identity service or an automatic reconciliation layer. A geographic test can provide evidence about incremental impact when its assignment, comparison, outcome measurement, and analysis are appropriate. It still needs checks for interference, other changes during the study, insufficient volume, and uncertainty. A product name cannot remove those conditions.
That is also why I separate platform capabilities from the operating work around them. In my research on agency operating models for 2027, the section on platforms absorbing the obvious work describes expanding native capabilities. The remaining job includes choosing the question, checking the inputs, resolving ownership, and deciding whether the result is strong enough to act on.
What I would let AI do with this data
Use code for identifier matching, deduplication rules, currency arithmetic, date normalization, and exact comparisons. Use AI where interpreting language helps: explaining a documented mismatch, classifying a sales objection, or proposing a question for review. Keep the records behind the explanation available.
Before changing a budget, the workflow should return the proposed action, controlling evidence, conflicting evidence, unresolved assumptions, and the person authorized to approve it. If the controlling source is stale or absent, hold the action. Do not quietly substitute the easiest report to retrieve.
Measure the reconciliation itself. Review a set of disagreements that people have resolved, then test whether the workflow reaches the same conclusion and identifies the same missing evidence. Count incorrect resolutions separately from cases it correctly sends for review. A confident explanation of the wrong customer is still a failure.
After an approved action, retain the result and the later outcome. If the original assumption fails, update the rule or workflow and retest it. That is how the system becomes more useful without rewriting the history of what the team believed.
Before an AI moves money, make someone accountable for which evidence it is allowed to believe.
Frequently asked questions
Why don't my CRM and Google Analytics numbers match?
CRM and Google Analytics can count different events, identities, dates, and attribution windows. Compare the metric definition, customer or account identifiers, time zone, filters, reporting settings, and data freshness before treating the difference as an integration failure.
Which marketing system should be the source of truth?
Choose the authoritative source for each decision. Finance controls reconciled financial outcomes, CRM controls agreed commercial stages, analytics controls validated tracked events, and ad platforms report their own attribution. Incremental impact requires suitable causal evidence rather than one universal source of truth.
Can AI fix inconsistent marketing data?
AI can flag mismatches, classify evidence, and help explain conflicts. People must define the business rules, and code should perform exact reconciliation. AI should preserve missing evidence and route unresolved conflicts for review rather than invent a complete customer journey.
What is evidence precedence in marketing measurement?
Evidence precedence is the agreed order of authority used to resolve conflicting records for a specific decision. It identifies which source controls the outcome, what supporting evidence is required, and who resolves exceptions. The order changes with the question being answered.
Should I trust platform-reported conversions?
Use platform-reported conversions within their documented attribution settings and coverage. They can guide account optimization, but adding totals across platforms can duplicate credit. They do not independently establish collected revenue or the incremental business caused by advertising.
Sources
Primary documentation reviewed October 1, 2026. Undated documentation is identified by review date, not presented as a new 2026 publication. Product announcements retain their publication dates.
Google Ads, data discrepancy guidance
Google Analytics, reporting identity
HubSpot, contact and company lifecycle stages
Google, May 5, 2026 measurement announcement
Google, September 4, 2026 Meridian release announcement
Google, Meridian GeoX source and methodology
About the author
Jeff Brokaw is a CMO and Certified Chief AI Officer who builds the commercial systems around AI, including measurement, workflow design, revenue operations, and human approval for consequential actions.