# Why AI answers on marketing data go wrong, and the seven fixes Canonical URL: https://www.martechdevs.com/blog/ai-ready-marketing-data By [Jafrin Akter Alo](https://www.martechdevs.com/authors/jafrin-akter-alo) Data Engineer and Co-founder, martechdevs Published: 2025-11-18 Last updated: 2026-10-12 AI tools answer from whatever your stack gives them. When events fire twice, one customer has three IDs, revenue means something different in each team and nobody wrote the rules down, Claude or ChatGPT answers confidently and gets it wrong. AI-ready marketing data passes seven checks: tracking, identity, warehouse models, metric definitions, context docs, access and MCP. An AI tool asked for last quarter's CAC picks a spend table, a customer table and a formula. If the purchase event fired twice in March, if the CRM and the product hold the same person under different IDs, or if marketing and finance count new customers differently, the model has no way to know. It returns one number in a confident sentence, and the number gets questioned in the meeting. At martechdevs, we fix these gaps as [AI-ready data projects](https://www.martechdevs.com/ai-ready-data). The free [AI readiness checklist](https://www.martechdevs.com/martech-ai-audit) lists the same seven checks, and the [AI readiness score](https://www.martechdevs.com/tools/ai-readiness-score) scores your stack against them in a few minutes. ## Why do AI answers on marketing data go wrong? Because the model fills every gap with a guess. It doesn't know which revenue column finance trusts, which events were double-counted before a tracking fix, or that MQL changed definition in July. Gartner expects organisations to abandon 60% of AI projects unsupported by AI-ready data through 2026, and found that 63% of organisations either don't have or are unsure they have the right data management practices for AI ([Gartner, February 2025](https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk)). The payoff runs the other way too. Gartner's 2026 research found that organisations with successful AI invest up to four times more, as a share of revenue, in foundations such as data quality and governance, and that the most AI-ready data and analytics teams report up to 65% greater business outcomes ([Gartner, April 2026](https://www.gartner.com/en/newsroom/press-releases/2026-04-16-gartner-says-organizations-with-successful-ai-initiatives-invest-up-to-four-times-more-in-data-and-analytics-foundations)). ## What are the seven checks for AI-ready marketing data? They follow the way data flows: collect it right, join it, model it, define it, describe it, then decide who and what can reach it. ### The seven checks, in fix order 1. **Tracking:** Events fire once 2. **Identity:** One profile per person 3. **Warehouse:** Joined and tested 4. **Definitions:** One CAC, one MQL 5. **Context:** Rules written down 6. **Access:** A login per AI tool 7. **MCP:** Read-only first ### 1. Tracking accuracy Key events and conversions have to fire once, with the right values. One common gap: a GA4 purchase sent with an empty transaction ID makes GA4 merge those purchases into one, and GA4 only removes duplicate purchases on web streams ([Google Analytics Help](https://support.google.com/analytics/answer/12313109)). **Good looks like:** a written tracking plan, every event tested end to end, and conversion counts that match the CRM, with any gap explained. **First step:** compare one week of purchases or signups across GA4, your ad platforms and your CRM. ### 2. Identity resolution One person is a cookie on the website, a device ID in the app and an email in the CRM. Product analytics tools join them at login, and small mistakes split or merge people. Mixpanel's identity docs, for example, say reset() has to run on logout and that two different user IDs can't be merged ([Mixpanel docs](https://docs.mixpanel.com/docs/tracking-methods/id-management/identifying-users-simplified)). **Good looks like:** shared IDs such as user ID, email and device ID, with merge rules that turn each person into one profile. **First step:** pick five recent customers and follow each one from first visit to payment. Count how many profiles each has in each tool. ### 3. Warehouse models An AI tool that queries raw exports redoes the joins every time, a little differently each time. A warehouse such as Snowflake or BigQuery, with models built in dbt, does the joins once and tests them. **Good looks like:** tested models for customers, accounts, campaigns and revenue, refreshed on a schedule, with checks for freshness and duplicates. **First step:** list the five questions leadership asks most, and note where each answer comes from today. ### 4. Metric definitions If CAC, MQL, pipeline and churn mean different things to marketing, sales and finance, an AI tool picks one of them, and you won't know which. A semantic layer defines each metric once, in a place every report and agent reads: the dbt Semantic Layer, Snowflake semantic views or LookML, for example. Through MCP, Snowflake's Cortex Analyst answers over semantic views only ([Snowflake docs](https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-mcp)). **Good looks like:** each metric defined once, written down, and built into the semantic layer every report and agent uses. **First step:** write the definitions of CAC, MQL and churn on one page, and get marketing, sales and finance to agree them. ### 5. Context docs Agents also need the rules nobody wrote down: your ICP, naming conventions, which campaigns count as brand, and what sales means by qualified. Without them, the agent guesses at the rules your team follows. **Good looks like:** a marketing knowledge base for AI, with definitions, naming conventions, ICP, playbooks, and brand and compliance rules, shipped as skills and knowledge files your agents load. **First step:** collect the docs you already have into one folder and mark what's out of date. ### 6. Access and permissions AI tools get the access of whoever connects them, and connectors have already leaked data. In June 2025 Asana found a flaw in its MCP server that had exposed data across organisations for over a month, affecting about 1,000 customers ([BleepingComputer, June 2025](https://www.bleepingcomputer.com/news/security/asana-warns-mcp-ai-feature-exposed-customer-data-to-other-orgs/)). In September 2025 a fake postmark-mcp package, never published by Postmark, copied outgoing emails to an attacker ([Postmark, September 2025](https://postmarkapp.com/blog/information-regarding-malicious-postmark-mcp-package)). **Good looks like:** access set per team and per tool, a separate login for AI tools, and an owner for each system. **First step:** list every AI tool and connector with access to your CRM, ad accounts and warehouse, and who approved each one. ### 7. MCP readiness MCP servers are how Claude, ChatGPT and your own agents reach your tools. Most of the stack has an official one now. When we checked Claude's connector directory on 11 October 2026, it listed HubSpot, Salesforce, Snowflake, BigQuery, Klaviyo, Braze, Customer.io, Intercom, Mixpanel, Amplitude, PostHog, Clay and Apollo. Google's own GA4 server is local, experimental and read-only. **Good looks like:** MCP servers on the tools that matter, an internal MCP server for your ops teams, and access you can switch off at any time. **First step:** connect one read-only server, for example the warehouse behind a single agent, and test it on questions you already know the answers to. The [Snowflake MCP server guide](https://www.martechdevs.com/blog/snowflake-mcp-server) walks through that setup. ## What order should you fix them in? Tracking and identity first, because every model, definition and answer is built on them. Then the warehouse and the definitions, then context, then access and MCP. Connecting AI to data that fails the first two checks gives you wrong answers faster. ## How do you test whether AI answers are right? Write ten golden questions with answers you already trust, and ask them through the AI tool every time something changes: a new connector, a new model, a renamed field. If an answer drifts, the check it fails tells you which of the seven to fix. - What was last quarter's CAC, by channel? - How much new MRR did we add each month this year? - What share of MQLs became SQLs last month? - Which campaigns created the most pipeline this quarter? - How many trials converted to paid in the last 30 days? ## What does this look like for a 15-person SaaS? An illustrative example, before any AI is connected, for a 15-person SaaS on HubSpot, GA4 and Stripe: - **Tracking:** the signup event fires on page load and on submit, so GA4 counts most signups twice. - **Identity:** HubSpot contacts and product users share an email, with no user ID passed at signup. - **Warehouse:** none yet. Stripe and HubSpot numbers meet in a spreadsheet. - **Definitions:** finance counts a customer at the first paid Stripe invoice, sales at closed-won. - **Context, access and MCP:** HubSpot connected to Claude by one person with admin rights. That stack would score low on the [AI readiness score](https://www.martechdevs.com/tools/ai-readiness-score). The fix order would be the signup event, a shared user ID, then one agreed definition of a customer, well before a warehouse or more connectors. ## Checklist: AI-ready marketing data - [ ] Key events fire once, with values that match the CRM - [ ] One profile per person across web, app and CRM - [ ] Tested warehouse models for customers, campaigns and revenue - [ ] CAC, MQL, pipeline and churn defined once, in a semantic layer - [ ] Definitions, ICP, naming and playbooks written down for agents - [ ] A separate, limited login for every AI tool, with an owner - [ ] Read-only MCP access on the tools that matter, tested on golden questions ### Score your stack, then fix the gaps The AI readiness score takes a few minutes and gives you the fix order. At martechdevs, we then fix the gaps at a fixed price, usually live in weeks, and your team owns everything after handover. ## Frequently asked questions ### What is AI-ready data? Data an AI tool can use without guessing: tracked accurately, joined by shared IDs, modelled in a warehouse, defined once in a semantic layer, described in context docs, and reachable through access you control. ### Can we use Claude or ChatGPT on our marketing data now? Yes, for questions your data already answers well. Start read-only, with one connector and golden questions you can check, and widen access as each check passes. ### Do we need a data warehouse to be AI-ready? Not for every question. One tool's MCP server can answer questions inside that tool. Questions that join tools, such as CAC or pipeline by channel, need the data joined somewhere, and that's the warehouse's job. ### What is a semantic layer? A place where each metric is defined once, with its formula, filters and grain, so every report and AI tool computes it the same way. Examples include the dbt Semantic Layer, Snowflake semantic views and LookML. ### How long does it take to make marketing data AI-ready? It depends on which checks fail. Tracking and identity fixes are usually the quickest. Warehouse models and definitions take longer, because several teams have to agree them. A free audit gives you the order, and the quote gives you a fixed price and timeline. ### Is it safe to connect AI tools to the CRM? It can be, with limits: a separate login per AI tool, read-only access first, no email-sending tools in the same session as customer data, and an owner who can switch access off. ## Sources 1. Gartner, [Lack of AI-ready data puts AI projects at risk](https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk), 26 February 2025. 2. Gartner, [Organizations with successful AI initiatives invest up to four times more in data and analytics foundations](https://www.gartner.com/en/newsroom/press-releases/2026-04-16-gartner-says-organizations-with-successful-ai-initiatives-invest-up-to-four-times-more-in-data-and-analytics-foundations), 16 April 2026. 3. Google Analytics Help, [Minimize duplicate key events with transaction IDs](https://support.google.com/analytics/answer/12313109), read 11 October 2026. 4. Mixpanel, [Identifying users (Simplified ID Merge)](https://docs.mixpanel.com/docs/tracking-methods/id-management/identifying-users-simplified), read 11 October 2026. 5. Snowflake, [Snowflake-managed MCP server (documentation)](https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-mcp), read 11 October 2026. 6. BleepingComputer, [Asana warns MCP AI feature exposed customer data to other orgs](https://www.bleepingcomputer.com/news/security/asana-warns-mcp-ai-feature-exposed-customer-data-to-other-orgs/), 18 June 2025. 7. Postmark, [Information regarding the malicious postmark-mcp package](https://postmarkapp.com/blog/information-regarding-malicious-postmark-mcp-package), 25 September 2025.