Short answer: start with the questions you want AI to answer, then fix only the data those questions depend on. In practice that’s seven steps: decide the questions, clean records and duplicates, lock down lifecycle stages, make pipelines meaningful, repair source and attribution, set ownership and property rules, then test with real questions before rollout.
Everyone agrees you should fix your data before adding AI. Almost nobody says how.
So teams do one of two things. They connect Claude to a messy portal and get confident, wrong answers. Or they start a clean-up project with no end, and AI never arrives.
There’s a better way. Work backwards from the decisions you want to make, and fix the data those decisions rely on. This is the method we use. It sits inside our complete guide to Claude and HubSpot, and inside the Enable stage of our PACE methodology: systems before AI.
The short version
- AI doesn’t fix bad data. It makes bad data faster and more convincing.
- You don’t need a perfect portal. You need trustworthy data for the first use case.
- Start with the questions, not the clean-up.
- Fix in order: records, stages, pipelines, attribution, ownership.
- Test with real questions and compare the answers with what the team already knows.
- Expect two to six weeks for a typical mid-sized portal, depending on how far it’s drifted.
What does AI-ready HubSpot data look like?
Data a leader would quote in a board meeting without a caveat.
That’s the test. If your team already hesitates before quoting a HubSpot number, AI won’t remove the hesitation. It’ll hide it behind fluent text.
| Area | Not ready | Ready enough for AI |
|---|---|---|
| Records | Duplicates, blank owners, missing company associations | Duplicates merged, owners assigned, contacts linked to companies |
| Lifecycle stages | Used differently by sales and marketing, or skipped | Defined, written down and set by rules |
| Pipelines | Stages based on internal admin, deals left open for months | Stages match buyer progress, with entry and exit rules |
| Close dates and amounts | Expired close dates, placeholder amounts | Kept current through a weekly hygiene routine |
| Source and attribution | A catch-all source swallowing most contacts | Original source reliable; self-reported source captured on key forms |
| Properties | Hundreds of unused or duplicate fields | Core properties documented, unused ones archived |
| Permissions | Everyone sees everything | Access matches role |
Step 1. Decide the questions first
Write down the five questions leadership wants answered every week. They decide which data matters.
For example:
- Are we on track for this quarter?
- Which deals are most at risk, and why?
- Which channels create pipeline that closes?
- Where are deals stalling?
- Which customers are showing signs of churn?
Each question maps to a small set of objects and properties. “Which deals are at risk?” depends on deal stage, close date, amount, last activity date and associated contacts. It doesn’t depend on your blog subscriber list. That’s where you focus.
If you want a starting set, use our 15 questions every revenue leader should ask HubSpot every week.
Step 2. Clean records and duplicates
Merge duplicates, assign owners and fix associations for the records your questions use.
- Duplicates. Use HubSpot’s duplicate management tool for contacts and companies, where your subscription includes it. Agree a rule for which record wins.
- Owners. Every open deal and every target company needs an owner. Unowned records drop out of accountability and out of analysis.
- Associations. Deals with no contact or company are invisible to account-level analysis. Fix them.
- Required fields. Make the properties your questions rely on required at the right stage, so the problem doesn’t come back.
Don’t try to clean every contact you’ve ever captured. Clean what the first use case touches.
Step 3. Lock down lifecycle stages
Define each stage in one sentence, agree who can move a record, and automate the rest.
Lifecycle stage is one of the most-used properties in any analysis. If marketing sets MQL one way and sales ignores it, every conversion rate Claude calculates is wrong.
- Write a one-line definition for each stage.
- Decide which stages are set by workflow and which by people.
- Stop stages being skipped or moved backwards without a reason.
- Write it down where the team can find it.
Step 4. Make pipelines mean something
Pipeline stages should describe buyer progress, not internal admin. And open deals should be real.
- Stages. Each stage should reflect something the buyer has done, like agreeing a discovery call or receiving a proposal. Add entry and exit criteria.
- Close dates. Clear expired close dates in one review. Then make updating them part of the weekly sales routine.
- Amounts. Replace placeholder values. A pipeline full of £1 deals or round guesses produces a meaningless forecast.
- Stale deals. Close out deals with no activity for an agreed period, or move them to a nurture stage.
This is where AI forecasting works or fails. Forecast confidence relies on stage age, activity and close-date history. Our article on advanced revenue reporting shows what becomes possible once this data is sound.
Step 5. Repair source and attribution
If you can’t say where your best deals came from, AI can’t either.
- Check original source on recent contacts. If most sit under one catch-all value, fix the tracking and form set-up behind it.
- Add a “How did you hear about us?” field to high-intent forms. Self-reported attribution captures demand that tracking misses.
- Make sure web forms create or update contacts in HubSpot, and that deals are associated with those contacts.
- Agree one attribution view for leadership. It doesn’t need to be perfect. It needs to be consistent.
Step 6. Set ownership and property rules
Clean data only stays clean if someone owns it.
- One owner per number. Pipeline comes from HubSpot. Invoiced revenue comes from finance. Decide which system owns which figure and say so.
- Property governance. Document your core properties. Archive unused ones. Agree naming conventions for new ones.
- Permissions. Check that HubSpot roles match what each person should see. The HubSpot connector for Claude uses those permissions, so they matter more once AI is involved. See our guide to connector security and permissions.
- A weekly hygiene routine. Fifteen minutes a week from each deal owner prevents the drift that caused the clean-up.
Step 7. Test with real questions
Before rollout, ask Claude the questions from step 1 and compare the answers with what the team already knows.
- Ask each question in a consistent format.
- Check the numbers against a HubSpot report built by hand.
- Where they differ, find out why. It’s usually a data issue, not an AI issue.
- Fix it, and ask again.
- Only share outputs with leadership once two weeks in a row match.
This step turns “we think the data’s fine” into evidence. It also gives you the first version of a prompt library.
Want the shortcut? Our AI-Powered HubSpot Audit has Claude read every record in your portal, read-only, and scores it out of 100 across 13 areas. You get a prioritised fix plan in about two weeks.
How long does it take?
Typically two to six weeks for a mid-sized B2B portal, focused on one use case. These ranges come from our own delivery work, not a published benchmark.
| Portal condition | What usually needs doing | Typical effort |
|---|---|---|
| Well maintained | Light dedupe, close-date clean-up, test questions | One to two weeks |
| Drifted | Stage definitions, pipeline rework, attribution fixes | Three to four weeks |
| Neglected or post-migration | Rebuild of stages and properties, mass association fixes, permissions review | Five to six weeks or more |
Readiness scorecard
Score one point for each statement that’s true. Five or more and you’re ready for a first use case.
| Statement | Score |
|---|---|
| Leadership trusts at least one HubSpot pipeline report without a caveat | 1 |
| Fewer than one in twenty open deals has an expired close date | 1 |
| Every open deal has an owner and an associated contact and company | 1 |
| Lifecycle stages are defined in writing and set by rules | 1 |
| Pipeline stages have entry and exit criteria | 1 |
| Original source is reliable for most new contacts | 1 |
| HubSpot permissions match each role | 1 |
| Someone owns data quality as part of their job | 1 |
Under five? The RevOps health check shows where the leaks are, and how to audit HubSpot for revenue leaks walks through a full review.
Can Claude help clean HubSpot data?
Yes, carefully. Claude is very good at finding problems and less suited to fixing them in bulk.
Through the HubSpot connector, Claude can read every record and flag duplicates, expired close dates, deals with no contacts and inconsistent stages. It can also create and update records, 10 at a time. With the write tools set to “Needs Approval”, you see the proposed changes before anything is written, and every change shows in HubSpot’s audit log.
That suits targeted fixes reviewed by a person. Big clean-ups are better done with HubSpot’s own tools, imports and workflows, with Claude finding and prioritising the work.
What should you do next?
Pick one use case. Fix the data it depends on. Test it. Then expand.
If you’re deciding which use case comes first, read where to start with AI in HubSpot. If you want to see the end state, read how we built an AI revenue analyst using Claude and HubSpot.
Not sure how ready your HubSpot is? Free until 12 January, book a 30-minute Day One session. We look at your HubSpot live with Claude and send you the three fixes worth making first, in writing, within 48 hours.
Key takeaways
- Start with the questions. They tell you which data to fix.
- Fix in order: records, lifecycle stages, pipelines, attribution, ownership.
- Keep data clean with owners, rules and a weekly routine.
- Test Claude’s answers against known numbers before sharing them.
- You need trustworthy data for the first use case, not a perfect portal.
Frequently asked questions
How clean does HubSpot data need to be before using AI? Clean enough that leadership trusts the reports the first use case depends on. Perfection isn’t required. Consistency is.
How long does it take to prepare HubSpot data for AI? Typically two to six weeks for a focused first use case, depending on how far the portal has drifted.
Can Claude help clean HubSpot data? Yes. It can find and prioritise issues across every record and make small, reviewed updates, 10 records at a time. Bulk clean-up is usually better handled with HubSpot’s native tools.
Which HubSpot properties matter most for AI? For revenue use cases: lifecycle stage, deal stage, close date, amount, owner, original source, last activity date and the associations between contacts, companies and deals.
Should we fix everything before we start? No. Fix what the first use case needs, prove the value, then extend the clean-up to the next use case.
