1) Field hygiene: tidy the few fields AI actually needs
Start with a short list of properties that most AI tasks use: email (or contact ID), first name, last name, company, job title, lifecycle stage, owner, created date, last contacted, UK postcode and a short "notes" or "lead description" property. In HubSpot, go to Settings → Properties → Contacts and make sure each exists and is the correct type (single-line text, dropdown, date).
Standardise formats with two quick HubSpot actions: create a saved list to find messy values, then bulk-edit or import to correct them. Example filters for the list: Email is unknown OR Email contains a space; Phone number does not start with +44; Postcode does not match pattern (use contains/does not contain to catch obvious problems). Export that list, fix in a spreadsheet (lowercase emails, strip spaces, normalise phone numbers to +44, uppercase postcodes with correct spacing), then Import → Update existing records using email or HubSpot contact ID.
Keep it small: aim to fix the top 500 recently-active records first (use a list filtered by Last Contacted or Created date). That gives a useful, clean slice for testing AI without needing a full clean-up.
2) Consent, provenance and quick dedupe
You need explicit flags that mark records as usable for AI. Create two lightweight contact properties: AI ready (single checkbox) and Data source (dropdown: webform, import, manual, partner). Put them in Settings → Properties → Create property and add them to the contact record layout so editors can see and set them.
For consent, check HubSpot's existing subscription and marketing contact fields (Subscriptions, Marketing contact, Legal basis). Use lists to find contacts missing clear consent: filter where Marketing contact is false OR Subscription status is unknown. For any records you intend to feed to AI, mark AI ready true only if consent is explicit or the data is non-identifiable.
Quick dedupe tactics (afternoon-ready): use HubSpot’s Manage Duplicates (Settings → Data Management → Duplicates) if available; otherwise export a CSV of recent contacts and run a spreadsheet dedupe on email, phone and company+name combinations. For priority merges, use HubSpot’s contact view: find pairs and Merge (select primary record, click Actions → Merge). Focus on high-value records first (largest deal amount, recent activity, or owners' top 50).
3) Build a safe test CSV and a simple prompt to validate outputs
Create a safe test set by making a HubSpot list: Contacts → Lists → Create list (Active). Filters to include: AI ready is true, Last contacted within 18 months (or Created within 2 years), and exclude anyone with sensitive tags. From the list view choose Actions → Export and select only the properties your test needs (contact ID, first name, company, job title, postcode, lifecycle stage, short notes). Exclude emails, phones and full addresses to avoid sharing PII.
In the exported CSV, anonymise remaining names if you prefer (change "Emma Smith" to "Contact 1") and remove any free-text notes that contain sensitive content. That file becomes your safe sample to feed an AI. Keep it to 50–200 rows so you can iterate fast.
A simple validation prompt you can run in any LLM: "You have this CSV with these columns: [list columns]. For each row, produce a one‑sentence buyer persona and list which fields you used. If the data is insufficient to create a persona, reply 'INSUFFICIENT DATA' and say which field is missing." Run this on your sample and check two things: 1) does the model cite the exact fields it used, and 2) does it say INS UFFICIENT DATA when obvious gaps exist. If the model hallucinates (adds invented facts) or uses fields it shouldn't, add those cases to your 'AI ready' gating and expand the sample fields.
If you want a short, practical pairing session to run these checks on your HubSpot instance, Optira can help run an afternoon workshop and hand the team a safe test CSV and prompt set.