AI data hygiene is the work of making business information accurate enough, consistent enough and appropriately governed before an automation relies on it. For a UK SME, that usually means cleaning a narrow set of enquiry, customer or sales fields rather than attempting a company-wide data project.

The commercial reason is simple: an automated follow-up cannot repair a missing owner, an invented status or three duplicate versions of the same lead. It can only move the uncertainty faster.

Quick Answer

Before connecting AI to a CRM or sales workflow, choose one decision the data must support. Define the minimum fields for that decision, remove duplicates, make allowed values explicit, record where each field came from and keep uncertain records in a review queue. Do not upload a full customer database merely because a tool can accept it.

A useful first slice might clean the last 50 enquiries so the team can answer four questions reliably: who owns the lead, what service is requested, what information is missing and what approved next action is due.

Start With The Decision, Not The Database

"Clean the CRM" is too broad to test. "Identify quote requests that have no owner after one working day" is specific.

Write the decision first, then work backwards:

  • Which records are in scope?
  • Which fields are genuinely required?
  • What values are allowed?
  • What makes a record uncertain?
  • Who resolves the exception?
  • What proof will show that the cleaned data helped?

This prevents a tidy-up from becoming months of taxonomy work with no visible effect on response speed or revenue.

Create A Minimum Field Contract

A field contract names what each field means and how it can be used. For an enquiry workflow, the minimum may be name, safe contact route, service requested, received time, owner, status, next action and source.

Do not let free text carry a decision that needs a controlled value. "Hot", "urgent-ish" and "call soon" are difficult for people and software to interpret consistently. A short allowed list such as new, waiting for details, quote in progress, won, lost and closed is easier to review.

Keep an "unknown" state. Forcing a guess into every blank field makes the data look complete while reducing trust.

Deduplicate Without Erasing Context

Duplicate records are common when forms, inboxes, spreadsheets and CRMs all capture the same person. The safe first workflow should suggest likely matches and preserve source context for review.

Useful comparison fields include:

  • Normalised email address or telephone number.
  • Company name and domain.
  • Form submission time.
  • Service requested.
  • Existing owner or account reference.

Do not merge automatically when two people share a household, generic inbox or company switchboard. A reviewer should decide which fields survive and whether the records are genuinely the same enquiry.

Keep Source And Freshness Visible

Every important field should have a source that a reviewer can understand: website form, customer email, call note, imported spreadsheet or staff update. Add a checked date for values that can go stale, such as budget, project timing, consent preference or company role.

This is provenance in practical form. It lets an AI summary say "timing was stated in the 24 July enquiry" rather than presenting an old assumption as a current fact.

The ICO UK GDPR guidance is the right starting point for understanding data-protection responsibilities. The NCSC data-security guidance provides useful security context. Halo does not replace legal or information-security advice.

Minimise Before You Automate

Use only the fields the workflow needs. A quote-routing assistant rarely needs a full customer history, private notes or unrelated payment information. Redact test data where possible and document retention, access and deletion before live records enter the route.

Keep account invitations, API keys, bulk exports and real customer records out of an initial audit. A redacted sample and the field contract are usually enough to identify the first useful fix.

Build An Exception Queue

Good data automation does not hide uncertainty. It routes it.

Send records to review when:

  • The owner is missing.
  • The same contact appears in conflicting records.
  • A required field cannot be inferred safely.
  • Consent or contact preference is unclear.
  • The requested action would overwrite a human decision.
  • The record contains sensitive or unexpected information.

The queue should name the reason and the person responsible. "AI failed" is not a useful exception label; "two possible account matches" is.

Test With A Small Representative Set

Choose examples that include clean records, blanks, duplicates, unusual enquiries and records that must not be changed. Compare the automation output with a reviewed answer and record false matches, missing fields and unsupported inferences.

A useful first proof signal is not "the model processed every row." It is fewer manual corrections, faster assignment or a measurable drop in unresolved records without an increase in wrong updates.

Use the AI CRM automation guide for the downstream workflow and the admin automation guide when the main problem is copying between inboxes, sheets and task lists.

What To Send Halo

Send one redacted record set, the current field names, the decision the workflow should support and examples of the mistakes that cost time now. Halo can map a minimum field contract, review queue and proof signal before any live connection is considered.

Use the free AI audit to identify the smallest data-quality slice worth fixing first.

FAQ

What is AI data hygiene?

AI data hygiene is the process of making the inputs to an AI workflow accurate, consistent, appropriately sourced and limited to what the workflow genuinely needs.

Do I need to clean my whole CRM before using AI?

No. Start with the records and fields needed for one decision, such as enquiry assignment or quote follow-up. Prove that narrow slice before expanding.

Can AI merge duplicate customer records automatically?

It can flag likely matches, but ambiguous records should go to human review. Shared emails, phone numbers and company details can create false matches.

What data should I send for an AI audit?

Use a redacted example, field list and workflow description. Do not send passwords, API keys, full customer exports or live account invitations in the first request.

How should a UK SME measure data quality?

Use a practical measure tied to the workflow: fewer duplicate leads, fewer missing owners, faster assignment, fewer corrections or a lower unresolved-exception count.