Clean Your CRM Before AI: How To

Why clean the CRM first, and what does clean actually mean?

Clean the CRM first because AI reads your records as fact and answers with confidence either way. A model pointed at a database where one customer exists three times, half the industry fields are blank, and a third of contacts left their jobs two years ago will produce fluent, plausible, wrong answers. We covered why this kills AI projects in your CRM might be killing your AI. This guide is the how: the actual cleanup, in the order that works.

Clean does not mean perfect, and chasing perfect is how cleanups stall. Clean means four measurable things, judged against what the AI will be asked to do:

• Unique: one real-world customer equals one record, with duplicates under roughly 5 percent.
• Complete: the fields the AI needs are filled on at least 80 percent of active records.
• Consistent: the same fact is recorded the same way everywhere, so grouping and filtering work.
• Current: records reflect this quarter, with dead and departed records archived out of the working set.

The order of work matters as much as the work itself. Deduplicate first, because there is no point polishing fields on records that are about to be merged. Standardise and fill fields second. Settle ownership and the master-system question third, because that is what stops the mess coming back. Then, and only then, add guardrails and connect AI.

Before starting, take measurements: total records, duplicate rate, completeness on your key fields, share of records untouched in 18 months. Ten minutes of exports gives you a baseline, and the baseline does two jobs. It sizes the work honestly, and it lets you prove afterwards that the cleanup happened, which matters when the next question is what the AI build should cost.

How do you deduplicate a CRM properly?

Deduplicate by matching on stable identifiers, merging with explicit survivorship rules, and working in passes from safest to riskiest. Done in that order, dedupe is careful clerical work. Done casually, it destroys data you cannot get back, so the method matters.

Match on what does not change. Names are the worst matching field: "ABC Pty Ltd", "ABC Proprietary Limited", and "ABC" are one company, while "J Smith" might be three different people. Match companies on ABN first, then website domain, then phone. Match contacts on email address first, then mobile. Only fall back to fuzzy name matching for records nothing else catches, and treat every fuzzy match as a suggestion for a human, not a decision.

Set survivorship rules before merging. When two records merge, which values win? A workable default: keep the most recently updated value for contact details, keep the earliest creation date, keep all activity history from both records, and never let a filled field be overwritten by a blank one. Write these rules down. The person doing the merging should never be inventing policy record by record.

Work in passes. First pass: exact matches on stable identifiers, safe to merge in bulk. Second pass: near matches flagged for quick human review, which for a few thousand records is hours of work, not weeks. Third pass: the genuinely ambiguous handful that needs someone who knows the customers. Do not auto-merge the ambiguous cases to finish faster. A wrong merge, two genuinely different customers collapsed into one, is worse than a duplicate and much harder to unpick.

Two traps to avoid. Parent and child companies are not duplicates: three branches of one group may be legitimately separate records that share an ABN, so decide the rule before the pass, not during it. And export a full backup before the first merge. Every experienced person does this. Every burned person wishes they had.

What does field hygiene involve?

Field hygiene means making the fields the AI will use complete, consistent, and constrained, and ignoring the fields it will not. This is where cleanups either get focused or die of scope, so start by shortening the job.

Pick the fields that matter. List what the AI will actually be asked to do, then work out which fields those tasks depend on. Summarising pipeline needs stage, value, owner, and close date. Segmenting customers needs industry, size, and status. That is usually eight to twelve fields. The other forty can wait, and most can be deleted: fields nobody has filled in for two years are clutter that confuses both staff and AI.

Replace free text with lists. Free-text fields are where consistency goes to die. If industry is a text box, you have "Construction", "construction", "Building", and "Bldg + civil" as four different industries. Convert the fields you group and filter by into fixed dropdown lists, then map the existing mess onto the list values. This single change does more for AI usefulness than any other field work, because grouping and filtering are what AI does constantly.

Standardise formats. One format for phone numbers, one for dates, states as consistent codes. Mechanical work, mostly automatable, and it makes matching and reporting reliable.

Fill strategically, not completely. Fill missing values on active, high-value records first, using data you already hold in invoices and email rather than guessing. Do not burn weeks completing fields on dormant records.

Archive the dead. Contacts gone for years and deals long closed should leave the working dataset. Archive rather than delete, so history survives for reporting, but get them out of what the AI reads as your current business. A model reasoning over 12,000 records where 5,000 are historical is answering questions about a business that no longer exists.

Who owns the data, and which system is the source of truth?

Ownership and mastery are the unglamorous decisions that determine whether your CRM is still clean in a year, and they are decisions, not tools. No software setting substitutes for them.

Name one owner. One person is accountable for CRM data quality: not the whole team, and ideally not the owner of the business. Their job is not typing in everyone's data. It is holding the standard: running a monthly duplicate check, chasing incomplete records back to whoever created them, and approving any new field before it gets added. Field creep is how CRMs got messy in the first place; somebody has to be allowed to say no. Without a named owner, entropy wins and you are re-running this cleanup in eighteen months.

Decide the master. Your customer data almost certainly also lives in your accounting package, your job system, and a spreadsheet or two. For each type of information, decide which system is the authority. A split that works for most SMEs: the CRM masters relationship data (contacts, pipeline, activity), while the accounting system masters financial facts (legal entity name, ABN, billing terms). Then make the systems agree with the master, ideally by automated sync, at minimum by a written rule about where each fact gets edited. The full method is in the single source of truth guide.

Set the sync direction deliberately. Data should flow from master outward. Two-way sync between systems that both accept edits is how conflicts breed, so if you cannot automate it cleanly, a one-way flow with a weekly exception report beats a clever loop that quietly corrupts both ends.

Why this matters for AI specifically: an assistant connected to the CRM answers from the CRM. If the CRM is not the agreed truth for customer data, the AI is authoritative-sounding and wrong by design, and no amount of record cleaning fixes a mastery problem.

How do you keep it clean, and what does this cost?

Keeping a CRM clean costs a fraction of cleaning it, but only if guardrails go in while the data is still good. Four that do most of the work:

• Validation at entry. Make key fields required, use the dropdown lists you built, and turn on duplicate warnings at record creation, which most CRMs include and few SMEs enable. Every bad record blocked at entry is one you never clean.
• A monthly hygiene pass. The data owner spends an hour on duplicate and completeness reports. Monthly, drift is trivial to correct. Yearly, it is a project again.
• Standards where staff work. A one-page rule sheet: who creates records, which fields are mandatory, where each fact is edited. New starters get it on day one, because untrained staff recreate mess faster than any owner can clean it.
• Quarterly measurement. Re-run the baseline numbers from before the cleanup. If duplicates or blanks are creeping, you find out in months, not after the AI starts answering strangely.

On cost and effort: a capable staff member can run this cleanup internally over a couple of months of part-time effort, and for a small, single-system CRM that is a reasonable path. It gets harder when records number in the tens of thousands, when several systems hold conflicting copies, or when nobody internally can be spared to hold the standard afterwards. Done as a focused engagement, this is exactly what a Data Readiness Sprint covers: two to six weeks from $6,000 plus GST, ending with measured, guarded data rather than a one-off tidy.

If you are not sure how bad your CRM actually is, measure before deciding anything. The Business Readiness Audit does that for $1,990 inc GST, fixed scope, with written findings on your real duplicate rates, completeness, and system overlaps, fully credited toward any engagement within 90 days. Either way, the sequence holds: clean first, guard it, then point AI at data it can be trusted with.

FAQ

Questions owners ask