Decide what one CRM account represents
Ask sales and finance to finish the sentence “These two records refer to the same…” Sales might mean the same shop, while finance means the company that owes the money. Two branches may share a billing company and still need separate delivery instructions. A franchise can share a brand with a neighbouring shop while paying its own bills.
Give each relationship its own place in the CRM. A site can have a delivery address, a trading name and a link to a billing company. Put the date and evidence on that link. If a shop changes hands, staff can record who placed the new order without silently moving last year’s debt to the new owner.
Use a name search to find candidates
Keep the spelling from the enquiry. Search a second, tidied version if punctuation, spacing or an abbreviation might hide a match. Companies House provides a company search that takes a search term and a separate profile lookup that takes a company number. A practical workflow uses the first to find possibilities and the second to inspect a particular record.
Where the buyer has supplied a company number on a purchase order, copy it as text. Preserve any leading zeroes or letters. Fetch that profile and compare the registered name with the order. Also check what the number describes. A group website might give the parent company’s details while the order names a subsidiary.
Give the assistant a field for the source of each detail and an option to return no supported match. A trading name alone may produce several results or none. If a search fails because the service is unavailable, record the failed lookup separately so staff can retry it. An empty shortlist and an unfinished search need different follow-up.
Illustrative example — two workshops at one yard
Suppose a supplier receives a purchase order headed “Yard Workshop”. The CRM contains a repairs account and an equipment rental account. Both use the same industrial yard, reception number and email domain. The names and circumstances here are fictional.
A name search produces two plausible companies. The assistant can show the accounts clerk the conflicting evidence in three lines. That saves reading time while leaving the choice of billing account open.
| Source | Evidence | Decision |
|---|---|---|
| Delivery note | The same yard and reception number appear on both accounts. | Keep both candidates. |
| Group website | The footer identifies the group company. The site advertises both repair and rental work. | Keep the group relationship separate. |
| Purchase order | The buyer block names the rental company and gives a number matching its Companies House profile. | Link this order to the rental company. Keep the repairs history on its own account. |
Count independent clues and show conflicts
A postcode and the town derived from it tell you much the same thing. Likewise, a website footer copied into three directories still has one source. A long list of agreements can make a weak match look convincing. Show where each clue came from and let the reviewer see which clues share an origin.
Splink, an open-source tool for probabilistic record linkage, compares records across fields and cautions about using highly correlated columns. For a small CRM, start with a shortlist and side-by-side evidence. More elaborate matching earns its keep when the volume of unresolved pairs justifies the work of tuning and checking it.
A useful review screen includes the strongest competing candidate and a specific question. In the workshop example, ask for the company number of the buyer on this order. A generic request to confirm the address would send staff back to the clue both companies already share.
Test wrong links and missed links separately
Take a sample of recent orders. Have a colleague establish the billing company from the transaction paperwork before showing them the AI result. Include same-name firms, branches, former customers and cases where the paperwork leaves the answer open. Keep those open cases visible when reporting how much work the system can handle.
For an illustrative trial of 100 orders, suppose the rule automatically links 70. A reviewer finds two wrong links among those 70. The observed accuracy of automatic links is 68 out of 70, about 97%. The other 30 orders still need someone’s attention. Inspect those too. A confirmed company missing from the shortlist calls for a better search, while two plausible candidates may call for a better question on the order form.
Start the pilot with suggestions only. Ask an accounts clerk to record the evidence that settled each case and the time spent correcting a bad suggestion. At the next review, choose one narrow condition for automatic linking, such as a verified buyer number on an order with consistent company details. Leave conflicts in the review queue and check a sample of accepted links after each change.