We built a prospecting tool for a UK business finance company. It scores about 34,000 CRM accounts against what the client’s past customers had in common, checks callable companies against Companies House, and gives the sales team a short call list with a reason for each one.
Our client’s sales team phones businesses all day, up to 100 calls each, from a Zoho CRM of about 34,000 company accounts. They wanted better targets for the team they already had, and the CRM held no evidence about which companies were most likely to become customers.
Only about 22,000 of those accounts could be called. In a sample of the book, about one in five had no phone number and one in six had no company number.
When we checked the top 2,093 prospects against Companies House, 131 had been dissolved, and 127 of those were still marked Active in the CRM.
We started with a proof of concept on the client’s own data, which became a short analysis for their leadership. We matched thousands of paid-out finance agreements back to the CRM to see which kinds of company had gone from prospect to customer. Across the book, about 2.5% of accounts had.
Read straight from the CRM, companies under three years old seemed to convert at 5.7 times the book average. Many of those records had been created on the day the deal was signed, so they had never really been prospects. Once a conversion only counted if the account existed 90 days before the first deal, the figure fell to 2.0 times. Company age is still in the model, at the lower figure.
Print and packaging companies converted at 2.5 times the book average. Construction was the largest group in the book and one of the weakest, so working through the list in the order it arrived would put most calls in the wrong place.
We tested eleven other candidate signals and dropped them, each with a recorded reason. Use of government-backed lending schemes looked promising, but customers and non-customers used them at almost the same rate (47.1% against 47.5%). A distance signal mostly reflected where the team had already been calling.
The model that remains has 18 signals. Three are measured from the client’s deals: sector, company age and region. Fifteen are flat awards, such as an equipment finance charge on the register or a won customer in the same postcode district. The weights are judgement calls, so they sit in settings where the team can see and change them. Uploading a new deals export re-measures the book, and any figure that moves by more than 25% is held for someone to review before it goes live.
The app sits beside the client’s Zoho CRM. On a schedule, it takes a batch of accounts and:
The client is measuring the list against success criteria we agreed with them before the build started.
The language models do two narrow jobs: researching companies and wording openers. The score itself is plain arithmetic.
The research agent reads public UK web pages only. It records the page and a quote of fifteen words or fewer for each finding, so the team can open the link and check it. Only recent findings with a confidence of 0.6 or more count towards the score.
Every capitalised name in a call opener is checked against the facts held for that company. An opener that names anything else is rewritten once, then dropped, and the company goes on the list without one.
Nothing reaches the CRM until someone on the team approves it, and the app never sends a value it doesn't hold, so it can't blank out a field in Zoho. Every model call is also costed and logged against a daily budget.