AI Prospect Intelligence for a UK finance team

Who gets the next call

The client is a UK company that finances other businesses. Reps phone companies all day. One person can make 100 calls. The names come from Zoho, where the book runs to about 34,000 company accounts. The ask was better targets for the people already on the team. Nothing in the CRM said which account was more likely to buy.

Only about 22,000 of the records could actually be called. Sampling the book showed how thin the data was: roughly a fifth of the sample had a blank phone field, and roughly a sixth had no company number. Companies House, the UK public register of companies, was then used on the leading 2,093 names. 131 of them had been dissolved. Of those, 127 were still sitting in Zoho as Active.

Before the app, the idea was tried on deals the client already held, and the findings were written up for the directors. Paid-out agreements, thousands of them, were joined back to the CRM. The question was which sorts of company had actually become customers. The rate across the book was about 2.5% of accounts.

A conversion multiple that shrank once the dates were honest

A straight read of Zoho made young companies look exceptional. Under three years old, the conversion multiple was 5.7 against a typical account in the book. Plenty of those records were opened in the CRM on the signing day itself, so they had never waited in the pipeline. The replacement rule ignores a win unless the account's create date is at least 90 days earlier than the first deal. That cut the multiple to 2.0. Age stayed in the model, at 2.0 times the book average.

Print and packaging converted at a multiple of 2.5 on the book average. Construction held more accounts than any other sector, and its conversion was among the worst. Dialling the book in arrival order would have poured most of the day's calls into construction.

Another eleven candidate signals were tried, then removed, and the reason for each removal was written down. Government-backed lending schemes were the tempting one. Customers had used them 47.1% of the time. Non-customers had used them 47.5% of the time. A distance signal mostly pointed at postcodes the team already worked.

Eighteen signals remain. Three of them, sector, age, and region, are calculated from the client's own deals. The rest are flat points, added in full when a condition is met. One condition is a registered charge for equipment finance. Another is an existing win in that same postcode district. Staff chose the weights, and the weights live in settings they can edit. Load a new deals export and the book is measured again. If a figure differs from its previous value by over 25%, it stays unpublished until someone has looked at it.

What goes back into Zoho

A scheduled job pulls the next batch from the CRM that sits beside the app. Records tagged do-not-call come out. So do current customers, accounts a broker owns, and businesses that have closed. Companies House is queried for what is left: whether the company is live, who the directors are, charges registered against it, and the balance sheet from the newest accounts on file.

An agent reads public web pages about the company. Each finding keeps the page, plus a quotation capped at fifteen words. Scores are calculated and the highest 100 are ordered. The reasons that made the score are shown separately. Once a person approves a row, Zoho receives the rank, those reasons, any new contacts, and, if the stored facts support it, an opener two sentences long.

Success criteria were fixed with the client before building started. Measurement against those criteria belongs to the client. No figure from it is reported on this page.

Notes a rep can open and check

Models are confined to company research and to the wording of openers. Arithmetic on the weights produces the score. The agent may use public UK pages and nothing else. A finding changes the score only if it is recent and the confidence on it is 0.6 or higher.

Capitalised names inside an opener are compared with the facts stored for that company. A name that is not among those facts triggers one rewrite. If the rewrite still contains an unsupported name, the opener is discarded. The company remains on the list with no opener attached.

CRM updates wait for a person on the team. The app also refuses to post a value it does not have, so an empty write cannot wipe a Zoho field. Model calls are priced as they happen, and the prices are written to a log that is held against a daily budget.

Australian firms with a book this size

Australian firms that finance other businesses can have the same day in front of them: a CRM full of company rows, and no stored reason to prefer one of them when the calls run out. The part an Australian team can copy is how the score is controlled. Weights are editable, a remeasure that moves by more than a quarter is held, and Zoho is updated only after a person says so. The client on this work is British. The register it queries is Companies House.

Company research sits with AI Agents. The approved write-back into Zoho sits with AI Integration. Putting the ranked list into the sales day sits with AI Enablement.