Case studies

The problem

Customer emails and cases landed in shared queues, and staff read every one to decide where it should go. An AI triage project meant to fix this had stalled at 53% accuracy. Nobody trusted it enough to switch it on.

What we built

We rebuilt the triage. Rules handle the clear-cut cases, AI handles the grey areas, and anything uncertain goes to a person.

The problem

Purchase VAT reporting is a key pillar of tax compliance. But the finance system report behind it came out wrong every time it was downloaded: duplicates, misaligned rows, mixed old and new tax codes. Two analysts fixed and reformatted it by hand, then reconciled the VAT figures to the general ledger, using judgement that lived only in their heads.

It was slow, so it only ran weekly. Each missed run was a chance for an error to slip through. That could mean tax overpaid or underpaid, or an audit by the tax authority with interest and penalties.

What we built

We wrote that judgement down as ordered tax rules and automated the whole pipeline to run every working day, with an audit trail. Each transaction is labelled include, exclude or human review, along with the rule that decided it. Nothing is released until the VAT totals reconcile to the general ledger.

The problem

Agents warned business customers about unexpected usage spikes by hand, reading each customer's wishes from a free-text note. Written that way, it couldn't be automated or scaled, and every customer got the same blunt service. Missed warnings ended in goodwill credits.

What we did

We redesigned the process and replaced the free-text note with structured preference fields. With those in place the alerts could be automated, the service was easy to manage, and the business could offer tiered and premium alerting.

The problem

New bill-pay customers had to upload identity and address documents before their order could go live. Many stopped at that step and never came back. Each dropped order was a contract, and its monthly revenue, lost after the sale was already made.

What we built

We built an automation that finds orders stuck at document checks and reminds customers to finish.

The problem

Supplier invoices arrived by email as PDFs in dozens of layouts. The accounts payable team keyed each one into the finance system by hand, then chased mismatches against purchase orders. It was slow, and every typo turned into a payment problem later.

What we built

Invoices are read automatically, matched to the purchase order, and posted. Anything the reader is unsure of goes to a person.