Modernising a Working Product with AI

A client came to us with a product that wasn’t broken — a B2B billing and market-analytics platform in the healthcare space, running reliably for years. Quarters closed on time. The numbers were right. Nobody was asking for a rewrite.

The cost had simply moved. It had migrated out of the software and into the people operating it.

Purchase data arrived from suppliers as spreadsheets attached to emails. Someone downloaded each one, worked out which supplier and quarter it belonged to, uploaded it, then scanned the rows for anomalies — a price that jumped, a duplicated line, a customer not yet in the system.

Commercial terms — tiered discounts, rebates, free-goods offers — formed a matrix that varied by supplier and by product, existed in two parallel versions, and changed every quarter. Most terms carried over; some didn’t; a handful of customers had negotiated exceptions. Every quarter, that matrix was largely re-entered by hand.

And answers lived behind dashboards. “Why is this customer’s invoice higher than last quarter?” was answerable — after ten minutes of filtering, drilling, and exporting.

One root cause: the system stored the data faithfully, then left all the reasoning to the operator.

1. Identify the manually complex tasks: Not the technically interesting ones. We look for work that is repetitive, judgement-heavy, and low in creativity — where a competent person follows a known procedure and the only scarce resource is attention. That is where AI pays back fastest.

2. Wrap the existing APIs in an MCP server: This is the step that keeps the project small. A mature product already has endpoints for everything that matters: fetch transactions, create a customer, apply a discount tier, generate a report. The Model Context Protocol turns those endpoints into tools an agent can call. No rewrite, no data migration — existing authentication, permissions and validation stay exactly where they are. The agent is simply another authorised client.

3. Design an intuitive UI on top of the existing one: Same layout, same palette, same components. New capability arrives as additional surfaces inside the familiar product: an inbox where incoming data lands already triaged, a chat panel on every screen, a plain-English summary of what changed in this quarter’s terms. Familiarity is the adoption strategy — users should recognise the product on sight and find it now does more.

Forward the supplier’s email, and the data arrives parsed, with three items flagged for review, each cleared in a click. Type “apply the new quarter’s terms to all customers” and see a preview of exactly what will change — negotiated exceptions preserved — before confirming. Ask a question in plain English and get the number, the chart, and the source.

One guardrail runs through all of it: the agent never writes without a human click.

It reuses years of hard-won business logic instead of discarding it. It ships in days rather than quarters. And it can be proven as a clickable prototype, in front of real users, before a line of production integration is written.

That is what we mean by modernising a product with AI, not replacing what works — teaching it to do the work.

Mobifilia helps product teams add an AI layer to software they already trust. If your product works but your people are still doing the work, we should talk.