Integrating AI with the systems you already run: a UK SME guide to legacy software.

A model that reads an invoice is a demo. A model that reads an invoice, matches it to the purchase order in the accounts system, posts it and flags the exception in the inbox the finance team already watches is a saving. The difference is integration, and it is the part of most AI projects that decides whether they reach production. Gartner’s causes of abandonment include escalating costs and unclear value; in our experience of stalled pilots both usually trace to an integration that was left for later. The UK surveys back the point: most adopters report productivity gains rather than new revenue, and productivity gains come from removing the retyping between systems, not from the model.

A published API is the first choice: reliable, auditable, supported by the vendor. Most modern cloud systems have one; many older ones have a partial one, and the partial one is often enough. Scheduled export and import is the second: the old system exports a file nightly, the automation reads it and writes results back through an import the system already supports. It is slower and it works with almost anything. Email and documents are the third and the most underrated: many legacy systems already send and receive email or accept PDFs, so an AI that reads what the system sends and replies in the format it accepts integrates without touching it. Screen-level automation, driving the system’s own interface, is the last resort: it works, and it breaks whenever the interface changes.

Pick the least fragile route the system supports. Fragility is paid for every month; effort is paid once.

The existing system stays the record of truth. The automation reads from it and writes to it; it does not become a second place where the records live. Firms that break this rule end up with two systems that disagree and a reconciliation job nobody asked for. It also keeps the scope honest: the build is “read what arrives, do the work, write the result back”, not “and while we are at it, replace the CRM”. Replacing the old system is sometimes right, and it is a separate project with its own case, and our guide on off-the-shelf against custom software sets out how to decide.

At fixed-price UK rates a first implementation that is mostly integration runs £4,000 to £15,000, with the running cost of the AI stated per feature before go-live; published market ranges sit around the same band with day-rate suppliers higher. Three questions separate a supplier who will do this well from one who will not. Which of the four routes will you use for each system, and why that one. What happens when the old system is down or changes, and who is paged. And who owns the integration code and the credentials afterwards, which should be you, with a runbook your own developers or your next supplier can follow.

Replacing the old system at the same time. Building a second database because the old one was awkward. Choosing screen-level automation because it demoed fastest. Hard-coding a credential that belonged to a person who then left. Skipping the failure path, so the first outage silently drops a week of work. And not instrumenting cost, so the integration that calls the model on every record runs up a bill nobody predicted. Each of these is visible in the discovery step, which is why we do that step before quoting.