Legacy integrations

“Integrate with the ERP” is the phrase that inflates budgets. In an SME the legacy is often a 2009 programme, a spreadsheet that is the real system, and an ERP vendor who takes three weeks to return a call.

AI does not need, in month one, to write inside the legacy. It needs to read a file or leave a draft in a queue. The rest is architecture vanity.

This is the order we use so the pilot does not become a twelve-month integration project.

The order: file, API, queue, screen

• File. CSV, XML, exportable Excel. Ugly and robust. If it exists, it is the path. Version the format.
• API. If the vendor has a documented one and the plan includes it, better. Ask for a test environment. Do not experiment in production on a Friday.
• Queue. A mailbox, a folder, a board. The model leaves a draft; a human posts it to the legacy. That is an integration. Stop apologising.
• Screen (RPA). Last resort: no export, no API, volume justifies the upkeep. Budget for it breaking on every update.

What not to do in month one

Do not ask for “real-time bidirectional write” if nobody has measured the flow. Do not stand up a bus, an iPaaS and an enterprise staging environment for four documents a day. Do not let the ERP vendor sell you their AI module just so you never talk about a CSV.

Also do not “connect the whole Drive so the model finds the data”. That is not integration. That is a disguised inventory and a GDPR risk.

A contract with the legacy owner

Before you quote, a call with whoever maintains the ERP: does it export? is there an API? what does a new field cost? who issues the service user? If that call does not exist, the integration budget is fiction.

Write down which system is the source of truth. Two writers with no owner produce ghost stock and duplicate invoices. AI must not become a third source of truth.

Draft design

The model (or extractor) produces a record with named fields and a “pending” state. A human confirms. Only then is the object created in the legacy. Keep the original (PDF, email) next to the id. When there is an error, you can see why.

If the legacy has no draft state, the queue lives outside: a sheet or a board. Forcing the final object on day one is how you dirty the database.

Cost, risk, programme

Integration is the line that slips most. That is why it sits apart in the budget, with an assumption (“daily CSV; if we must open an API, we re-estimate”). Kit Digital can cover process management; it does not cover an ERP rewrite. Do not promise that in the memo.

Questions for the ERP maintainer (sample mail)

“We need to export [fields] from [documents] every [day/hour], or a read-only user for [API]. Does it exist? Cost and lead time to switch it on? Is there a test environment? Who creates the service user? We will not write to production in month one.”

If the answer takes three weeks, the plan B (queue + sheet) starts now, not after three weeks of polite waiting. The pilot cannot be held hostage by the vendor.

Keep that answer: it is evidence of why the deliverable lands in a queue and not “in the ERP”. A reviewer understands a vendor no. They do not understand a ghost integration.

Field mapping, without “ontology” theatre

A two-column table: our field → legacy field (or “does not exist”). Three example rows with real values. If a field does not exist, it stays in the queue; it is not invented. That table is the integration contract. The box diagram is not.

Version the table (date). When the vendor changes a name, you will see it. When a salesperson promises “smart mapping”, ask them to fill the same table. It usually stays half done.

Do not accept a shared “integration” user with admin rights. A service user, least privilege, password in the manager, an alert if it is used off-hours. The legacy is already fragile; do not add a master key in a chat.

Frequently asked questions

Will the ERP vendor force us onto their AI?

They may push. Ask price, residency and whether you can leave. Sometimes their module is fine. Sometimes it is a lock. The CSV remains an adult plan B.

What if there is not even a CSV?

A human queue plus a photo or PDF. You extract to a draft. Typing into the legacy remains, but less. That is already a gain. RPA is a later discussion.

How long is a “simple” API?

The simple API is the one already documented with a test user. The other one is not simple. Budget the vendor call as a risk, not a footnote.

Next: automating operations and assisted billing. Tell us the system and whether it exports.