Our main business software has no usable API. Is a forward-deployed AI project still possible without replacing it?
Yes, but start with the least fragile connection your existing software supports. Do not make replacing a core system the default price of improving one workflow.
Have the embedded engineer investigate supported exports, imports, scheduled reports, approved database access, and vendor-supported integration options. Confirm permissions and contractual restrictions before building around any of them.
A useful order of preference is:
- Use a supported integration where one exists.
- Use structured file exchange when the workflow can tolerate delayed updates.
- Build a sidecar tool that prepares work for a person to review and enter.
- Consider browser automation only after evaluating its maintenance burden and failure modes.
For example, a Lovable or Replit application could turn an approved export into a work queue without writing back to the original system. That is a proposed design pattern, not a claim that every legacy product supports it.
Keep the existing system authoritative for the records it already owns. Make data freshness visible so employees do not mistake yesterday's export for a live operational view. Decide how failed imports, changed file formats, and missing records will be detected and handled.
This approach preserves EOS as the operational backbone while improving the supporting technology. Capture integration constraints on the Issues List, and give any implementation Rock a clear finish line rather than an open-ended mandate to modernize everything.
My recommendation is to prove one supported data path before committing to the full build. If the only option is brittle screen automation for a critical workflow, reconsider the scope or seek a supported vendor solution.
Category: Forward-Deployed Engineering