Production AI deployments need four roles. Most pilots have three. The missing one is why they fail.
The three roles every team gets right:
- The model owner — the ML engineer or AI vendor responsible for the technology itself. The prompts, the fine-tuning, the inference latency, the accuracy benchmarks.
- The integration engineer — the developer who wires the model into operational systems. The ERP API, the practice management ODBC connection, the CRM webhook, the field service platform.
- The operations SME — the controller, the operations director, or the field manager who knows what is supposed to happen end-to-end. They sit in the kick-off, sign off on the discovery document, and are referenced when a question arises.
The fourth role is the one that determines whether the deployment ever ships: the translator.
The translator's only job is to convert operational logic into model-grade specifications. A controller knows that receivables invoices over a certain threshold for a specific payor class need a different posting treatment. The translator's job is to capture that knowledge as a five-line spec, exception logic, confidence threshold, and audit log entry — not as a sentence in a Slack message that the integration engineer reads but does not implement.
The translator is not a project manager. The translator is not a business analyst. A project manager moves the work forward. A business analyst documents the existing state. The translator produces something neither does: the artifact that the model can execute against, with every edge case named.
In a recent multi-entity finance transformation involving forty-four entities and roughly 1,843 chart-of-account lines, the translator's deliverable was a thirty-page specification mapping every operational rule to a specific posting workflow. The integration engineer wrote zero code until that document was signed. Once it was, the build took six weeks.
If you cannot point to the translator on your current AI project — by name, with a calendar that protects sixty percent of their week — your deployment is going to take longer than budgeted and produce something that does not match what operations actually needs.
The three other roles can be filled by vendors. The translator usually cannot. They have to come from inside the business, or be embedded by someone who has lived inside this kind of operation before.
Filed under
