Custom AI implementation: from business process to working software
Custom AI implementation means building software around a specific business process and connecting it to the tools people already use. A forward deployed engineer works with the operating team to define the task, implement the software and evaluate it in context. DotGrey does that implementation for the client.
Buy, automate or build?
Choose existing software when it covers the process and integrations you need at an acceptable total cost. Use conventional automation for stable rules and structured inputs. Consider custom AI when variable documents or language require interpretation and the output must enter your own workflow. An agent is useful only when choosing the next action adds value; a fixed sequence is often easier to operate.
Example: preparing a quote from an incoming request
An illustrative implementation reads an incoming request, extracts the required fields and checks them against an approved product catalogue. Fixed rules calculate prices. The software prepares a draft in the existing quoting system and highlights missing information for a person to review. Sending the quote remains a separate approval. This example describes a possible scope, not an existing customer deployment.
What should the pilot deliver?
The deliverable is a usable process with explicit limits. Agree the sample inputs, destination system and expected outcomes before development. Keep representative cases out of development so the final evaluation tests unfamiliar examples. If a required field is missing or a destination is unavailable, the system should report an exception instead of quietly recording an incomplete result.
- Acceptance: field accuracy, human correction rate, completion time and cost per completed task.
- Failure checks: repeated requests, missing documents, access restrictions and unavailable integrations.
- Handover: source code as agreed, deployment instructions, monitoring and an operator responsible for exceptions.
Budget for running the software
Compare implementation fees with model usage, hosting, integration licences, maintenance and human review. Estimate value from the time actually saved after corrections, not the time of the original task alone. For an illustrative task reduced from 12 minutes to 5 minutes including review, the saving is 7 minutes per completed task. Check that improvement against the same quality standard and include exceptions before expanding the rollout.