The situation
Most operations teams have one workflow that everyone knows is wasteful. A request arrives by email or form, someone reads it, looks up the customer in the CRM, checks a spreadsheet, updates a ticket and sends a reply. It happens dozens or hundreds of times a week, the inputs are fairly predictable, and the person doing it would rather be doing something else.
AI makes this kind of workflow newly automatable, because the messy part (reading free text, classifying a request, extracting the fields that matter) is now something a model can do reasonably well. The difficulty is everything around the model: permissions across several systems, what happens when a step fails halfway, how to stop a retry from updating a record twice, and who notices when the workflow quietly starts getting things wrong.
This page is for commissioning that workflow as a bounded delivery: one workflow, implemented in your environment, accepted against agreed measures and handed to an owner.
What the work involves
I start with the people who do the work today, not with a model. Together we write the input/output contract: what arrives, in what forms, what the finished output looks like, which systems are read and written, and which decisions need a person. Then I sample real past cases, including the awkward ones, and measure the baseline: volume, handling time, error and rework rate.
The build follows a few rules that come from running AI systems in production:
- Scoped credentials per workflow. The automation gets the minimum access needed for its steps, not a shared admin account.
- Approval before consequential writes. Anything that sends an external message, changes money or changes a customer-facing record can be held for review. The model proposes; the workflow checks; a person approves where agreed.
- Idempotent writes and safe retries. Each case carries a stable key, so a timeout and retry cannot create duplicate updates.
- An exception queue, not silent failure. Cases the workflow is unsure about, or that fail validation, go to a named queue with the reason attached.
- Monitoring the owner can read. Volume, exception rate, approval backlog and failure alerts, in a tool your team already watches.
Where agents need to call several internal tools, I may expose them through a controlled tool layer such as the MCP integration pattern, with each tool scoped and logged.
The signature deliverable
You receive the workflow implementation and handover: working code or configuration in your environment, the input/output contract, the exception queue, monitoring, and a runbook. Illustrative example of an input/output contract extract:
| Field | Source | Rule | On failure |
|---|---|---|---|
| Customer ID | CRM lookup by sender domain | Exactly one match required | Exception queue: “ambiguous customer” |
| Request type | Model classification of email body | One of seven agreed types; confidence threshold agreed with owner | Below threshold: human triage |
| Ticket update | Ticketing API | Idempotency key = message ID | Retry up to agreed limit, then alert |
| Customer reply | Drafted by model | Held for approval for refund and cancellation types | Never sent without approval |
Illustrative example showing the format, not a client deliverable.
How acceptance is judged
Acceptance criteria are agreed in writing after the starting assessment, before the build. Typically: the workflow runs in parallel with the manual process on real cases; outputs are compared field by field; the exception rate and error rate are measured against the baseline; and retry and failure tests show no duplicate writes. The business owner signs off. If the measures are not met, the gap and its cause are written up rather than smoothed over.
Ownership and handover
The workflow is yours from the first commit: your repositories, your accounts, your credentials. At handover, the named owner receives the runbook (how to pause the workflow, clear the exception queue, rotate credentials and change approval rules), and we walk through a live exception together. I do the work personally; any specialist help is disclosed and approved by you first.
Choosing the narrower route
If your workflow is mostly reading documents and preparing reviewed reports, the document processing and reviewed reporting workflows page is more specific. If it happens on calls, see voice AI workflow engineering, which is taken on selectively. If the real blocker is that data is scattered and access is unclear, start with AI data readiness. If you want staff to change how they work with AI tools rather than commission a system, business AI enablement is the better fit.