Build the Workflows
This is the sixth step in The Adoption Path. The policy from Write the AI Policy says what is allowed in general terms; this step turns each approved use into a specific, repeatable procedure — the actual prompt or prompt pattern, the specific review point, and where the output goes next. A policy without workflows leaves every adviser reinventing the same approved use differently each time, which is exactly the inconsistency the policy was supposed to end.
Why it matters
A policy that says client emails may be drafted with an approved tool tells you almost nothing about how to actually do that well. One adviser writes a careful, specific prompt and reads every line before sending; another types a vague request and skims the output. Both are technically following the policy, and the results are nothing alike. Workflows close that gap — they are where the policy actually becomes the way work gets done, rather than a document that sits next to it.
What actually happens
Turn one approved use into one written, findable workflow
Take one approved use from the policy and write down exactly how it works: the input, the exact prompt or prompt template used — with client-identifying information already handled the way Sort Your Data requires, since this step assumes that sorting has already happened and does not redo it — the specific person who reviews the output, and what happens to it after review: filed, sent, discarded, or escalated. Write it down somewhere the whole firm can find it, not in one adviser’s head and not in a chat history the vendor can delete anytime. A workflow that only exists in someone’s memory cannot be reviewed in Review and Keep Records, which makes it functionally the same as no workflow at all.
Build a small library of prompt templates
Build a small library of prompt templates for the firm’s recurring approved uses, rather than starting from a blank prompt each time. A template does more than save time — it is where the data-handling rule from the data matrix gets baked in structurally. A template with a field that reads “client-identifying details: none” is harder to violate by accident than a blank box that quietly invites whatever the adviser happens to type.
Name the review point precisely
Name the review point precisely for each workflow — not “someone should check this,” but who checks it, against what, and what they are actually confirming. A check for factual accuracy, a check that a number matches the source, and a check that the tone is appropriate are three different reviews that happen to look the same from the outside; a workflow that just says “review before sending” leaves the reviewer to guess which one is being asked for.
Decide what happens when a workflow doesn’t fit
Some requests won’t fit any built workflow — output that is clearly wrong, or a use nobody anticipated. Decide what happens then, and make the answer “route it back to a person,” not “let the adviser freelance a fix inside the tool.” A workflow with no answer for its own edge cases quietly pushes every hard case back into the same undocumented, ad hoc territory the whole Path exists to move the firm out of.
Revisit workflows on the same cadence
Revisit workflows on the same cadence as the tools and the policy — the re-check trigger from Vet and Choose Tools and the review cadence from the policy — rather than putting workflows on a separate schedule of their own. A tool’s terms changing, or a new use added to the policy, is exactly the kind of event that should trigger a look at whether the workflows built around it still hold.
Questions to ask yourself
- For each approved use in the policy, is there an actual written workflow, or does each adviser do it a little differently?
- Where does the workflow live, and could a colleague find it without asking you?
- For the review point in each workflow, could you say in one sentence what the reviewer is actually checking for?
- What happens when a request doesn’t fit any built workflow — is there an answer, or does it get improvised?
- When a tool’s terms changed last, did the workflows built around it get revisited?
Common mistakes
- Approving a use in the policy but never writing the actual workflow, so each adviser does it differently.
- A review point so vague nobody knows what they’re actually checking for.
- Storing the only copy of a working prompt in one person’s head.
- Building workflows before the policy exists rather than after.
- Never revisiting a workflow once it’s built.
Which Firm Functions apply
- Prompting and Review Skills: Writing a precise prompt template and naming a specific review point are exactly the skills this Firm Function covers.
- Records and Documentation: A workflow that is not written down and findable is not a record, and cannot be reviewed later.
At a larger firm, workflows are usually built and maintained centrally, sometimes inside the platform itself. The job is to use the workflow as built, not to make an individual version of it — an individual variant is exactly the inconsistency a centralized workflow exists to prevent, even when the individual version feels like an improvement.
Next step
Handle Client-Facing Use: Apply extra scrutiny to any workflow whose output a client can see.