Where AI belongs in the workflow.
Most AI workflow ideas should not start with a model. They should start with the work: the handoffs, the failure modes, the evidence, the person who has to own the outcome when something goes sideways.
This is how I make that judgment repeatable: gather real examples, check the risk early, compare AI against simpler fixes, and leave with a next step someone can defend. Not a vibes-based "yes, automate this."
00 — decision map
A fast read on the filter.
Evidence
What does the work look like today? Show examples, tools, owners, volume, rework, rules, and consequences.
Risk check
Can a bad answer hurt someone, create legal trouble, expose regulated data, or move too fast for review?
Real problem
Is the pain coming from the process, the policy, the data, the training, the tools, or the task itself?
AI fit
AI fits best when the task repeats, uses text, has sources to check, and produces work a person can review.
Next move
The answer might be a draft tool, a decision aid, a limited agent, plain automation, cleanup work, or no AI for now.
01 — how I decide
The route comes after the evidence.
Start with the workflow
I want to know who owns the work, what starts it, which tools are involved, where it breaks, and what a better week would look like. If that part is fuzzy, the AI idea is too early.
Look for the boring fix
Sometimes the answer is a clearer handoff, a cleaned-up spreadsheet, a policy decision, or training on a tool the team already has. I want to find that before anyone buys another model subscription.
Check the risk
If the work touches safety, legal exposure, regulated data, irreversible actions, or weak review paths, that comes first. A sponsor being excited does not make the workflow safe.
Then test whether AI helps
AI is useful when the work repeats, uses a lot of text, can be checked by a person, has source material to rely on, and can be measured. Drafting, comparing, sorting, and summarizing are often better fits than final decisions.
02 — proof first
The input I care about.
A useful workflow conversation has proof attached. Interview notes help. Written instructions help more. Real examples are best: the form someone fills out, the email they send, the spreadsheet they fix every Friday. The paperwork is not the point. The point is to stop guesses from quietly turning into software.
- Real examples of what goes in and what comes out
- The tools people use today, and who owns each step
- How often the work happens, how long it takes, and where rework shows up
- The policies, instructions, examples, and rules people already follow
- What happens when the answer is wrong, and whether a person can review first
- The unclear parts: missing facts, disputed rules, and adoption problems
03 — possible routes
A good answer might not be AI.
AI draft or review aid
AI can summarize source material, draft replies, sort work into queues, or compare a case against a checklist while a person stays in charge.
Limited agent assist
An agent may help with several small steps if its permissions are narrow, approvals are built in, logs are kept, and it knows when to stop.
Regular automation
If the work follows clear rules, a script or integration may be better: create records, copy fields, route tasks, & hand work to another system.
Fix the foundation first
Sometimes the useful answer is process cleanup, better data, a policy call, training, more discovery, or no build.
04 — judgment
The part I do not outsource.
Agents can draft packets, summarize notes, write test cases, and poke holes in a recommendation. They do not get to erase uncertainty. They do not get to treat a vendor claim as proof. They do not get to turn a sorting suggestion into approval to act.
That is the work I like: finding the thin line between "this is a great AI pilot" and "please just fix the source of truth." The second answer is less impressive in a demo, but it saves teams from building a smart wrapper around a broken process.
05 — weak fit
When I push back.
The policy is unsettled, but the team wants AI to decide anyway.
The data has no owner, no source of truth, or too many quiet exceptions.
A human could review in theory, but nobody has the time, context, or authority.
The task is structured enough that rules or an integration would be cheaper and safer.