Repeated
The work happens often enough that learning and setup could pay back.
AI at Work
Find one good workflow before building an AI empire.
A practical path from “this task is annoying” to a small pilot with clear inputs, human review, useful measurements, and a stop button.
The AI fit test
A task is a better candidate when four things are true.
Inputs exist
The person can provide approved examples, rules, files, or structured information.
Visible output
A human can inspect the result before it causes harm or becomes a record.
Low consequence
An early mistake is correctable, contained, and does not decide someone’s rights, care, money, or access.
The first workflow
Document the work before automating it.
An unclear process plus AI becomes a faster unclear process.
- Map the current task. Trigger, inputs, steps, decision points, output, owner, delays, and common errors.
- Choose the narrow AI job. Draft, summarize, classify, extract, compare, reformat, or suggest. Do not hand it the entire department.
- Set the data boundary. Approved source material, prohibited information, allowed tools, storage, and deletion.
- Place the human checkpoint. Name who reviews, what they verify, and what happens when the output is uncertain.
- Run a small sample. Use representative cases, including awkward examples and known failure cases.
- Measure the result. Time, correction rate, completeness, staff effort, user impact, and whether the work became easier.
- Decide deliberately. Adopt, revise, pause, or reject. “We already paid for it” is not an evaluation metric.
Good first candidates
Useful before impressive.
These are examples, not automatic approvals. Apply your information rules and review process.
Draft from an approved template
Turn structured notes into a first draft while a person remains responsible for accuracy, tone, and release.
Summarize public or approved text
Create a shorter working version with links back to the source and an explicit factual review.
Reformat information
Convert the same approved content into a checklist, table, outline, FAQ, or training handout.
Classify low-risk requests
Suggest categories or routing while a person reviews low-confidence or unusual cases.
Compare documents
Surface differences for a human reviewer rather than declaring which policy or contract is correct.
Generate practice scenarios
Create fictional examples for training, then review them for realism, bias, and accidental sensitive details.
Keep these human-led
Higher consequence needs stronger control.
Some work may use AI support, but it should not become an unsupervised decision lane.
Eligibility and access
Benefits, housing, employment, discipline, credit, insurance, education, and services affecting rights or essential needs.
Health, safety, and crisis response
Diagnosis, treatment, emergency guidance, safeguarding, physical security, or decisions with immediate harm potential.
Money and binding commitments
Payments, contracts, legal positions, bank changes, financial advice, or communications that obligate the organization.
Pilot scorecard
Measure more than speed.
A system that saves six minutes and creates three new errors did not save six minutes.
| Measure | Question |
|---|---|
| Time | Did total task time fall after review and correction? |
| Quality | Did completeness, consistency, or clarity improve? |
| Correction | How much human work was required to make the output usable? |
| Risk | Did the pilot create new exposure, access, bias, or incident paths? |
| Experience | Did staff and service recipients find the process easier or more frustrating? |
| Recovery | Could the team continue when the tool failed or was unavailable? |
Minimum operating rules
Write the rules before the pilot spreads.
- Approved tools and account types
- Allowed and prohibited information
- Required human review by consequence
- Rules for generated facts, citations, and disclosure
- Permission limits for connected systems
- Incident reporting without punishment for early reporting
- Named owner, review date, and stop authority
Next useful step
Turn the workflow into a six-session learning lab.
The Community AI Readiness Lab helps an organization move from shared language to one documented, low-risk workflow with a real decision at the end.