“We should use AI” is not a project brief. A useful brief names the work that keeps repeating, the material needed to do it, the cost of a bad answer and the person who will decide what happens next. That is where practical AI consulting should begin.
Score the workflow before the software
Before choosing ChatGPT, Claude, Gemini, Copilot or another tool, score the work itself. The U.S. Small Business Administration recommends starting small, testing whether a tool adds value and keeping human review around generated work. NIST’s AI Risk Management Framework similarly asks organizations to define intended use, oversight, benefits, costs and how performance will be measured.
A strong first pilot is frequent, bounded and reversible. Payroll, hiring decisions, legal commitments, health information and unattended financial actions are not good places to learn the basics.
1. Inbox triage and follow-up drafts
A crowded inbox creates two different problems: understanding what matters and doing something about it. A narrow assistant can summarize a long thread, identify unanswered questions, group messages by project and prepare a reply in the user’s voice.
A sensible first version
- Reads only selected mailboxes or labels.
- Produces a morning list of messages that need a decision.
- Drafts replies but does not send them.
- Links every summary back to the original thread.
- Escalates uncertain or sensitive messages instead of improvising.
Current Gmail and Outlook products already support forms of summarization and drafting on eligible plans. The consulting work is not merely turning the feature on; it is deciding what the assistant may read, what “important” means for your role, how drafts should sound and which messages always stay human.
2. Meetings into decisions and next steps
Meeting notes are useful only when they become confirmed decisions. An assistant can turn a transcript or note set into a proposed list of owners, dates, open questions and a short status update. The participants then confirm the record before tasks or downstream messages are created.
The system prepares the first pass. The people in the meeting confirm what was actually decided.
3. Repeatable customer questions
Small teams often answer the same questions about shipping, scheduling, products, policies or next steps. AI can prepare answers grounded in approved material and surface the document it used. That can reduce searching without giving a bot permission to invent policy.
Refunds, legal issues, safety questions, angry customers and answers with weak evidence should route to a person. A good pilot tests both the happy path and the escalation path.
4. Routine documents and sales follow-up
Proposals, project updates, research briefs, outreach drafts and FAQ material often reuse the same approved facts in slightly different forms. An assistant can assemble a first draft from those sources, while a person remains responsible for pricing, claims, commitments and who receives the message.
This is also where unsupported promises become dangerous. The Federal Trade Commission has repeatedly made clear that performance and accuracy claims need evidence. If the system cannot trace a statement to approved source material, it should omit the statement or flag it for review.
5. Reporting and exception review
Instead of asking a person to rebuild the same report every week, a system can prepare a consistent view of sales, operations or marketing data, call out changes and list records that need attention. Spreadsheet tools can already assist with formulas, summaries and charts when the data is structured.
The generated interpretation is not the accounting record. The output should preserve the source, calculation logic and date range so the operator can verify what changed.
Run a pilot you can actually judge
- Write the baseline. Record how the work happens now, including elapsed time, handoffs, rework and missed items.
- Choose a small sample. Use representative cases, including a few that should be escalated.
- Define the checkpoint. Name the person who reviews each output and what they are checking.
- Log failure modes. Capture wrong facts, missing context, tone problems and permission surprises.
- Compare the result. Decide whether the pilot made the work faster, clearer or more dependable without creating new risk.
Only then should the workflow expand. The point is not to prove that AI can produce something. The point is to prove that the whole process works better.