A meeting ends. Someone needs to summarize the discussion, confirm decisions, assign follow-up, and send an update.
It is a familiar process. It also contains several different problems: organizing information, interpreting what people meant, deciding who is responsible, and communicating clearly.
That makes it a useful example of the question behind a first AI project: which part of the work should AI help with, and how will you know whether it helped?
Our recommendation is to start with one recurring process that your team understands well enough to improve and check. Here are five questions to help you choose it.
Does the work happen often enough to matter?
Look for work that repeatedly takes someone away from other responsibilities: preparing meeting follow-up, assembling a recurring report, drafting proposals, or finding answers in internal documents.
Frequency alone does not make something worth automating. Ask what the repetition actually costs: time spent collecting information, delays while someone waits for an answer, or corrections after an incomplete handoff.
“Improve administration” is too broad to investigate. “Prepare the first draft of our weekly project update” gives you a process you can observe. Before choosing a tool, describe where that process begins and what a completed result looks like.
Is the information usable?
Consider what someone needs to do the job correctly. For meeting follow-up, that might include notes, a participant list, and the relevant project context. For a proposal, it might include a clear brief, current service descriptions, and approved examples.
Check whether those sources are current, accessible to the intended users, and appropriate to use in the selected system. Resolve access and confidentiality questions before putting information into a new tool.
If two documents disagree, someone needs to decide which is authoritative. If essential information exists only in one person’s head, the first improvement may be documenting it. Include that preparation in the project; it is part of making the workflow work.
Can a person check the result reliably?
For a first project, favor an output that someone can review before it affects a customer, colleague, or external partner.
A meeting summary is useful only if the reviewer can confirm the decisions and identify invented or missing action items. A proposal draft needs someone who can verify the scope and commitments.
Be specific: who checks the output, what do they compare it against, what would make them reject it, and what happens when information is missing? If checking the result takes longer than doing the original work, reconsider the approach.
- 01Approved notes
Use an agreed source.
- 02AI-assisted draft
Prepare the summary.
- 03Owner review
Check, correct, then share.
Missing information or an unreliable draft? Return to the source before sharing.
Can you test a smaller version?
Imagine a team considering meeting follow-up, proposal preparation, and internal document search. Each is a candidate, not an automatic recommendation. The right choice depends on the team’s actual workload and information.
Start with one document set, one recurring meeting, or one proposal section. Keep a clear path back to the existing process, and let the person responsible decide whether an output is ready to use.
Worked example · hypothetical
Three ideas. Three smaller starting points.
Meeting follow-up
Draft an internal summary from approved notes.
Human check: The meeting owner confirms decisions and action items before sharing.
Proposal preparation
Draft one section using a current brief and approved service information.
Human check: The proposal owner verifies scope, language, and commitments.
Internal document search
Answer questions from a small, maintained document set, showing sources.
Human check: The document owner checks answers against the current source material.
If this team already keeps reliable meeting notes and a meeting owner can review the draft, follow-up would be a reasonable first trial. If its notes are incomplete, that preparation comes first. The choice follows the team’s circumstances.
This principle also appears in Anthropic’s engineering guidance: begin with the simplest solution that works and add complexity when it demonstrably improves the outcome.
What would count as an improvement?
Record the current process before testing the new one. Include time spent gathering inputs, producing the result, reviewing it, and correcting mistakes. Decide what quality the result must meet and who owns that judgment.
For a meeting-follow-up trial, ask: can the meeting owner produce an accurate, approved update with less total effort, including review and corrections? That is more useful than measuring how quickly AI generates a draft.
Choose a condition for stopping or revising the trial. Repeatedly inventing commitments, missing important decisions, or requiring extensive corrections would all deserve attention.
Time released can be valuable, but it does not automatically become cash savings. Be clear about what improved and what your team actually does with the capacity.
Put it to work
Choose one candidate this week
Use these prompts with your team to turn an idea into a test.
- Workflow
- What recurring process are we improving?
- Owner
- Who is responsible for the result?
- Inputs
- Which information and permissions are required?
- Baseline
- What effort, delays, and errors occur today?
- Trial
- What is the smallest useful version?
- Review
- Who checks it, and against what?
- Success
- What improvement would justify continuing?
- Stop condition
- What would make us pause or change course?
A simple text worksheet to copy, fill in, and share with your team.
You may discover that a template, clearer ownership, or a simpler process solves the problem. That is a useful outcome.
At Fog Strategy, we recommend beginning with a specific operational question and testing the answer against real work.
What recurring process would you choose first?
Bring that example to a conversation with us.
Talk through your first workflow Explore our services