Session 1 · October 6, 2026
From AI conversations to useful work
The short version of what we covered, then the part you asked for: concrete examples, told step by step.
01Three ways to put AI to work
Most confusion about AI comes from mixing these up. The question that separates them is simple.
Who decides what happens next?
Conversation
You work through it with AI. You ask, it answers, you push back.
Decides nextYou
Scripted workflow
AI helps carry out a process you designed. The steps are fixed.
Decides nextThe process
Agent
You hand over a goal and some discretion. It chooses its own next step.
Decides nextAI, within limits
These aren't a ladder, and one isn't better than another. A schedule or a tool connection doesn't make something an agent. Deciding what to look at next does.
One task, three ways: the weekly project update. As a conversation, you paste in approved notes and work on a draft. As a workflow, every Friday the system pulls the updates, drafts them in your format, and sends the draft to you. As an agent, you say “prepare an accurate update for this project,” and it checks approved sources, notices two conflicting dates, and decides where to look next.
02Three questions for any AI setup
Before you connect, paste, or approve anything, ask these about the tool.
- Where does the information go?
- Who can access it?
- What can the system do?
Where it runs can be a public tool, an enterprise service, private infrastructure, or a model on your own hardware. Local isn't automatically secure. Storage, access, and connected tools still matter. And don't paste in anything you wouldn't email to a stranger.
03What to tell AI first
AI doesn't automatically know what your team takes for granted. Tell it what you'd tell a new colleague: the goal, the facts, what good looks like, and the limits. The right information, not all the information.
goalfactswhat good looks likelimits
Before“Write a project update.”
AfterUsing these approved project notes, draft a short client update. Separate completed work, blockers, and decisions needed. Use the latest dated source when information changed. Flag conflicts instead of guessing. Don’t invent commitments or send anything.
04Your first experiment
Pick one task you do every day or every week. Fill in five lines. Rough is fine.
- ProblemWhere does the work get stuck, slow, or depend on someone remembering?
- OwnerWho does this work today, and would check the result?
- ExperimentThe smallest real test. Which examples, which tool you already have?
- BoundaryWhat information can it use? What is it not allowed to do?
- EvidenceHow will you know? Time against today, errors, omissions, corrections. Then revise, stop, or test further.
Try it Wednesday, Oct 7
Review Friday, Oct 9
No company access? Use your own non-sensitive notes. Five examples is a learning exercise, not a validated sample. The worksheet's back side has a worked example for the weekly project update.
Open the worksheet (PDF, 2 pages)
05Specific applications
The feedback I heard was that the examples should be more concrete and unfold step by step. So here they are, each one in six beats, with the lesson last.
Real, details changed happened with a real business; I changed the details. Illustrative composite built to show one mode; not a client story.
Real, details changed
The owner working 80-hour weeks
No AI in this one
Monday morning
The owner of a home-services business starts the week already behind. Quotes are piling up, and the week is heading toward 80 hours.
Where it breaks
Every quote waits on one person. Nothing moves until the owner gets to it, so the pile grows while the hours do.
First attempt
Work more. That's how the week got to 80 hours.
What AI did
Nothing. This one has no AI in it.
What the human did
Built a pipeline, so each quote has a stage and nothing waits on one person to remember it. The owner's week came down to about 30 hours, with more time at home with family.
The lesson
Sometimes the first fix isn't AI.
Bad process plus AI usually gives you faster bad process.
Real, details changed
Crews, job sites, and the per diem question
A custom app, prototyped fast
Monday morning
A construction company has to decide who gets assigned to which job site. Who qualifies for travel or per diem depends on the drive distance and company policy.
Where it breaks
The distance checks live in spreadsheets, done by hand. The lookups hit the API limit and the sheet fills with errors.
| Employee | Site | Before | After |
|---|
| Crew A | Site 14 | #ERROR | Within policy |
| Crew B | Site 14 | #ERROR | Over threshold |
Illustration, not client data.
First attempt
A distance-checking spreadsheet. It's the one that hit the limits.
What AI did
Helped build a custom app quickly, as a working prototype instead of a long requirements document. That let the team react to something real: “is this what you mean?” The app imports employee and job data, calculates driving distances, shows the threshold information, and exports it for the people who decide. Most of it is ordinary software: the distance math and the thresholds. AI helps with one step, cleaning up messy addresses, and a person confirms the result.
What the human did
People still make the assignments. The project manager decides, and an administrator handles any booking. The app doesn't assign anyone, book hotels, or send notifications. Real use found problems a demo wouldn't, like re-imports creating duplicate people, and those got fixed. The person who uses it learned it in about five minutes and taught a colleague in five more.
The lesson
Figure out the decision first. Pick the tool second.
Illustrative composite
Renewal call prep
Mode: Conversation · you hold the pen
Monday morning
An account manager has a renewal call on Thursday. The customer has been quiet lately, and there are two open support issues.
Where it breaks
The notes are scattered across the CRM, old emails, and the last call summary. Prep turns into skimming and hoping.
First attempt
The vague prompt“Help me prepare for my renewal call.”
The answer is fluent and generic: build rapport, highlight value, ask about goals. It says nothing about this customer.
What AI did
The context-rich promptPrepare me for Thursday's renewal call with this customer. Use the approved account notes below: contract terms, the last two support tickets, and the summary of our last call. Give me the three objections I'm most likely to hear, each tied to a specific note, and two questions I should ask. If the notes don't support something, say so instead of guessing. Don't draft anything to send.
This time it returns three objections, each pointing at a line in the notes, plus two questions.
What the human did
Pushed back on the second objection (“that's not what they're worried about, it's the support issue”), checked each claim against the account notes, and decided what to say on the call. Only approved notes went in.
The lesson
The value is in how well you ask and how hard you check.
Illustrative composite
Sales call to proposal
Mode: Scripted workflow · the process holds the pen
Monday morning
A sales rep finishes a run of discovery calls. Each one needs a proposal, and the proposals will wait until the rep finds the time.
Where it breaks
Every proposal is built from scratch, copied from the call notes, a pricing sheet, and the last proposal that looked similar. It's slow, and details get dropped or carried over from the wrong one.
First attempt
Paste a transcript into a chat window and ask for a proposal. It works, but it comes out different each time and depends on the rep remembering to do it.
What AI did
The same fixed steps every time. AI has one job, in one place.
- TriggerA call transcript is saved
- GatherTranscript, pricing sheet, proposal template
- AI stepDraft the proposal sections from the call
- Rep reviewsChecks the draft against the call
- Rep sendsAI never sends
What the human did
Read the draft against what was said on the call, and caught the one planted error in this example: a start date that was never discussed, carried over from the template. The rep fixed it and sent it. Nobody else touches the send step.
The lesson
A workflow is only as good as the process you wrote down.
It fails quietly when the input changes shape, so somebody has to notice. That's what the review step is for.
Illustrative composite
The invoice that came in 12% over the PO
Mode: Agent · AI holds the pen, inside a fence
Monday morning
An invoice arrives 12% over its purchase order. Normally someone in accounts payable opens the PO, the receiving record, and the email thread with the vendor, one at a time.
Where it breaks
What to check next depends on what the last check turned up. A fixed list of steps doesn't fit, and the usual result is a pile of invoices waiting for someone to dig in.
First attempt
A rule that flags any invoice over its PO. It catches this one, and it can't tell you why.
What AI did
The goal is “resolve or escalate.” It reads the PO, the receiving record, and the vendor emails, and finds an approved freight charge that accounts for the difference. It drafts a note: here is what I checked, here is the charge, here is the approval, here is what I couldn't confirm.
AllowedRead approved sources. Draft a note.
Not allowedApprove payment. Change records. Send anything.
What the human did
The AP reviewer reads the note, checks the freight approval, and approves the payment. The agent can't. If it hadn't found an approval, the goal says to escalate to the buyer.
The lesson
The fence is the permissions, not the prompt.
“Please don't approve payment” isn't a boundary. Having no way to approve payment is.