Most CFOs I have talked to who say “AI didn’t stick” made the same mistake. They tried to bolt LLMs onto their existing workflow instead of restructuring the workflow around what LLMs are good at.

The pattern below is the one that has actually held for me across portfolio companies. It is a five-day cadence with a distinct mode each day, a distinct AI use each day, and a distinct output each day.

The weekly cadence at a glance

Monday: Weekly flash review

Monday morning is not the time to think creatively. It is the time to catch up on what actually happened last week and get in front of anything that is off. The AI job is fast synthesis from a set of inputs you drop in.

Attached: prior week's revenue file, cash file, and pipeline export.

Write a 1-page Monday flash for the leadership team:
1. Revenue last week vs weekly plan (single number).
2. Cash on Friday close vs Monday-prior.
3. Pipeline movement: new opps added, stage progressions, losses.
4. Two things I need to look at this week that were not obvious last week.

Rules: 200 words max, no adjectives, name the driver on any variance.

By 9am you have a flash, and by 9:30 you have circulated it. The rest of the day is meetings and follow-ups triggered by whatever the flash surfaced.

Tuesday: Prep for the leadership meeting, reason through open decisions

Tuesday is the day I use AI most for actual thinking. Every week I have 2-3 open decisions that are not yet resolved. Renewals, hires, vendor consolidations, capex asks. I write a short decision memo prompt for each.

Decision to make: {ONE-LINE DECISION}
Context: {2-3 SENTENCES ON WHAT IS TRUE TODAY}
Options I am considering: {A, B, C}
What I know:
- {fact 1}
- {fact 2}
- {fact 3}
What I do not know:
- {open question 1}
- {open question 2}

Write:
1. A one-paragraph steel-man for each option (best case for that choice).
2. The two questions I need to answer before I can pick.
3. A recommendation with the specific tradeoff you are choosing.

Be direct. Push back if my framing is off.

Push-back is the key phrase. Without it, the model will validate whatever you propose. With it, you get a real second opinion.

Wednesday: Mid-week P&L pulse

By Wednesday you have enough data to know whether the month is going to make plan. This is the “no surprises” check.

Attached: month-to-date revenue, month-to-date opex, month-to-date cash.

For the current month with {N} business days remaining:
1. Project full-month revenue at the current run rate.
2. Project full-month EBITDA using MTD opex trend plus my known remaining commits.
3. Flag any line item where MTD spend is more than 10% ahead of a straight-line pace of monthly plan.
4. Give me the top 2 course corrections I could still make with {N} days left.

Do not smooth the projection. Show me actual current pace.

Two lines of that output usually trigger a Wednesday afternoon Slack: “hey, we need to slow down the {X} spend for the next 8 days.”

Thursday: Deep work

Thursday is protected for one long thing. A proposal, a spec, a strategic memo, a JD for a new role, a competitor teardown. Whatever the one thing is, it gets 3-4 uninterrupted hours.

AI is the drafting partner, not the finished writer. My pattern:

1. Ten minutes on a whiteboard (paper or Miro). Write the outline by hand.
2. Feed the outline into Claude with a “write section 1 in {my voice}, focused on {audience}” prompt.
3. Edit heavily.
4. Move to section 2. Repeat.

Thursday is when the AI produces the most words but is doing the least of the actual thinking. That is the correct division of labor.

Friday: Weekly wrap and next-week ask

Friday afternoon is when you close the loop with your sponsor and set the agenda for next week. This is the second most-valuable AI use of the week after Tuesday decision memos.

Attached: this week's Monday flash, my Tuesday decision memos, Wednesday pulse, Thursday writeup.

Write a Friday email to my sponsor / CEO with:
1. The two most important things that happened this week (specific, with numbers).
2. The one thing that is not going to plan and what I am doing.
3. What I need from them next week (a decision, an intro, or "nothing").

Rules:
- Under 150 words.
- No "reach out with questions."
- No "have a great weekend."
- If nothing needs a response, say "no response needed" at the top.

The “no response needed” line at the top is what separates senior-CFO email from junior-CFO email. Sponsors get 40 updates a week. Telling them upfront that they do not need to reply is a courtesy.

The 30-minute Monday setup that saves 3 hours by Friday

The single highest-ROI thing you can do is spend 30 minutes on Monday setting up the week’s Claude Project or Custom GPT with the current week’s files pre-loaded.

1. New Project (or clean up your standing one).
2. Upload: last month’s board deck, the current 13-week cash forecast, the current forecast file, the last 4 weekly flashes.
3. Paste in your system prompt: “You are helping me run finance for a PE-backed {industry} company at {revenue{‘}’} run rate. My style is direct. Push back. Cite the file when you make a claim.”
4. Test with one throwaway question. Confirm it is pulling from the files.

Every subsequent conversation that week is faster because the context is already there. By Friday you have avoided re-explaining your business three or four times.

What doesn’t belong in the cadence

• Real-time trading decisions. AI is not a copilot for time-sensitive treasury moves.
• Personnel decisions where you feel the answer already. The model will find rationale for whatever you propose. Trust your read.
• Auditor-facing memos. Draft outside the LLM for anything that will be in a workpaper.

FAQ

Is this actually the right cadence or just what happens to work for you? Some of it is my rhythm. But the pattern of review-prep-pulse-deep-wrap is not new. What is new is which AI use fits each mode.

What if I have a board meeting on Wednesday? Slide the deep work to Tuesday, move prep to Monday, do the wrap Thursday. Keep the modes in the same order.

Should I use the same tool all week? If you can. Context that persists across the week is what makes this fast. See Custom GPT vs Claude Project vs Copilot Agent for tradeoffs.

What about model choice? Any of the frontier models handle this workload. See Claude vs ChatGPT vs Copilot.

How long until this feels natural? About 3 weeks. Week 1 is clunky. Week 2 you skip a step and notice. Week 3 you rely on it.

Related reading

• The 5-Prompt Weekly Financial Review for CFOs
• The Prompt Library Every CFO Should Steal
• Build Your Board Deck with AI
• The First 90 Days as a Portfolio Company CFO
• Restaurant Bottom Line for operators running weekly cadence in hospitality

Sources

• Anthropic, Claude documentation, docs.anthropic.com
• OpenAI, ChatGPT productivity guides, help.openai.com
• AFP, 2026 CFO Time Allocation Survey, afponline.org

Written by The Pragmatic CFO. 15+ years running P&Ls and building AI-native finance workflows across portfolio companies.