AI Working Session
Built for Tom Lutz
A two hour working session built around the way AI capability actually compounds in a senior leader's work life. Three rungs. Four themes. One real document of yours that follows you up the ladder.
Each rung asks more of you, and gives more back. The chart below shows how value compounds as you climb. You don't need to be at the top this week. You need to know where you are, and what the next move looks like.
Paste a CIM, get a thesis. Draft an email, sharpen a sticky. One off prompts on one off tasks. Real productivity gains, but the AI forgets you the second you close the tab.
Start hereSet up a Project. Build a Custom GPT. Give it your deals, your voice, your frameworks once, then it remembers. The same prompt does ten times the work.
Climb upAn agent that monitors your inbox for new CIMs, builds the first pass diligence by morning, and surfaces only what needs your judgment. You don't run it. It runs.
See the horizonWe'll run the live blocks together. The solo and optional paths live on this site as your leave behind. Pick them up this week, or hand them to a PL.
Eight principles that stay true whether you're on rung one or rung three. Different formats to keep them sticky. Flip the cards, study the examples, look at the split.
The judgment is yours. The AI drafts the memo, but the conviction in it is yours. You sign your name, not the model's.
Compress, don't replace. The win is collapsing the hours between question and first draft so you spend the saved time on what only you can do.
BCG enterprise accounts only. Spot check facts. Trust frameworks, verify figures. Every time. The model is confident even when it's wrong.
LLMs are nothing like Google. Keywords get you the most probable interpretation. Context gets you the one you actually meant.
Your first prompt is a draft, not a contract. The output is the start of the conversation, not the end. Follow up, refine, push back.
The more structured the ask, the more structured the answer. Specify role, context, task, format. Vague in, vague out.
The model is the easy part. The data is the harder part. The people and process changes are where most value lives, and where most projects fail.
Telling the AI "write in my voice" rarely works. Showing it 3-5 examples of your past writing always does. Recipes beat rules.
One prompt. One output. One task done in a tenth of the time. This is where everyone starts. Before we pick an adventure, ten minutes on the mechanics that separate a good prompt from a great one.
Every good prompt has these four parts. Pick the right tool for the job. And keep one prompt in your back pocket for when you need to kill a page of dense output.
Miss one and the output is mediocre. Use them as section labels in your prompt for the first 10 tries.
All five are useful. Pick by where the work is and what the work needs.
| TOOL | BEST FOR |
|---|---|
| Claude | Writing, reasoning, long docs (1M+ context), careful tone |
| ChatGPT | All-rounder, voice mode, broad plugins, strong research |
| Gemini | Native in Google Workspace, sees Drive/Docs natively |
| Copilot | Native in M365 (Outlook, Word, Excel, Teams) |
| Perplexity | Live web research with citations, fast lookups |
Pick the one closest to what's on your desk this week. Whichever you pick, the workflow shape is the same. Synthesize, structure, draft, compress.
Walk a real diligence chain start to finish on a CIM you have on your desk right now.
Golden record audit, microsegment generator, personalized campaign concept, the conversation with a skeptical COO.
Who to reach out to, why now, what to say. Alumni event prep dossier in 12 minutes instead of two hours.
What's sitting on your desk right now that took you longer than it should have? Bring it. We build the chain for it together.
Five chained moves. Each one produces an artifact you'd actually send.
Your own BCG article as the workflow. The 10/20/70 conversation made concrete.
Sarah, Mike,
Quick pre-read so we use the 30 minutes well tomorrow. Three things we'll cover.
1. Where you actually are on data. You have 4M loyalty members and 2 years of POS, which is a real foundation. Two gaps: order-level item attribution is partial, and you don't yet link app behavior to in-store visits. Closeable in a quarter, but it's the gate to 1:1.
2. Three segments worth activating in Q1.
3. What changes for your teams. The model is 10% of this. The other 90% is your campaign team moving from monthly batch sends to weekly behavior-triggered ones, and your store ops team handling more redemption variability. I'd be lying if I said that was free.
Looking forward to tomorrow.
Most senior leaders' networks decay through neglect, not intent. AI is exceptionally good at the remembering, which is most of the work.
David,
Saw the Lone Star portfolio expansion into industrials, congrats. The Reliance carve-out looks like exactly your kind of play.
I'll be in Houston Nov 12-13 for a board meeting. Free either night? Would love to catch up properly, has been too long since the Carlyle days.
No agenda, just dinner.
The most powerful workflow is the one you'd actually use tomorrow. Tell me what's on your desk, and we build the chain together.
Stop re-introducing yourself to the AI every morning. A Project, Custom GPT, or Skill remembers your deals, your voice, your frameworks. The same prompt does ten times the work. This is where leverage compounds.
A real teammate has four ingredients. Skip any of the four and you're back to a generic chatbot. Build them where your data already lives, not where you wish it lived.
Skip any of the four and you're back to a generic chatbot.
Build where your data and your work already live.
| TOOL | BUILD IT AS |
|---|---|
| Claude | Projects (private workspace with docs) or Skills (modular instructions that travel) |
| ChatGPT | Custom GPTs with files plus actions. Shareable in your org's GPT store. |
| Gemini | Gems that see your Drive, Docs, Sheets, Calendar natively |
| Copilot | Agents in M365: Outlook, Word, Teams. Best for IT-governed orgs. |
Whatever you ran in Tool, you can promote here. The workflow that worked once becomes a Project, GPT, or Skill that runs forever. We set up PE live. The others are templated for you.
A reusable Project that ingests any CIM, runs the 5 step chain, and produces the IC one pager. Built once, used on every deal.
Pre-loaded with 10/20/70, your case examples, and the 2024 BCG piece. For any restaurant or consumer client conversation.
Your alumni roster, dinner notes, career milestones. One Project that becomes your relationship memory.
Whatever chain you built in Rung 1. Turn it into a persistent Project so it works on every next instance, not just today's.
Same 5 step chain you ran in Rung 1, now living in a Project that remembers your approach.
Same idea as the PE Project, ChatGPT flavor. Pre-loaded with 10/20/70, the 2024 BCG piece, and your sanitized restaurant cases.
The one most senior leaders will use weekly. Your alumni roster, dinner notes, career milestones, all in a Project that becomes external relationship memory.
Whatever you built in Rung 1's Bring Your Own. We promote it together. The chain you ran once becomes the Project you'll run forever.
You don't run the AI. It runs. Each sketch below has prescribed steps for you (what to do, click by click, who to call) and a technical appendix for the team that builds it. Start with Alumni. It's buildable in Cowork this month.
Three real sketches and one custom. Each has prescribed click-by-click steps for you, plus a collapsible technical appendix for your engineering team. Start with Alumni. Lower stakes, lower complexity, buildable in Cowork today.
Click-by-click in Cowork. Watches LinkedIn and news for your network. Sunday 6pm: 3 names, 3 reasons, 3 drafted notes in your draft folder. Live in 90 minutes.
New CIM arrives in your inbox, agent ingests overnight, drafts the IC one pager, flags the three judgment calls only you can make. Waiting in your inbox by 7am.
Not for you to build for yourself. For clients to build with BCG's help. The reference architecture for an autonomous personalization engine.
What do you do every week that follows the same shape? That's your candidate for agentic automation. We sketch it together.
Buildable in Cowork in 90 minutes. Click-by-click below. This is the one you go home and do tonight.
90 minutes total. Sunday after dinner.
name, company, role, linkedin_url, last_touch_date, notes. 50-150 names is plenty. 10 minutes to build the list if you don't have it.Sundays at 6:00 PM Central. Recipient: your email. Save. 1 minute.Tom,
Three names from your alumni list worth reaching out to this week. All drafted, all in your "reconnect-drafts" folder, ready to review.
1. Priya Menon (Trilogy Brands) · 14 months since last touch. Just announced she's leading their AI-driven category management pilot. Good reason to congratulate and mention BCG retail AI work. Draft tone: warm, brief, soft mention of a 30-min call.
2. Mark Yoshida (Apollo) · 8 months since last touch. Apollo announced a $400M add-on to one of their portfolio QSRs last week. Mark led it. Draft tone: congratulate, soft float on restaurant personalization work.
3. Karen Whitman (Insight Health) · 22 months since last touch, way overdue. Company hit $1B valuation in their Q3 release. Draft tone: personal only, "thinking of you, no agenda."
Hit review-and-send in 15 minutes or skip any that don't feel right.
0 18 * * 0 in Central time. No separate scheduler infrastructure.claude-opus-4-7 for ranking who matters most (judgment-heavy). claude-sonnet-4-6 for the bulk drafting. Cowork picks defaults; override in advanced settings.You don't build this in Cowork. You commission it. This page is what to ask for, who to call, and what to demand before you give the green light.
Five moves. None involve writing code. All take real time.
Tom,
340-unit casual dining CIM arrived from Apollo at 6:47 PM. I've worked through it overnight. Spend 20 minutes on this email, not 4 hours on the CIM.
1-page summary attached. Below: the three things you specifically need to look at before noon.
Top 10 diligence questions ranked by deal-killing potential are in the attached one-pager. Highest-priority three line up to the three flags above.
One question I couldn't answer from the materials: What's the unit-level break-even for the new prototype stores? Flagged for management interview.
microsoft.com/graph/api/mail-rules.POST /v1/projects/{id}/files.claude-opus-4-7 for reasoning depth. Total chain: 5-8 API calls.Not for you to build for yourself. For you to sell. The reference architecture for an autonomous personalization engine inside a restaurant or consumer client.
Four moves to make this a sellable BCG offering, not just a vision slide.
The test: what do you do every week or every month that follows the same shape, with the same inputs, producing the same kind of output? That's your agentic candidate.
Adoption beats perfection. The leaders who get fluent at this are the ones who pick two specific things from the list below and actually do them this week. Tap a card to mark it done.
Take an actual CIM, proposal, or memo on your desk this week. Run the Rung 1 chain end to end. Time yourself.
Run the 4-step alumni workflow on your real network this Sunday. Identify 12, draft 4 outreach notes, send at least 2.
Pick a recent recommendation you made. Ask Claude to argue the opposite case as hard as it can. Notice what surfaces.
15 minutes. Create the Project, paste the system prompt, upload 2 sanitized past one-pagers as knowledge. Save 5 starter prompts.
Highest weekly-use case for you. Set it up Sunday, use it every week. The setup template is in Rung 2.
Pull the PE Diligence SKILL.md from this site as a template. Adapt it. 20 minutes. Now you have a portable skill that travels with you to any Claude account.
Send one prompt from this session to a PL on your team with two lines on how to use it. Adoption beats perfection.
Show your team one workflow live. Not slides about AI. The thing running on your screen. 30 minutes.
If the PE one-pager output works, propose it as the standard format for the practice. The Project becomes infrastructure.
When something does not work this week, send 3 lines: what you tried, what failed, what you expected. We will diagnose together.
Friday afternoon, 10 minutes. How many times did you use Claude or ChatGPT? On what? What was the highest-leverage use?
Most realistic Rung 3 move. 90 minutes one Sunday. Click-by-click instructions are in the Operating Layer page.
What's the one task you'd most want to not do tomorrow morning, that an agent could do for you?
Which workflow today, if your whole practice ran it, would change how BCG shows up?
When does this stop feeling like a tool and start feeling like infrastructure?
38 prompts, organized by the work you actually do. None of these were in the session above. Bookmark this page. Use the filters to find what you need. Copy, paste, tweak, run.