NotebookLM Audio Podcast — Steering Prompt
Workflow
1. Go to notebooklm.google.com
2. Create new notebook: "Private AI for Mechanical Contractors — Ajay Tyagi"
3. Upload these three source files from the vault:
industrial-contractor-ai-faq.md2026-04-28-your-operations-isnt-slow-ajay.md2026-05-05-private-ai-three-tests-ajay.md
4. Click Audio Overview → Customize
5. Paste the prompt below into the customization field
6. Generate (takes 3-5 minutes)
7. Download the MP3, save to 06-assets/audio/
---
The steering prompt (copy-paste into NotebookLM "Customize" field)
This audio overview is a business podcast aimed at executives inside mechanical, industrial, and manufacturing contracting companies — specifically COOs, CFOs, VPs of Operations, Presidents of Construction, and CIOs at $50M–$1B contractors. Picture your listener as a time-poor operator commuting to a job site or reviewing a board deck on Sunday night. They've been pitched a dozen AI platforms in the last six months and are tired of generic vendor language. Tone: executive, specific, unhyped. Two hosts who are smart and skeptical, not celebratory. No jingles, no fluff, no "in today's fast-paced world" openings. Treat the listener as someone running a real P&L who will tune out the moment you generalize. What to focus on — in roughly this proportion: FIRST 40% — the cognitive paperwork problem in mechanical contracting. Open by unpacking the 7,000-hour AP story (a $400M mechanical contractor whose AP team spent 7,000 hours a year on exception reconciliation, which collapsed to 2,100 hours in twelve weeks after an agentic AI layer absorbed the first-pass work). Use that as a scaffold to walk through the five paperwork streams draining cognitive time at every contractor over $50M: submittals and RFIs, bid-document scope capture, change order drafting, AP exception reconciliation, and safety/QA-QC/close-out documentation. Quantify the total drain at 40,000–60,000 hours per year per contractor. Make it clear this is judgment work being squandered on reading — not a staffing problem. NEXT 25% — what agentic AI actually does to this pipeline. Explain the architectural difference between a chatbot (a tool your team opens) and an agent (something that lives inside the workflows your team already runs — Outlook, Bluebeam, the ERP, Procore). Walk through specific before/after mechanics for submittal triage, bid-document scope capture, AP exception reconciliation, and change order drafting. Emphasize the division of labor: agent absorbs low-judgment high-volume work, human retains judgment, every decision with financial or safety consequence has a human approver by design. NEXT 20% — the three tests every AI vendor must pass. This is the trust and compliance half of the story. Use the $1.4M story (a Fortune 500 industrial firm that spent $1.4M over 18 months with a major AI vendor before compliance discovered prompts were traveling to a third-party data center). Unpack the three tests clearly: (1) Does the prompt leave your VPC at any point during inference? (2) Who controls the model weights — you or the vendor, on whose schedule? (3) Can you reconstruct any output, in under 30 seconds, six months later with the exact model version and retrieval context? Explain why each test specifically matters to mechanical contractors doing federal work, prevailing-wage data handling, or joint-venture engagements. Name that DPAs are contracts, not controls. NEXT 10% — the revenue multiplication math. The reclaimed 2,000 cognitive hours per year compound into $400K–$800K of margin-adjusted revenue at a typical $200M mechanical contractor — not through productivity theatre, but through specific redeployment: 3-7 additional project awards per estimator from bid velocity capacity, 8-12% more service calls per technician per week, $600K-$1.2M in change-order prevention from preconstruction scope review capacity freed up. Be specific with the numbers. The listener's CFO will ask where they come from. FINAL 5% — a clean, direct call to action. Reference the AI Readiness Index by name. The listener should hear this verbatim or very close: "If you want to know where your 2,000 hours are hiding — before your next board meeting — take the AI Readiness Index at wnbpc.app. Thirty questions. Five minutes. You get back a scoped report identifying the two or three places in your operation where a 12-week pilot would pay back fastest, based on your actual systems, your team, and your project mix. No sales call required to get it. Just hand it to your CFO and see if the numbers defend themselves." Critical guardrails: DO NOT use these phrases or any variant: "revolutionize," "unlock value," "game-changer," "leverage" as a verb, "cutting-edge," "enterprise-grade" (except in scare quotes, criticizing vendor-speak), "in today's fast-paced world," "transformation journey," "synergize." If either host drifts toward generic corporate language, pull back hard with a specific number or a specific operational detail. DO NOT generalize. Every claim should be anchored in a specific number, a specific industry role, a specific workflow (submittals, RFIs, AP exceptions, change orders, bid documents, safety, close-out). If one host starts a sentence with "many companies are…" the other should interrupt with "let's be specific — at mechanical contractors over $50M, what we're seeing is…" DO NOT talk in the abstract about "AI transformation." Talk about specific document shapes — a Medius AP exception queue, a Bluebeam submittal markup, a Viewpoint Spectrum change order, a Procore RFI log, a Sage 300 CRE scope item. DO NOT make it a celebration of AI. The skepticism is the credibility. Acknowledge that most AI deployments at industrial contractors have underdelivered. Explain why this architecture (private, in-VPC, agentic, scoped) is structurally different from what the listener has already tried. Host dynamic: one host is the experienced operator (thinks in P&L, asks "what does this cost us not to do?"), the other is the skeptic (thinks in compliance, asks "what happens when the auditor shows up?"). Their back-and-forth should feel like two executives at a trade conference happy hour, not like a marketing conversation. Target length: 12–16 minutes. Long enough for substance. Short enough for a commute. Closing line (the very last thing the listener hears): "What's your patience tax?"
---
After generation — what to do with the audio
1. Download MP3 from NotebookLM → save to 06-assets/audio/2026-05-01-paperwork-to-revenue.mp3
2. Upload to a podcast host (Transistor / Captivate / Spotify for Podcasters / Buzzsprout)
3. Generate episode page with:
- Title: *Private AI for Mechanical Contractors: The 7,000 Hours Hiding in Your Operations*
- Description: pull from the blog intros + three-test framework
- Show notes: link to wnbpc.app (ARI), link to both Medium articles, link to the FAQ
4. Cross-post episode link on Ajay's LinkedIn with a 3-sentence teaser
5. Send to 3 podcast networks (mechanical contractor podcasts, construction-tech podcasts) as "an internal-for-now episode we'd open up if you want to cross-post"
Quality check before distribution
Listen to the generated audio. Kill signals to regenerate:
- ❌ Hosts use any banned phrase
- ❌ The $1.4M story or 7,000-hour story is dropped or paraphrased incorrectly
- ❌ The ARI CTA is not delivered clearly with the wnbpc.app URL spoken
- ❌ One host monologues for >90 seconds without the other interjecting
- ❌ The three tests are not enumerated clearly (prompts/weights/audit-trail)
- ❌ Generic language creeps in ("AI is transforming the industry…")
Pass signals to distribute:
- ✓ At least one host says "let's be specific" or similar mid-episode
- ✓ A specific ERP product is named (Viewpoint Spectrum / Sage 300 CRE / JD Edwards / Medius / Procore / Bluebeam)
- ✓ The revenue multiplication math is delivered with dollar figures
- ✓ Closing line lands: "What's your patience tax?"
- ✓ The ARI URL is spoken at least twice in the final minute