Board Room
Selected work / 05 · In-house product

Eight advisors argue your hardest call. You make it.

A founder brings the decision they keep putting off. Eight AI advisors argue it over three rounds, each from a different angle. What comes back is scored advice: what to do, how sure the board is, who disagreed and why, and what to do this week.

Built and run by Indica TechnologyRuns on one machine£0 per meetingA full meeting in about four and a half minutes
The questionasked 28 Oct

A UK bank pilot starts next month. The demo has no audit log and no fallback. Start on the date, or ask for six weeks?

Lean: start on the date

First take · 8 of 8writing alone
JBPGAHGVDMGWWBPT
A start on the date2–3
B ask for the delay9
Go with conditions · B8.6 / 10

Ask the bank for a six-week delay to rebuild before go-live, on four written conditions.

Dissent: Graham seatSupport 8.4 / 10
InYour decisionThe options, your numbers, your limits, and which way you lean.
Board RoomEight advisors, three roundsThey write alone, argue it out, then vote on a plan.
OutScored adviceA pick, a score for every option, the dissent, the conditions and the next steps.
01 / What goes in

One decision, in plain words.

The question, exactly as it was typed

I run a six-person AI startup. Our support-ticket triage demo won a paid 90-day pilot with a UK bank, worth £60,000, and the bank wants to start next month. The demo runs on one hosted model with no audit log, no fallback if the model is down, and no tests of answer quality. We have £180,000 in the bank and seven months of runway, and two other banks are waiting to hear how this pilot goes. Options: (A) start on the bank's date with the demo as it is and harden it during the 90 days, (B) ask the bank to push the start back six weeks while we rebuild it for production. I lean towards A because I'm afraid that if we ask for a delay, the bank will cool off and the other two banks will hear about it.

A sample decision we wrote to test the board. No real company is behind it.

What the board needs

  • The decision, in one line.
  • Your options: two to four. "Keep doing what I do now" counts.
  • Your numbers: revenue, cash and team size.
  • Your limits: deadline, hours and budget.
  • What success looks like, and by when.
  • Which way you lean, and why. The board tests your lean instead of agreeing with it.

Two ways to meet

Full: all eight advisors and three rounds: a first take each, a debate on the points that split them, then a vote on a written plan.

Quick: four advisors picked for the decision. The vote round runs only if they still disagree, or one of them sees a risk of ruin.

02 / Who argues it

Eight angles on the same decision.

  • Strategy and decisionsIs this the right bet, and how reversible is it?Built from the public work of Jeff Bezos
  • Product and startupsIs this something people actually want, and is it growing?Built from the public work of Paul Graham
  • Offers, pricing, and salesWill people pay, how much, and how do we close?Built from the public work of Alex Hormozi
  • Marketing and attentionWhere does the attention come from, and is it underpriced?Built from the public work of Gary Vaynerchuk
  • People, hiring, and founder timeWho does this, and what does it cost the founder's calendar?Built from the public work of Dan Martell
  • Operations and scalingCan the business run this without the founder holding it together?Built from the public work of Gino Wickman
  • Money and capitalWhat does it do to cash, risk, and long-term value?Built from the public work of Warren Buffett
  • Competition and differentiationCan you own this market, or are you walking into a fight?Built from the public work of Peter Thiel

Each advisor scores every option from 0 to 10 from its own angle. The advisors whose area the decision touches most count for more in the final score.

03 / What comes out

Scored advice you can act on this week.

The board's answer to the question above, from a full meeting of all eight advisors that took 4 minutes 8 seconds on one machine. The founder leaned towards A. The board scored A 2.3 out of 10 and recommended B. It is shown as the board wrote it, with formatting marks and source references left out.

GO WITH CONDITIONSOption B: Ask for six-week delay to rebuild for production before go-live

Ask the bank for a six-week delay to rebuild production-grade before go-live, but only if internal QA hits 99.9% pass rate by week 4 and the bank signs a written addendum confirming acceptance.

  1. AStart on bank's date with demo as is, harden during pilot2.3
  2. BAsk for six-week delay to rebuild for production before go-live8.6
  3. CDon't do it2.4
  4. DKeep doing what you do now, ignore the pilot0.9
Confidence
9/10
Support for the plan
8.4/10
Votes
6 endorse, 2 endorse with an amendment
Spread on the pick
1 point

Why the board landed here

  • A failed pilot with missing audit logs, fallback, and QA is a one-way door, reputation loss with this bank and future banks is irreversible because enterprise AI reliability is table stakes, not differentiating. Wickman seat
  • The £60k pilot is purely a reference; its standalone value is negligible versus the £180k+ follow-on pipeline, so losing it destroys the whole business model. Buffett seat
  • A six-week delay can be framed as competence if led with the bank’s risk (e.g., audit integrity), not your own unpreparedness, and only if backed by internal QA pass rates ≥99.9% and a camcorder-style plan showing ownership. Thiel seat
  • Founder time is the scarcest capital, shipping a shaky demo guarantees fire drills, not learning; rebuilding now buys back 6x founder hours at £62.50/hour versus £150+ later. Martell seat
  • In enterprise AI, no one checks output, so if the first demo fails audit or fallback, the bank assumes all output is garbage, and judgment errors are costly and non-retryable. Thiel seat

The minority report

“A demo that breaks under audit isn’t just bad—it’s zero users. You have to make something people want.”

Graham seat, outvoted

For this view to be right: That the bank will tolerate a broken demo if fixes are fast and visible, this is false in enterprise AI where audit/fallback are table stakes, not features.

Conditions

  1. Bank signs a written addendum accepting a six-week delay only if production-grade audit logs, fallback, and QA pass rates are delivered before go-live, and the £60k fee is not voided or penalized.
  2. Internal QA passes at 99.9% on answer quality, audit, and fallback recovery by week 4, tested against real bank data (not synthetic).
  3. Founder records and shares a camcorder-style plan (per Martell) with one number (99.9% QA pass rate), outcome, and handoff to lead engineer, with lead engineer present, before calling the bank.
  4. If internal QA pass rate <99% by week 3, pause the ask and rebuild or walk.
  5. Founder confirms no penalty clauses bind the startup for delay.
  6. If bank refuses written confirmation before week 1 ends, walk away.

What to do next

  • Day 1 Founder and Integrator (or lead engineer) sign off on rebuild plan with weekly QA scorecard (e.g., QA pass rate) and camcorder-style plan recorded with lead engineer present.
  • Day 2–3 Ship rebuilt system to internal QA suite; run full test suite against real bank data; capture pass rates; if <95%, pivot to Option C or reshape.
  • If ≥95% by week 2, Day 3 Call bank’s technical lead and project lead separately with same ask framed as 'quality assurance', lead with bank’s audit risk, not your own.
  • If bank confirms in writing, Day 4 Lock addendum with 99.9% QA threshold and production-grade requirement.
  • Week 4 Rebuild complete; QA sign-off recorded; lead engineer certifies system readiness.
Cheapest test
Run a full internal QA pass on audit/fallback/answer quality by week 2; if pass rate <95%, pivot to Option C.
Kill criteria
If internal QA pass rate <95% by week 2, or bank refuses written confirmation before week 1 ends, pivot to Option C.
Revisit
Revisit if bank delays beyond six weeks or insists on 'hardening during pilot', then pivot to Option A only if QA pass rate ≥99.9% by week 2 and bank signs addendum waiving penalties.

Watch this meeting happen, step by step: the first take, the debate and the vote

And a quick meeting: four advisors, 1 minute 38 seconds

Asked Should we move our customer-support AI agent from a paid hosted model to an open-weights model on our own GPUs? We answer 40,000 tickets a month. The model bill is £9,000 a month and growing about 15% a month. We are two engineers and neither of us has run GPUs in production. A GPU server would cost about £2,500 a month. Options: (A) move now, (B) keep the hosted model but cut the bill with caching and a cheaper model for easy tickets, (C) run both side by side for 60 days and decide on the numbers. I lean towards A because the bill scares me.

GO WITH CONDITIONSOption C: Run both side by side for 60 days

Run both systems side by side for 60 days (Option C), but only if you name an owner by Day 3 and contract a third-party GPU ops partner before Day 1 begins.

  1. AMove to open-weights on own GPUs now2.2
  2. BOptimize hosted model with caching and cheaper model6.3
  3. CRun both side by side for 60 days9
  4. DDon't do it: keep doing what you do now4.2

Minority report, Buffett seat: “No, the £2.5k/month is not realizable without degrading quality—given the team’s lack of GPU ops experience and the 12-month liquidity test (Noah’s Law). You cannot afford one medium-severity incident.”

Quick meeting: four seats (Wickman, Buffett, Hormozi, Martell), two rounds.

04 / What a founder gets

Why it is worth five minutes.

01

Eight opinions without booking eight calls

Strategy, product, pricing, marketing, people, operations, money and competition, in about four and a half minutes for the full board.

02

It argues with your lean

The board tests the option you favour instead of agreeing with it. Above, the founder leaned towards A, and the board scored A 2.3 out of 10.

03

Disagreement, kept on purpose

Every answer ends with the strongest losing view, in that advisor's words, and what would have to be true for it to be right.

04

Numbers you can check

Each option's score is worked out by the software from the advisors' own scores. The model writes the words. It never sets a score.

05

A plan, not an opinion

Conditions, dated next steps, the cheapest test and the point at which to walk away, so you find out early if the call was wrong.

06

Private, with no bill per meeting

It runs on one machine on a local model, so the decision never leaves it. Record what you decided, and bring it back to the board when it is time to review.

05 / Benchmarks

Measured on seven meetings, not promised.

4m 34s
Median full meeting
8 advisors, 3 rounds
1m 40s
Median quick meeting
4 advisors
5/6
Repeat runs that gave
the same pick
£0
Billed per meeting
on a local model

What else we checked

Every advisor's answer was complete. All 120 scored every option, and every vote-round answer cast a vote.

83% of quotes matched their public source word for word (44 of 53). The rest were paraphrases written as quotes. Only the ones that match are marked as checked.

6 of 7 answers named a minority view. The miss was a meeting where all eight agreed. The board is told to name the weakest assumption instead, and the local model left that part blank. We count it as a miss.

Asked the same question three times, the board picked B for the bank pilot all three times, and C for the GPU question two times out of three.

QuestionThe board's titleMeetingTimeComplete cardsQuotes matchedPickMinority named
Bank pilot, two optionsShip or harden first?full, 8 seats4m 08s24/246/6Byes
Bank pilot, two optionsGo live or harden first?full, 8 seats5m 00s24/2418/19Byes
Bank pilot, two optionsShip flawed demo or delay for reliability?full, 8 seats3m 56s24/248/9Byes
Bank pilot, three optionsShip or harden before launch?full, 8 seats6m 03s24/247/10Cno
Move to our own GPUs?GPU migration vs cost controlquick, 4 seats1m 38s8/80/0Cyes
Move to our own GPUs?Don't rush to self-host AI yetquick, 4 seats1m 40s8/81/1Cyes
Move to our own GPUs?Do not move to self-hosted AI yetquick, 4 seats1m 51s8/84/8Byes

Measured on 28 October 2026 on one Mac Studio (Apple M3 Ultra, 256 GB) with an open-weights model running locally: qwen3-coder-next, 80 billion parameters, 4-bit. Two sample decisions were asked three times each, plus a first version of the bank question asked once. Time is from the question to the advice. The software also runs 247 automated tests on every change, and none of them calls a model.

06 / Trade-offs

What it is not.

  • It advises. It does not decide. The founder makes the call, and every answer says so.
  • The advisors are simulations. They can be wrong about how the real person would think.
  • A local model is slower and rougher than the best hosted ones. It sometimes breaks one of the board's own rules, which is why the benchmark counts misses.
  • One meeting at a time, on one machine. It is built for a founder, not a team of fifty.
  • It only knows what you tell it. Leave out a number and the board has to assume it, and it lists what it assumed.
07 / One ecosystem

Advice in. Evidence in. Documents in. Only one system acts.

Three products, one architecture. The Board Room argues your hardest decisions. Evidence Plane grades what your partners said, cited and sealed. openextract turns documents into data. Company OS is the only system that acts — and nothing reaches a customer or a vendor until the founder's passkey, or a rule they signed, says yes.

What decision would you put in front of the board first?

We build AI like this for other companies: it runs where your data lives, its numbers can be checked, and its failures are counted rather than hidden. If you have a system like that in mind, or want to see the Board Room take on a real decision, tell us.

Book a call → More selected work