A council of AI advisors · a journal of your decisions
Decide with a council. Learn from the outcome.
Ask OpenAI, Anthropic, Google and xAI the same question at once. Get one attributed recommendation, save it as a decision, and come back to record how it turned out.
Our billing service had three incidents this year, all in retry handling. Should we postpone its rewrite until after the Q2 launch and harden the monolith instead?
Postpone. A rewrite during the launch competes with it for the same engineers, and the incidents you describe are in retry handling, not the architecture. Harden the monolith now: idempotent retries, a reconciliation job, alerts on failed charges.
Postpone, but set a date. The rewrite may be worth doing; doing it under launch pressure is what makes it risky. Spend two weeks putting a clean billing interface inside the monolith so the later move is mechanical.
Postpone. All three incidents were in retry handling, which you can fix without a rewrite. Fix the retries first. Revisit the split if invoice volume doubles or you add a second payment provider.
Only if you can freeze billing changes during the launch. If it adds plans or currencies you will change billing anyway, and doing that twice costs more than doing the rewrite once.
Postpone the rewrite until after the Q2 launch. Spend the next two weeks hardening retries and drawing a clean billing interface inside the monolith.
AI self-assessment, not a calibrated probability.
Launch adds new plans or currencies, so billing changes twice.
Hardening slips and the rewrite never gets a date.
Add idempotency keys and a nightly reconciliation job.
Put a billing interface in front of the existing calls.
Book the split for the first sprint after launch.
Launch scope adds a plan, currency or payment provider.
Invoice volume doubles.
A billing incident traces to the architecture, not retries.
Example council of 4 seats. The free plan seats 2.
Why a council
One model sounds sure. Four models show you where to look.
Different models disagree in useful ways.
They are trained differently and weigh risks differently. Side by side, the gaps in any single answer show.
Disagreement tells you where to dig.
When one advisor says "only if", that condition is often the real decision. Ask it a follow-up with an @mention.
Agreement across providers is a stronger signal.
Three providers reaching the same call independently says more than one model's confidence in itself.
Before and after
From four tabs and copy-paste to one thread.
- Paste the same question into four tabs.
- Copy the answers into a doc and reconcile them yourself.
- By next quarter, nobody remembers why you decided.
- Ask once. Every answer streams in side by side.
- One recommendation, each line attributed.
- Saved as a decision, with a reminder to review it.
The decision journal
The council answers. The journal remembers.
Save the synthesis as a decision with a title, a review date and the signposts that would change it. When the date comes, record whether you followed it and how it turned out. Over time you learn which calls held up.
How it works
Four steps, from question to lesson.
Seat your council
Pick 2 to 4 advisors, one each from OpenAI, Anthropic, Google and xAI, and choose each one's model.
Ask
One question goes to every seat. Answers stream in side by side. Follow up with one advisor using @mention.
Call the secretary
@secretary returns one recommendation, its risks, next actions and signposts, each line attributed. It also records each advisor's stance, which gives the consensus count.
Save and review
Keep it as a decision with a review date. A reminder brings you back to record the outcome.
Who it's for
Built for consequential calls. Not for everything.
A good fit if you
- Lead a product, engineering team or company, or advise one.
- Face calls that are costly to reverse: budget, hiring, architecture, vendors, strategy.
- Want to see where models disagree before you commit.
- Want a record of why you decided, and whether it held up.
- BudgetDo we put next quarter's $40k into paid acquisition or into onboarding?
- Technical trade-offRead replicas now, or a cache layer first?
- HiringHire a senior platform engineer, or contract the migration for six months?
- Vendor or contractSign three years at 20% off, or stay on annual terms?
Not the right tool for
- Quick factual lookups. A search engine or one model is faster and cheaper.
- Casual chat, or writing help with no decision at the end.
- Tasks you want done for you. The council advises; it doesn't act.
- Replacing a lawyer, doctor or financial advisor.
Documents and images
Put the evidence on the question.
Attach a PDF, a Word document or a screenshot. Every seated advisor works from the same material. Models that read images are marked as such.
Web search
One search per question, with sources.
The council can draw on one web search for a question. Sources are listed under the answer, so you can check them.
Privacy and trust
Your hard questions stay yours.
Plain statements about what happens to what you ask. The Privacy Policy has the details.
Sent only where it needs to go
Your question goes to the AI providers you seat. If you turn on web search for a question, the search terms go to our search provider. If you report an answer, that answer and its question may be sent to the provider whose model wrote it.
Not in our logs
Questions, answers and decisions are never written to our logs.
Payments by Paddle
When paid plans launch, Paddle is the merchant of record. It handles payment, tax and invoices; we never see your card.
Export or delete, any time
Take your account with you, or delete it, from Settings.
Pricing
Start free. Go Pro for a full council.
FAQ
Questions, answered plainly.
Which models sit on the council?
One model each from OpenAI, Anthropic, Google (Gemini) and xAI (Grok). A council has 2 to 4 seats, and you choose the model for each seat from the tiers your plan includes.
What does the confidence figure mean?
Each advisor rates its own confidence; the figure summarizes those ratings. It is an AI self-assessment, not a calibrated probability, and not a measure of accuracy. Treat the consensus count and the disagreements as the stronger signals.
What does the secretary do?
When you ask @secretary, it reads the council's answers and writes one synthesis: a recommendation, 2–3 risks, 2–3 next actions and 1–3 signposts. Every line names the advisors it came from. It also records each advisor's stance, which gives the consensus count, such as 3/4 favor.
Is my data used to train models?
We don't train models on your content. Your question is sent only to the providers you seat, through their APIs, under their API data terms. The Privacy Policy lists each provider's terms.
What is the decision journal?
Where saved decisions live, with their title, review date and signposts. When a review is due, you get an email or push reminder, and you record whether you followed the decision and how it turned out. Search it, or export a decision as Markdown, JSON or print.
Are there mobile apps?
Yes. The iPhone and Android apps are in beta. They use the same account, conversations and journal as the website.
Can I cancel?
Once paid plans launch, cancel any time from Settings, in Paddle's customer portal. Pro stays active until the end of the period you paid for. See the refund policy.
Is this a chat app?
Not really. It is built for questions that end in a decision. For quick lookups or casual chat, a single model is a better fit.
Your next hard call is a good first question.
Start on the free plan. Seat two advisors and ask.