Skip to content
Back to results
Anonymized client

Aurora

Council — multi-agent decision system

Aurora is the Council in production: a decision system that convenes a panel of specialist AI agents, has each reason independently, debates and peer-ranks them, then synthesizes one recommendation — with confidence, evidence, and a full audit trail.

Industry
High-stakes professional advisory
Year
2025
What we ran
Product architecture, AI engineering, delivery
ASD · COUNCILLive

Confidence score

0%avg

Live debate rounds

2 / 3

Verification checks

0%passed

QUERYShould we launch the SMB self-serve plan in Q3, or hold for the rebrand?
POLICYORCHESTRATOR
Live council sessionRound 2 / 3
StrategySoft-launch beats hold on expected value.+0.00
FinanceMargin holds at the proposed price.+0.00
LegalSelf-serve ToS needs sign-off first.0.00
Verdict snapshot
0.84

Recommended

Soft-launch — gated to a 10% cohort

Overall confidence: 0.84

Faster to a recommendation you can defend

Full

Audit trail on every recommendation

60%

Less time spent chasing scattered context

The problem

Experts were making high-stakes calls under time pressure, with the context they needed scattered across systems and no second opinion on hand. A single model gave confident but unaccountable answers — a non-starter when the cost of being wrong is measured in real consequences.

How we solved it

We replaced the single model with a panel. Each specialist evaluates the case from its own discipline; a synthesis layer weighs their reasoning, surfaces where they disagree, and produces one recommendation with the rationale attached — never a verdict you have to take on faith.

What it does

01

A multi-model council where each specialist reasons independently before synthesis

02

A rationale trail on every recommendation — you can see exactly how it was reached

03

Confidence scoring, with disagreement flagged rather than averaged away

04

A reviewer workspace that keeps a human firmly in command of the call

Under the hood

  • Multi-model orchestration
  • Retrieval & memory
  • Reasoning synthesis
  • Audit logging
How it works

How multi-agent decisions work.

From prompt to verified recommendation.

01

Input

We take the question and the context.

02

Specialist selection

A chief model picks the 4–7 advisors that fit.

03

Live debate

Advisors reason and challenge each other across models.

04

Peer ranking

Each advisor ranks the others — consensus emerges.

05

Synthesis

Positions converge into one grounded recommendation.

06

Plan-B verification

Independent models re-check the call.

07

Final verdict

A recommendation with confidence, reasons, and sources.

Traceable reasoning

Every step supplied

Cross-model debate

Diverse perspectives

Confidence scoring

Calibrated, not guessed

Evidence pack

Sources in one place

Audit trail

Decisions you can defend

Transparent. Verifiable. Defensible.

The experts' shared brain

Every expert. The best models for the task.

Your AI experts aren't tied to one model. Each role reaches for the right model on every sub-task, then ASD AI synthesizes their answers — grounded in your data, backed by evidence.

Fable 5

Analysis

GPT-5.5

Reasoning

Gemini 3.5 Pro

Synthesis

Perplexity

Research

Synthesis

Grok 4.3

Logic

Llama 4

Local

Embedding

Context

Mistral

Efficiency

Right model, right task

Each expert picks the best model per sub-task

No single-model bias

Strengths combined, weaknesses cancelled out

Grounded in your data

Answers drawn from your systems, not guesses

Evidence on every answer

Citations and reasoning you can verify

How it works

1

Each expert routes

Every role sends each sub-task to its strongest model

2

Models do the work

Specialized models run in parallel against your data

3

ASD AI synthesizes

One grounded answer, with confidence and sources

Many models. One synthesized answer.

Not locked to a single model — each expert draws on the best, grounded in your business.

Verified · evidenced · governed

Trust every decision your team makes.

Every decision your AI expert team makes is verified, evidenced, and governed — so you can act with confidence.

0.87

Confidence score

High confidence

Well-supported by data and verified reasoning.

Top 13% of decisions

Evidence pack

24

Key inputs

18

Supporting docs

12

Model citations

Open evidence pack
Audit trail

100%

Traceable

27

Events recorded

0

Manual overrides

Compliance checksCompliant

All checks passed

12 / 12 controls validated

Disagreement

2 advisors disagreed

Reviewed & resolved · rationale documented

Plan-B verified

Ready for activation

End-to-end encryption
Immutable logs
Tamper-evident

What would your expert team take on?

Tell us the work that eats your week. We'll show you which AI expert takes it off your plate — working inside your existing systems — and what it saves you.