200+ clients, 500+ team members, 8 AI tools. How the operating system works.

When people ask how AC Solutions runs a growing business with a lean team across 30 countries, the honest answer is: AI councils for decisions, AI agents for information retrieval, humans for execution and relationship management.

This is the full operating system.

THE DECISION LAYER: AI COUNCILS

Every significant decision — pricing changes, market entry, hiring above a certain level, product roadmap shifts — goes through a structured multi-agent debate before any human commits to it.

The council architecture: 5 expert agents with assigned roles (CFO, CMO, CTO, Devil's Advocate, Domain Expert) debating in parallel on DeepSeek-chat. Gemini 2.5 Pro with extended thinking synthesizing the outputs. DeepSeek Reasoner as a final sanity check for high-stakes decisions.

Council sessions run under 10 minutes for most questions. The output is a structured document: recommendation, key disagreements between agents, open questions, minority views. I read it. I decide.

THE INFORMATION LAYER: VIRTUAL CEO

The decision layer only works if it has context. Our Virtual CEO system ingests all company data — CRM activity, meeting transcripts, financial reports, project updates, client health signals — into a vector database (Pinecone). When I or any executive asks a question, it retrieves relevant context from all sources and produces a grounded answer.

"Which clients are at churn risk this month?" returns an actual list from CRM data, not a generic framework for thinking about churn.

This is what prevents councils from reasoning from general principles when they should be reasoning from specific company reality.

THE EXECUTION LAYER: HUMANS PLUS TOOLS

Councils decide. Virtual CEO informs. Humans execute.

The execution layer is where we have NOT automated aggressively. Client calls, relationship management, negotiation, complex support escalations — these stay with humans. The failure mode of automated execution in high-stakes external interactions is too costly.

Where we have automated execution: routine communications, campaign scheduling, compliance reporting, candidate screening. Rule-based automation with human review at defined checkpoints.

THE OPERATING CADENCE

Daily: overnight batch analysis arrives at 8 AM — competitive signals, client health scores, campaign performance anomalies. Team reviews and flags items needing action.

Weekly: council session on one major strategic question. 10 minutes of AI debate, 20 minutes of human review and decision.

Monthly: Virtual CEO generates a structured company state document — pipeline, client health, team capacity, financial position. Used as context for the quarterly planning council.

THE NUMBERS

- Strategic decision time: 40% reduction vs 2024 - Decision reversal rate (within 30 days): from 22% to 8% - Information retrieval time per executive: from 14 hours/week to 5 hours/week - External consulting spend significantly reduced

THE FAILURE MODES

Autonomous execution agents: we tried them in Q2 2024. Confident plans, no correct execution. Abandoned.

Council sessions without context: a council reasoning from general principles, without knowing our specific client base and operational state, produces generic advice. Quality depends directly on briefing quality.

Overconfidence in synthesis: Gemini's synthesis is compelling and well-structured. It can be wrong. The council is an input to decision-making, not a replacement for it.

THE PLAYBOOK IN ONE SENTENCE

Give AI agents the information they need, assign them adversarial roles, let them argue, synthesize the argument into a structured document, and then make the decision yourself.

→ The full stack is documented and partially open source. nealkhis.com