How We Built a 30-Action AI Agent Platform in 6 Months

Most "AI agents" are ChatGPT with a wrapper. We built something different: a production AI operations platform with 30 autonomous actions, 9 connected data sources, and a 3-model consensus engine. It runs at a professional services firm with real users and real data.

This article covers the architecture, the hard decisions, and what we learned building AI systems that people actually trust.

The Problem: 10 Tabs, Zero Context

Every professional services firm has the same shape. They use email (Front, Outlook), task management (Asana, Linear), documents (SharePoint, Google Drive), billing (QuickBooks), signatures (DocuSign), chat (Slack, Teams), CRM (Airtable, HubSpot), and meetings (Granola, Zoom).

That's 8+ systems. None of them talk to each other. The operations team is the glue — copying context between tabs, checking deadlines across platforms, manually routing requests.

We asked: what if an AI agent sat between all of them?

The Architecture: 5 Layers

Layer 1: The Brain (Intent Classification)

When a user asks something — "find overdue tasks for Acme Corp" — the Brain classifies intent in under 200ms using a fast LLM. It returns a confidence score from 0 to 1:

• 0.9+ → Execute the matched action immediately
• 0.7–0.89 → Ask a clarifying question through a decision tree
• 0.3–0.69 → Fall back to general Q&A with semantic search
• Below 0.3 → Not an action request, just chat

The key insight: confidence scoring is more important than accuracy. A system that's 95% accurate but doesn't know when it's wrong is dangerous. A system that's 85% accurate but flags uncertainty is trustworthy.

Layer 2: Graph-Based Narrowing

When intent is ambiguous, the system doesn't guess — it asks. But not through prompt chains (which are unreliable). Through deterministic decision trees.

Example: a user says "can't log in." Is this a client or a team member? The system asks. Based on the answer, it routes to completely different actions with different parameters.

This is a graph, not a prompt. Same input always produces the same path. Fully auditable.

Layer 3: Multi-Model Consensus

For critical decisions, one AI model isn't enough. We run Claude, GPT and Gemini in parallel on the same question, then synthesize:

• All three agree: high confidence
• Two agree, one disagrees: moderate confidence, worth a second look
• All three disagree: low confidence, flag for human review

This costs 3x per query. For routine questions, single-model is fine. For contract review, compliance analysis, billing reconciliation — the extra confidence is worth every cent.

Layer 4: The Approval Queue

Here's what most AI agent builders get wrong: they optimize for automation. We optimize for trust.

Every action the AI wants to take goes through an approval queue. The human sees:

• What the AI wants to do
• Why (reasoning)
• How confident it is (0–100%)
• What data sources it checked
• One-click approve or reject

Low-risk actions (data verification, flagging) auto-complete. High-risk actions (sending emails, creating documents, billing entries) require explicit approval.

"AI proposes. Human decides." This is the pattern that makes enterprise adoption possible.

Layer 5: The Data Pipeline

The AI is only as good as its data access. We built connectors to 9 systems with automatic sync: task management every 10 minutes, email every 15, documents every 60, chat messages every 30.

Everything is searchable via vector embeddings (38,000+ vectors). Ask "how did we handle this before?" and the system searches across all 9 sources with cited results.

What We Built: 30 Actions

Not features. Actions — things the AI actually does:

• Revenue protection: finds completed work that was never billed. Cross-references tasks against invoices. Flags gaps.
• Time savings: logs billable hours from natural language. Extracts action items from meetings. Drafts email replies with full context.
• Error prevention: reviews documents against standards. Verifies compliance across systems. Checks data consistency between platforms.
• Deal acceleration: creates signature envelopes, checks signing status, sends reminders for stale documents.

Each action has intent patterns (what triggers it), anti-patterns (what should NOT trigger it), required parameters, a confirmation template, and formatted output.

The Numbers

What We Learned

1. Trust is the product. The approval queue, confidence scores, and audit trail aren't features — they're the product. Without them, nobody uses the AI for real work.
2. Graph routing beats prompt chains. Prompt chains are non-deterministic. Graphs are deterministic: same input, same path, every time. Auditable.
3. Three models are better than one. Not for accuracy — for disagreement detection. When models disagree, that's the signal to involve a human.
4. Start in shadow mode. The AI watched and reported for weeks before taking any action. This built trust and caught edge cases without risk.
5. Anti-patterns matter more than patterns. Knowing what NOT to match prevents the most dangerous errors — doing the wrong thing confidently.

Where This Architecture Works

The architecture is industry-agnostic. The same brain, routing, consensus, and approval queue work for governance and proxy voting, ESG research, asset management, consulting, tax and accounting. The actions change. The data sources change. The brain stays the same.

If your team has analysts, clients, and data across multiple systems — this architecture works for you.