How We Built an AI Discovery Agent That Scales Strategic Consulting Without Replacing Consultants

How We Built an AI Discovery Agent That Scales Strategic Consulting Without Replacing Consultants

The discovery phase of a digital transformation is where the most important thinking happens — and where the economics break down fastest. A typical 10-week discovery engagement for a mid-market organization runs $75K–$150K, most of it senior consultant time: CTO-level thinking, business analysis, technical architecture, and facilitation happening simultaneously.

The scaling problem is structural. You can’t hire more senior consultants without degrading their seniority. You can’t automate away the judgment, facilitation, and relationship work. And you can’t ask overloaded staff to run structured workshops while also documenting requirements and synthesizing decisions across weeks of conversation.

We asked a different question: What if you could decompose discovery into specific, orchestrated roles and let an AI system manage the workflow between them?

The Insight: Discovery Is Six Jobs, Not One

The breakthrough was recognizing that discovery work isn’t monolithic. It decomposes into six distinct roles with clear boundaries:

1. Meta-Orchestrator — Coordinates the other agents, manages workflow sequencing
2. Discovery Lead — Owns facilitation, stakeholder engagement, political dynamics
3. Business Analyst — Captures requirements, maintains living requirements documents, performs gap analysis
4. Digital Strategist — Maps the technology landscape, evaluates vendor options
5. Technical Architect — Designs solution architectures, evaluates tradeoffs
6. Reference Librarian — Maintains the knowledge base, surfaces relevant context across workshops

Most discovery teams have one or two people doing all six jobs. What if each role had an AI agent with specific expertise, defined inputs and outputs, and access to all prior workshop context?

The agents run as an orchestrated system — not in isolation. The Discovery Lead agent prepares for workshops by pulling context from the Reference Librarian. The Business Analyst captures requirements and immediately flags feasibility concerns to the Technical Architect. The Digital Strategist continuously updates the technology landscape as new information surfaces.

The human discovery lead still runs the room. But instead of transcribing, synthesizing, and cross-referencing alone, they orchestrate AI agents through the workflow.

How It Works: 10 Workshops, Structured Outputs

We structured the discovery into 10 workshops with defined inputs and outputs:

W1 (Kickoff): Scope alignment, stakeholder introductions, current-state walkthrough. Heavily facilitated, minimal AI involvement — this is where psychological safety and trust get established.

W2–W5 (Requirements): One workshop per major function. The BA agent prepares requirements templates beforehand; during the session, the human facilitates while the AI captures and structures. Post-workshop, the agent synthesizes findings and flags conflicts to the discovery lead.

W6–W7 (Technology Evaluation): The Digital Strategist agent pre-researches vendor options and prepares comparison matrices. Workshops focus on tradeoff analysis — cost vs. capability, integration effort vs. built-in functionality, vendor lock-in vs. standardization.

W8–W9 (Architecture): The Technical Architect agent proposes solution architectures based on requirements and technology evaluation. Workshops stress-test the proposals against deployment complexity, data migration, and performance.

W10 (Executive Readout): The recommendation is built from 9 weeks of AI-synthesized inputs — not a manual compilation assembled at 10 p.m. the night before.

Every workshop output auto-syncs to a living knowledge base, cross-referenced and searchable. By W8, stakeholders are actually referencing earlier decisions and changing positions based on documented tradeoffs.

What Went Wrong Along the Way

We’d be less than honest if we didn’t share the friction. The first-generation agent prompts were too generic — a "Digital Strategist" agent that works for healthcare has fundamentally different evaluation criteria than one for events management. We burned the first few days of each engagement tuning prompts to organizational context.

The stateless-vs-stateful decision was also hard-won. Early versions ran agents as stateless calls — they’d re-propose alternatives that had been rejected in earlier workshops. Moving to stateful agents with full workshop history access eliminated the problem but added complexity around context management and token budgets.

And the first workshop remains stubbornly human. W1 establishes psychological safety and scope agreement. No amount of AI orchestration replaces the discovery lead reading the room, sensing fear about organizational change, or recognizing when a stakeholder says "tool problem" but means "political problem."

Where Humans Still Lead

This is the critical part: the AI agent did not replace the human discovery consultant.

The discovery lead was in every workshop — facilitating conversations, managing political dynamics, making judgment calls about when to push back and when to let things breathe. When a development team expressed fear about losing donor data, that wasn’t a requirements gap — it was a trust issue. The human consultant recognized this and asked the architect agent to add a security and data governance section directly addressing the concern.

The strategic insight that emerged in the final weeks — "What if we flip the data model from event-centric to stakeholder-centric?" — came from a human sitting in nine workshops, absorbing patterns, and having a conversation that no individual agent could initiate.

The AI accelerated the work. It didn’t eliminate the need for an expert in the room.

What This Means for Agencies and Clients

For agencies: you can run more discovery engagements with the same senior consultants. The work of discovery becomes more explicitly about judgment, facilitation, and strategic sense-making — less about transcription and documentation overhead.

For clients: you get discovery that’s more rigorously documented, consistently cross-referenced, and thoroughly synthesized. You also get a living knowledge base you can reference for the next three years of transformation.

The biggest win isn’t speed — it’s consistency. An orchestrated system generates consistent output every week. You can plan downstream work with more confidence because every recommendation traces back to a specific requirement, stakeholder need, or technology tradeoff that was surfaced and discussed.

We published the agent prompts, workshop templates, and knowledge base structures on GitHub. The innovation isn’t in the code — it’s in the methodology: the insight that discovery can be decomposed into six specific roles with clear inputs and outputs, and that orchestrating them with an LLM produces better outcomes than traditional human-only discovery.

CTA

If you’re managing a complex multi-stakeholder transformation and want to explore whether this discovery model fits your context, let’s talk.

Interested in the methodology? We’ve published the agent prompts and workshop templates on GitHub.