What a Fixed-Fee AI Discovery Engagement Actually Looks Like

Why Fixed-Fee for AI Discovery Work?

Fixed-fee discovery eliminates the friction that kills AI adoption projects before they start. The typical pattern: CEO wants AI assessment, gets quoted $150-300/hour with no timeline cap, decides to "wait until Q2" instead of approving an open-ended expense. Six months later, the same cycle repeats.

Hourly billing creates perverse incentives in discovery work. The consultant benefits from extending the assessment phase. More interviews, more system audits, more analysis - all billable hours. The client pays for activity, not answers. Fixed-fee forces both parties to focus on outcomes: specific recommendations, quantified opportunities, and a clear path forward.

According to McKinsey's research on AI adoption barriers, 47% of executives cite "unclear ROI projections" and "undefined implementation costs" as primary obstacles. Fixed-fee discovery addresses both concerns upfront. The CEO approves a known budget for known deliverables. No scope creep, no surprise invoices, no consulting theater.

The traditional consulting model produces 120-page reports that executives skim and file away. Fixed-fee discovery produces working software against real data. The CEO sees Claude pull actual invoices from NetSuite, cross-reference them with project costs in Procore, and generate a cash flow analysis in real-time. That demonstration is worth more than any PowerPoint deck.

The Four-Week Discovery Timeline

Discovery runs exactly four weeks from kickoff to final presentation. Each week has specific, measurable deliverables that build toward the live demo.

Week 1 focuses on stakeholder mapping and business context. Schedule interviews with 3-5 key stakeholders: the CEO, the operations leader, the person who handles the most manual processes, and whoever manages IT decisions. Each conversation is 60-90 minutes. The goal is identifying which manual processes consume the most time and create the most frustration. Document current workflows, data sources, and integration points between systems.

Week 2 builds the first Model Context Protocol server against the primary business system. For most clients, this means financial data from NetSuite, project data from Procore, or customer data from Salesforce. The MCP server is read-only during discovery - Claude can query data but cannot modify anything. This eliminates executive anxiety while demonstrating capability. By week's end, Claude should answer basic questions about the client's real business data.

Week 3 adds the second integration and enables cross-system scenarios. Build an MCP server for the secondary system identified in Week 1. The breakthrough moment: Claude pulls data from both systems in a single conversation, cross-references information, and generates insights that require manual work today. Test edge cases, handle authentication issues, and refine the system prompt that guides Claude's responses.

Week 4 delivers the live demonstration and Phase 2 proposal. The CEO watches Claude perform 2-3 real business tasks using their actual data. The presentation includes quantified ROI projections, a roadmap of additional opportunities, and fixed-fee pricing for the build phase.

According to Anthropic's Economic Index, organizations that complete AI pilots within 30 days have 3.2x higher deployment rates than those with longer assessment cycles. The four-week timeline creates momentum and maintains executive attention.

What You Actually Get: The Live Demo Against Your Data

The centerpiece deliverable is a working AI system that queries real business data across multiple systems. This is not a proof-of-concept with sample data. Claude accesses the client's actual NetSuite instance, their actual Procore projects, their actual Salesforce records.

The read-only principle builds trust while demonstrating value. Claude can look at invoices but cannot modify them. It can analyze project costs but cannot change budget allocations. Executive concerns about "AI touching our systems" disappear when they understand the constraints. The system provides insights without creating risk.

A typical demo shows 2-3 use cases that represent patterns, not one-off solutions. For a construction client: Claude pulls committed costs from Procore, matches them against purchase orders in NetSuite, and identifies budget overruns by project phase. For a professional services firm: Claude analyzes time entries across projects, flags clients approaching budget limits, and generates utilization reports by practice area.

The technical architecture follows Model Context Protocol standards, enabling seamless integration between Claude and existing business systems. Each MCP server exposes specific resources (data reads) and tools (actions) that Claude can use. The system prompt defines the AI agent's role, available capabilities, and response format expectations.

Cross-system integration is where manual processes break down today and where AI delivers the highest ROI. Every executive recognizes the "swivel chair" pattern: pull data from System A, manipulate it in Excel, upload results to System B, send email updates to stakeholders. Claude eliminates those manual handoffs.

The ROI Model You Actually Need

The financial model quantifies current costs and projects savings in language boards understand. Instead of vague "efficiency gains," the model calculates specific dollar impacts based on measured time costs and error rates.

Start with headcount hours. How many hours does the finance team spend each month reconciling invoices? How many hours does operations spend generating project status reports? How many hours does the executive team spend preparing for board meetings? Multiply those hours by loaded rates (salary plus benefits plus overhead) to establish current costs.

According to research from Anthropic's Economic Index, organizations that quantify AI ROI in "hours saved per month" rather than percentage improvements achieve 2.4x higher approval rates for implementation budgets. Executives think in terms of people costs and time allocation, not abstract productivity metrics.

Error costs often exceed time costs. Manual data entry has measurable error rates. Invoice processing errors trigger payment delays, vendor relationship issues, and cash flow problems. Project budget overruns due to poor visibility cost more than the hours spent on reporting. The ROI model quantifies both direct time savings and avoided error costs.

The payback calculation is straightforward: implementation cost divided by annual savings equals payback period in months. Most mid-market AI implementations achieve payback within 6-12 months. Simple enough for board presentations, detailed enough for CFO approval.

Project API costs and build costs for the full implementation. Claude API usage scales with query volume, but costs are typically 5-10% of the labor savings. Implementation costs include MCP server development, system integration, and user training. Ongoing maintenance costs include monitoring, support, and periodic system updates.

Technical Architecture Decisions Made in Discovery

Discovery establishes the technical foundation for Phase 2 implementation. Every architectural choice balances capability against complexity, favoring approaches that ship quickly and scale reliably.

Direct Claude API integration eliminates framework overhead and vendor lock-in. LangChain adds complexity without providing material benefits for most business use cases. The Claude API provides conversation management, structured output, and tool calling natively. Building directly against the API ensures compatibility with future Anthropic releases and avoids framework-specific technical debt.

Model Context Protocol servers handle all external system integrations. Each MCP server focuses on a single business system: one for NetSuite, one for Procore, one for Salesforce. This separation of concerns simplifies development, testing, and maintenance. Authentication is handled at the MCP server level, isolating credentials and API keys from the main application.

System prompt architecture defines the AI agent's role and capabilities. The prompt includes domain knowledge specific to the client's business, available tools and when to use them, and response format expectations. Prompt development is iterative - refined through Weeks 2-3 as you learn what the model needs to know about the client's context.

Context stuffing versus RAG depends on corpus size and query patterns. For document-heavy use cases with large knowledge bases, implement retrieval-augmented generation with proper chunking, embedding, and reranking. For smaller document sets or structured data queries, load relevant context directly into the system prompt. RAG adds complexity and latency - use it only when context window limitations require it.

Authentication patterns favor OAuth 2.0 for enterprise systems like NetSuite and Salesforce, API keys for simpler systems. Never store credentials in application code. Use environment variables or cloud key management services. Implement proper token refresh handling for OAuth flows.

What Happens After the Demo

Phase 2 transforms the working demo into production-grade software with write access, user interfaces, and deployment infrastructure. The scope and timeline are defined by the use cases validated in discovery.

The Phase 2 proposal includes specific deliverables based on discovery findings: which MCP servers need write capabilities, what user interfaces are required, how the system integrates with existing workflows. Pricing is fixed-fee based on defined scope. Most Phase 2 implementations run 8-12 weeks depending on complexity.

Write access implementation requires additional safeguards beyond the read-only demo. Approval workflows for high-value transactions, audit logs for all system changes, rollback capabilities for problematic updates. The technical architecture established in discovery scales to support these production requirements.

User interface development ranges from simple Streamlit dashboards to Slack bot integrations to embedded widgets within existing business applications. The choice depends on user preferences identified during stakeholder interviews and the technical constraints of existing systems.

Deployment infrastructure includes monitoring, alerting, and ongoing maintenance capabilities. Helicone provides observability for API usage and performance metrics. Custom dashboards track business metrics like hours saved, error rates, and user adoption. See our approach to operational metrics and monitoring for detailed implementation guidance.

Training and change management ensure successful adoption. Document new workflows, train power users, establish support processes. The goal is sustainable usage, not just successful deployment. Phase 2 includes user training and documentation as standard deliverables.

Discovery Engagement Pricing and Scope Options

Two discovery tiers serve different market segments with appropriately scaled scope and pricing.

Essentials Discovery targets smaller organizations with 25-100 employees and straightforward technical environments. The engagement includes 2-3 stakeholder interviews, mapping of 1-2 core business systems, and identification of 5-10 manual processes. The deliverable is a concise assessment document with ROI projections and implementation recommendations. No proof-of-concept build - that moves to Phase 2. Pricing ranges from $3,500 for simple scenarios to $12,000 for complex multi-system environments.

Enterprise Discovery serves mid-market organizations with 100-500 employees and complex technical landscapes. The full four-week timeline includes extensive stakeholder interviews, comprehensive systems audit, detailed process documentation, and the live demo against real data. Pricing ranges from $15,000 to $50,000 based on organizational complexity, number of systems, and stakeholder count.

The key differentiator is the working prototype. Essentials Discovery provides analysis and recommendations. Enterprise Discovery provides a functioning system that demonstrates value. Both include quantified ROI projections and fixed-fee Phase 2 proposals.

ROI threshold determines which tier is appropriate. If the identified opportunities justify $50,000+ in implementation investment, Enterprise Discovery provides the validation and stakeholder alignment needed for a successful build phase. If the total implementation budget is under $25,000, Essentials Discovery provides sufficient analysis for decision-making.

Organizations with multiple departments, complex approval processes, or significant technical debt typically require Enterprise Discovery. The longer timeline and deeper analysis justify the higher investment by reducing implementation risk and ensuring stakeholder buy-in.

Time-sensitive opportunities favor Essentials Discovery. If competitive pressure or operational pain creates urgency, the faster cycle provides quicker decisions and implementation starts. Enterprise Discovery is appropriate when thoroughness matters more than speed.

Questions about what you've read? Ready to see how this applies to your specific situation? Reach out — we'd be happy to walk through what a discovery engagement would look like for your organization.