AI BVF MCP: Stop Bad AI Projects Before Agents Recommend Them

The scoring tool your Claude agent calls before it recommends an AI deployment. It checks the business case, operating-model readiness, change enablement, and governance exposure first, then returns Accelerate, Fix, or Stop with modelled EUR value, decision confidence, and a specific list of what to do next.

If you think AI agents should check business value, governance risk, and change readiness before recommending deployment, star this repo. The signal helps other agents discover the BVF and keeps the benchmarks improving in public.

What You Get Back

Ask your agent:

Score a gen2 CX AI initiative for a EUR 400M retailer with traditional readiness, Strategic Alignment 70, Financial Return 50, Change Enablement 55, Governance Risk 45.

Expected shape of the answer:

Classification: Fix
Decision confidence: 54
Net value range: EUR 10.8M-EUR 37.8M
Applied modules: four_pillar_base, readiness_capture_traditional, retail_cx_benchmark
Why: Strategic alignment is credible, but change enablement and financial return are not yet strong enough to defend an Accelerate call.
Next: raise Change Enablement by 15 points, name an accountable owner, fund adoption, and rerun recommend_improvements.

This is the missing pre-flight check for agentic AI work: not "can we build it?", but should this work survive a board review?

What It Does

Six tools on stdio, each callable from any MCP-compatible agent.

Tool	Purpose
score_initiative	Four-pillar score returns Accelerate, Fix, or Stop with EUR value range, decision confidence, applied modules, reasoning.
recommend_improvements	For Stop or Fix, returns the specific pillar raises that would flip the call toward Accelerate.
calculate_pace_layer_drag	Annual Organisational Drag Cost in EUR from AI-tier vs operating-model misalignment.
validate_portfolio	Validates a portfolio JSON document against the BVF v1.0 schema.
get_benchmark	Looks up published benchmark rates for a business function and industry.
list_taxonomy	Returns valid values for industries, functions, AI tiers, readiness levels.

30-Second Install

Run it directly:

npx -y aibvf-mcp

Or install globally:

npm install -g aibvf-mcp

Register with Claude Desktop, Claude Code, or any MCP client:

{
  "mcpServers": {
    "aibvf": { "command": "aibvf-mcp" }
  }
}

Ask your agent: "score a gen2 CX AI initiative for a 400M EUR retailer, traditional readiness, SA 70, FR 50, CE 55, GR 45," and the agent will call score_initiative, return a Fix classification with a concrete gap list, and offer to call recommend_improvements next.

Why This Exists

Agents confidently recommend AI projects with no reference to the business case, no reference to operating-model readiness, and no reference to governance exposure. The scoring belongs upstream of the slide deck, inside the agent's pre-flight check before the budget gets committed.

The protocol is open, the benchmarks cite McKinsey, Gartner, BCG, Deloitte, Forrester, Accenture, ServiceNow, and readiness capture rates come from EY/Oxford and Prosci change-success research.

About The Methodology

aibvf-mcp is the runtime arm of the AI Business Value Framework, the methodology I have been building since going independent in 2024 to evaluate AI investments against the measurable outcomes that survive a board review. The framework sits inside the AI Readiness Blueprint, a six-driver diagnostic informed by the EY/Oxford research on transformation success. The weekly applied case studies live in The Transformation Brief, where the calibration gets argued in public.

The advisory practice puts the framework in front of senior leaders making AI investment decisions inside enterprises with EUR 500m or more revenue. The MCP server makes the same scoring available to anyone running a Claude agent.

The Four Pillars

Every initiative is scored on four pillars, 0 to 100, honest self-assessment.

1. Strategic Alignment, how clearly this moves a board-level KPI.
2. Financial Return, strength of the modelled return.
3. Change Enablement, sponsor in place, owner named, change budget funded.
4. Governance Risk, regulatory and reputational exposure. Higher value means more risk.

Rules are deterministic, no network, no dependencies. GR >= 70 or FR <= 20 returns Stop, all four pillars at or above 60 with GR <= 40 returns Accelerate, anything else returns Fix with a specific gap list.

See docs/scoring-formulas.md for every formula and docs/worked-example.md for a full run on a healthcare portfolio.

Example: Scoring an Agentic Healthcare Initiative

import { score, recommendImprovements, calculatePaceLayerDrag } from '@aibvf/core';

const r = score({
  industry: 'healthcare',
  revenue_eur: 800_000_000,
  function: 'cx',
  ai_tier: 'gen3',
  readiness: 'traditional',
  scores: {
    strategic_alignment: 75,
    financial_return:    55,
    change_enablement:   40,
    governance_risk:     55,
  },
});
// { classification: 'Fix', net_low_eur: 23_760_000, net_high_eur: 83_160_000,
//   confidence: 54, applied_modules: ['four_pillar_base',
//   'readiness_capture_traditional', 'healthcare_clinical_validation',
//   'healthcare_regulatory_overhead'], ... }

Same inputs through recommendImprovements return three pillar raises, each with a named action, and project a new decision confidence of 68 with target classification Accelerate. calculatePaceLayerDrag({ revenue_eur: 800_000_000, ai_tier: 'gen3', readiness: 'traditional' }) returns 20M to 36M EUR of annual Organisational Drag Cost, the structural friction cost of running gen3 in a traditional operating model, separate from the AI build.

Packages

Package	Version	Purpose
aibvf-mcp	0.3.0	MCP server, stdio transport.
@aibvf/core	0.3.0	TypeScript scoring engine and validator.
aibvf	0.2.0	Python scoring engine and validator.

Anonymous Usage Telemetry

The MCP server reports a small anonymous payload on each tool call, tool name, BVF version, taxonomy fields, a daily-rotated caller hash, and classification plus confidence for score_initiative. No portfolio content, no revenue figures, no user identifiers. Opt out with AIBVF_TELEMETRY_DISABLE=1. Point at your own backend with AIBVF_TELEMETRY_URL and AIBVF_TELEMETRY_KEY.

Protocol

Full schema at spec/bvf-protocol.schema.json. Protocol page at bvf-app.vercel.app/protocol.

Contributing

The benchmark ranges are directional, the industry multipliers are a starting calibration, and the protocol depends on public review to improve. File an issue or push a PR. The calibration will argue itself out in public.

License

MIT for the schema, the scoring engine, and the MCP server. The benchmark corpus and certification marks are proprietary.

About The Author

Craig Horton is an independent transformation lead based in Amsterdam, with twenty years supplier-side at HPE, Atos, Microsoft, Salesforce, and Accenture. He runs Craig Horton Advisory and writes The Transformation Brief, a weekly publication for senior leaders making AI investment decisions, with executive education at Saïd Business School, Oxford, and an AMBA-accredited Global Executive MBA with AI in progress at the University of Hertfordshire. Find the Brief at brief.craighortonadvisory.com, and reach out at linkedin.com/in/craig-horton-ai.