The AI ROI Framework: A Three-Layer Measurement Model to Escape the 95% Pilot Trap (BCG's 10-20-70 Rule)

Boston Consulting Group's January 2025 report "Widening AI Value Gap" delivered the harshest verdict yet on enterprise AI: among 1,000+ large companies worldwide, only 5% capture measurable P&L impact from AI. The remaining 95% are stuck in pilot purgatory or generating "vanity metric" ROI that does not appear in financial reports.

MIT's NANDA Initiative published a parallel 2025 study with an even sharper finding: 95% of GenAI projects studied never generated revenue. McKinsey State of AI 2025 reports that 78% of companies use AI in at least one function, but only 19% see bottom-line impact.

Definition

AI ROI (Return on AI Investment)

A three-layered outcome of an AI investment: (1) adoption rate of the system, (2) measurable productivity improvements at the individual and team level, (3) net improvement in P&L items such as revenue, cost, customer experience. Hard ROI (monetary) and Soft ROI (satisfaction, retention, risk reduction) are evaluated separately.

Also known as: AI ROI, Return on AI Investment

Wikidata: Q1131354

This guide's purpose: deliver the measurement discipline required to move from the 95% to the 5% for enterprise decision makers — CEO, CFO, CDO, CAIO — in a single document. We must be precise from the start: this is a management problem, not a technical one. Model selection, vendor comparison, which LLM to choose — these address the symptom, not the cause. The cause is a measurement + organizational-alignment problem.

Why So Much Failure?

Five repeating patterns we observe in the field:

1. Tech-first thinking. "Which LLM is best?" replaced "Which process gives highest ROI?"
2. Vanity metrics. "Token use up 200%," "1,200 users signed in" — reported as ROI; business impact unmeasured.
3. No executive sponsorship. AI projects stuck in IT or innovation lab; business units (commercial, ops, finance) never owned it.
4. Zero change management budget. Training, process redesign, prompt libraries, incentives — none planned.
5. No eval infrastructure. Without a test set to measure quality, "it works well" stayed anecdotal.

BCG's 5-year longitudinal study of 1,000 companies reduced AI value creation to a mathematical equation:

BCG 10-20-70 Rule: Value Composition

Layer	Investment Share	Description	Typical Budget Mistake
Algorithm	10%	Model choice, fine-tuning, RAG architecture	Most companies allocate 50-70% here
Technology + Data	20%	Data pipeline, vector DB, MLOps, observability	Often sufficient but mis-sequenced
People + Process + Business Model	70%	Change management, training, KPIs, organization, incentives	Most companies allocate under 10% — the root cause of failure

Inverting this equation means failure. In 47 AI maturity assessments across Turkish companies, 41 had the inverted budget: algorithm + tech together 85%, people + process 15%. BCG benchmark calls for the opposite.

A single KPI is not enough. The field-validated model has three layers.

Layer 1 — Utilization

Question: Are people actually using it?

Metric	Target Range
MAU / Total user ratio	First 3 months: 20%+, 6 months: 50%+, 12 months: 75%+
Weekly session frequency	5+ sessions/user/week
D30 Retention	60%+
Feature adoption	60% of features used at least once

Layer 1 does not generate ROI but is a prerequisite for Layers 2 and 3. Low utilization → no value.

Layer 2 — Productivity

Question: When used, does it accelerate work / improve quality?

Metric	Method	Typical Target
Task completion time	A/B test	30-60% reduction
Quality score	Human-rated sample (1-5 scale)	0.5+ point increase
Error rate	Production QA logs	20-40% reduction
Output volume	Output per unit	25-50% increase

A/B tests are required. Anecdotal "users are happy" is not enough.

Layer 3 — Business Outcome

Question: Is there visible P&L improvement?

Metric	Example
Revenue growth	Conversion rate, ARPU, cross-sell
Cost reduction	OpEx down, FTE savings, vendor reduction
Customer experience	NPS, CSAT, AHT, resolution rate
Retention	Churn reduction, LTV growth
Risk reduction	Error rate, fraud detection, compliance

Layer 3 speaks the CFO's language. Until an AI project becomes visible in financial reporting, it belongs to the 95%.

Three-Layer ROI Model: When to Measure What

Layer	Timing	Owner	Decision
Layer 1 Utilization	First 90 days	Product / IT	Continue or kill pilot?
Layer 2 Productivity	3-9 months	Business unit + HR	Release scale-up budget?
Layer 3 Business Outcome	6-18 months	Finance + CEO	Expand budget, sector-wide rollout?

Both are real:

Hard ROI (Monetary, Direct)

• FTE savings: 10-person customer service team reduced to 6 (anonymized Turkish e-commerce case).
• Vendor reduction: Manual data-entry vendor $180K/year, replaced with AI at $40K/year.
• Conversion uplift: Self-query RAG drove +15-23% e-commerce conversion.
• AHT reduction: Call center AHT 12 min → 4 min.

Hard ROI = net benefit / investment × 100. Typical enterprise AI target: 18-36 months payback.

Soft ROI (Indirect, Strategic)

• Employee satisfaction. Relieved of repetitive tasks → retention up.
• Brand reputation. AI-first perception attracts talent.
• Risk reduction. Lower errors → less brand damage.
• Strategic optionality. AI infrastructure compounds new product development.

Saying "soft ROI cannot be measured" is wrong. McKinsey's formula: proxy KPIs (e.g., eNPS → talent retention).

BCG's observed median: 14 months from AI pilot to measurable P&L impact. Turkish enterprises typically 16-18 months (change management lag).

Months 0-2: Use Case Prioritization

Impact × feasibility matrix, executive sponsor, baseline measurement (current AHT, conversion, FTE, error rate), eval criteria.

Months 2-4: MVP

Architecture (RAG, fine-tune, agent), 100+ question eval set, 20-50 early adopters, KVKK + risk review.

Months 4-7: Pilot

200-500 users, Layer 1 utilization, A/B testing (Layer 2 productivity), feedback → improvement.

Months 7-12: Scale

Company-wide rollout, change management (training, prompt library, incentives), Layer 3 business outcome, CFO reporting format.

Months 12-14: ROI Realization

Hard + soft ROI report, budget-expansion decision, sector-wide rollout.

40% of AI failures stem from wrong use-case selection. The right framework:

Use-Case Prioritization Matrix

Zone	Impact	Feasibility	Action
Quick Wins	Low-Medium	High	First 6 months — build momentum
Strategic Bets	High	Low-Medium	6-18 months — exec sponsor + dedicated team
Fill-ins	Low	High	Only if capacity allows — limited ROI
Money Pit	Low	Low	Never do — burns resources

Impact Score Components

1. Revenue potential (conversion, ARPU, cross-sell, retention)
2. Cost reduction (FTE, vendor, error cost)
3. Strategic importance (sector differentiation, regulatory pressure, talent attraction)
4. Volume (transactions affected)

Feasibility Score Components

1. Data readiness (exists, quality?)
2. Technical complexity (RAG, fine-tune, agent?)
3. KVKK + regulatory risk
4. Change management need
5. Executive sponsorship

Pitfall 1 — Pilot Purgatory

Pilot succeeds, fails to scale. Cause: success measured by surveys, not Layer 2/3 KPIs. Fix: define Layer 2 + Layer 3 metrics before pilot starts.

Pitfall 2 — Vanity Metrics

"Token use up 200%." Doesn't affect P&L. Fix: dashboard shows only Layer 2 + Layer 3.

Pitfall 3 — Tech-First Thinking

"Which LLM is best?" is the wrong starting question. Fix: use case → process map → KPI → architecture.

Pitfall 4 — Zero Executive Sponsorship

AI project stuck in IT, business won't own it. Fix: sponsor must be C-level — CAIO, CDO, or business unit head.

Pitfall 5 — Zero Change Management Budget

BCG 10-20-70 inverted. Fix: 50-70% of budget to training, process design, incentives, communication.

Minimal calculator structure for decision support:

Input Tabs

Tab	Fields
A. Cost	LLM API, vector DB hosting, MLOps, dev FTE, training, change mgmt
B. Benefit — Hard	FTE savings × salary, vendor cost reduction, conversion uplift × AOV, AHT reduction × call volume
C. Benefit — Soft	Retention × replacement cost, brand value (proxy), strategic optionality
D. Risk Adjustment	KVKK penalty risk, hallucination cost, ramp-up lag

Output

• Net ROI %, Payback months, NPV (3 years), IRR
• Sensitivity analysis: utilization %, productivity %, business outcome %

Sector ROI Expectations

Sector	Typical Hard ROI	Payback	Priority Use Case
Banking	150-300% (3 years)	10-14 months	Customer service RAG, fraud detection, internal copilot
Retail	100-250%	9-14 months	Product search RAG, personalization, call center
Manufacturing	80-180%	12-18 months	Predictive maintenance, QC, supply chain
Healthcare	120-200%	12-20 months	Clinical decision support, documentation
Professional services	200-400%	6-12 months	Document analysis, research, contracts
Telecom	150-250%	10-14 months	Network optimization, call center, churn

KVKK + BDDK — Cost or Multiplier?

Short answer: if designed correctly, multiplier. Turkish companies treat KVKK compliance as cost; in ROI math, KVKK penalty risk (up to €20M) is a potential loss to be modeled. Compliant design reduces it to zero = +€2-20M risk adjustment.

Talent Cost

Senior AI engineer in Turkey: $4,000-8,000/month (full-loaded). A 6-12 person internal team × 12-18 months = $400K-1.2M. Must be in ROI math.

Vendor Ecosystem

KVKK-compliant vendors are limited. 15-25% of budget goes to vendor + licensing. Under-counted ROI math is unrealistic.

Turkish ROI Maturity Levels

• Level 0 (~38%): No measurement. "Feels good."
• Level 1 (~34%): Vanity metrics. Tokens, logins, signups.
• Level 2 (~18%): Layer 2 productivity measured, no A/B test.
• Level 3 (~8%): Layer 1 + 2 + 3 measured together.
• Level 4 (~2%): Real-time AI ROI on CFO dashboard.

Target: Move from Level 0/1 to Level 2/3 within 12 months.

Case 1 — Turkish Retail Group: +23% Conversion

Problem. 8,000-SKU online catalog, customers issue unstructured queries; classic filters fail; conversion suffers.

Approach. Self-query RAG (LLM decomposes query into metadata filter + semantic search). Embedding: jina-v3 multilingual + Turkish e-commerce fine-tune. L1: 80K queries/day at month 8. L2: 1.4 sessions/customer (was 4.2). L3: conversion +23%, AOV +12%.

ROI Math. Investment: $310K dev + $48K/year ops. Hard ROI: $1.4M/year additional revenue, $48K/year vendor cost cut. Payback: 11 months. Soft ROI: +11 NPS points.

Key Decision. 70% of budget allocated to change management: product team retrained, taxonomy redesigned, content writing supported by prompt library — full BCG 10-20-70 alignment.

Case 2 — Turkish Bank (Top 5): NPS +12, AHT 12 min → 3 min

Problem. 6,000-agent call center, 8-15 min query research time. Weekly catalog, campaign, and regulation refresh.

Approach. Hybrid RAG (BGE-M3 + Qdrant on-prem + BM25). 50 chunks retrieved → BGE reranker → top-5 → GPT-5 EU instance. PII anonymization (KVKK). Eval harness: 500 questions, RAGAS faithfulness.

Results. L1: MAU 6K agents, D30 retention 78%. L2: AHT 12→3 min (-75%). L3: call resolution +18%, NPS +12, customer effort -28%.

ROI Math. Investment: $880K dev + eval + KVKK audit, $180K/year ops. Hard ROI: deferred hiring saves $1.8M/year. Payback: 9 months. Soft ROI: NPS +12 = $4M/year proxy.

Key Decision. 14-week change management program: agent training, prompt library, "AI buddy" mentorship, KPI shifted from AHT to quality + customer satisfaction.

Risk-Adjusted NPV

Speak CFO language: don't report expected NPV; report risk-adjusted NPV. Scenarios:

• Best case (20%): Layer 3 exceeded, ROI 250%.
• Base case (50%): Targets met, ROI 120%.
• Worst case (30%): Layer 2 holds but Layer 3 weak, ROI 30%.

Risk-adjusted ROI = 0.2 × 250 + 0.5 × 120 + 0.3 × 30 = 119%. Far more credible to a board than the 250% headline.

To set up the AI ROI measurement framework in your company:

1. ROI Diagnostic. Layer 1/2/3 measurement of your existing AI portfolio, BCG 10-20-70 alignment audit, lost-ROI identification. 3-week deep dive.
2. Use Case Prioritization Workshop. Map all potential use cases on the impact × feasibility matrix; detailed ROI projection for top 5. 4-hour exec workshop + 2-week analysis.
3. CFO Dashboard Design. Real-time AI ROI dashboard for CFO. KPI definitions + reporting cadence. 6-week implementation.

Reach out via the contact form on the site.

References

1. Closing the AI Impact Gap (Widening AI Value Gap) — Boston Consulting Group, BCG · 2025-01-15
2. Scaling AI Pays Off — How Leaders Capture Value — Boston Consulting Group, BCG · 2024-11-10
3. The State of AI 2025 — McKinsey & Company, McKinsey QuantumBlack · 2025-03-12
4. The Economic Potential of Generative AI — McKinsey & Company, McKinsey Digital · 2023-06-14
5. IBM Institute for Business Value — AI ROI Report 2025 — IBM IBV, IBM · 2025-04-22
6. Masterofcode — AI ROI Calculator and Framework — Masterofcode, Masterofcode Global · 2025-02-08
7. DeepHumanX — Measuring AI Business Value — DeepHumanX, DeepHumanX · 2025-05-15
8. MIT NANDA Initiative — GenAI Value Realization Study — MIT Media Lab, MIT · 2025-06-01
9. Gartner AI Maturity Model 2025 — Gartner, Gartner · 2025-05-20
10. Deloitte State of Generative AI in the Enterprise Q4 2024 — Deloitte, Deloitte · 2024-12-10
11. PwC AI Predictions 2026 — PwC, PwC · 2025-11-04
12. HBR — How to Measure AI ROI — Harvard Business Review, HBR · 2024-07-22
13. Accenture Technology Vision 2025 — Accenture, Accenture · 2025-01-30
14. Forrester AI Investment Benchmarks 2025 — Forrester, Forrester · 2025-03-05
15. BCG — Where''s the Value in AI? — Boston Consulting Group, BCG · 2024-08-12
16. Databricks State of Data + AI 2025 — Databricks, Databricks · 2025-04-18
17. Andreessen Horowitz — Enterprise AI Spend Survey 2025 — a16z, Andreessen Horowitz · 2025-05-22
18. World Economic Forum — Future of Jobs Report 2025 — WEF, World Economic Forum · 2025-01-08
19. TÜBİTAK BİLGEM Türkiye AI Maturity Report — TÜBİTAK BİLGEM, Republic of Türkiye TÜBİTAK · 2024-12
20. TRAI Türkiye AI Initiative — Sector Report 2025 — TRAI, Türkiye AI Initiative · 2025-06
21. Stanford HAI — AI Index Report 2025 — Stanford HAI, Stanford University · 2025-04
22. KPMG — Generative AI Risk and Value Survey 2025 — KPMG, KPMG · 2025-05
23. EY — How AI Will Reshape the Enterprise 2025 — EY, Ernst & Young · 2025-03
24. BCG — AI at Scale Survey 2024 — Boston Consulting Group, BCG · 2024-11
25. KVKK - Law No. 6698 on Protection of Personal Data — Republic of Türkiye - KVKK, Republic of Türkiye · 2016-04-07
26. EU Artificial Intelligence Act — European Commission, EU · 2024-03-13

---

This is a living document; BCG, McKinsey, and PwC reports refresh each quarter, so it is updated quarterly.