← Agent Frameworks 🕐 20 min read
Agent Frameworks

AI Cost ROI Measurement and Business Case Frameworks for Enterprise (2026)

Executive Summary · March 2026

Most enterprise AI business cases fail before they reach the board — not because AI lacks value, but because finance teams apply the wrong measurement model.

Most enterprise AI business cases fail before they reach the board — not because AI lacks value, but because finance teams apply the wrong measurement model. Standard NPV calculations were designed for capital equipment and software licenses. AI investments do not fit that shape. This note provides a CFO-grade measurement framework: why traditional ROI fails for AI, a three-horizon value model, specific calculation formulas with real benchmarks, full total cost of ownership with validated multipliers, and a one-page board scorecard template.

The central finding from Deloitte’s 2025 survey of 1,854 executives: the median AI investment produces satisfactory ROI in two to four years — against a technology investment norm of seven to twelve months. That gap is not a sign of poor investments. It is a sign that standard payback-period models are measuring the wrong thing at the wrong time.

1. Why Standard ROI Frameworks Fail for AI

1.1 The Payback Period Mismatch

Traditional technology ROI models assume a primary investment followed by a steady-state productivity improvement. AI does not work that way. Value accumulates non-linearly: slow in the first six months as adoption builds, accelerating through months twelve to thirty-six as workflows are redesigned, and compounding beyond that as data flywheel effects and cross-functional integrations accrue.

Deloitte’s 2025 survey found that only 6% of enterprises report payback in under a year. The median is two to four years. This does not mean the investment is poor — an investment that returns a 180% NPV over three years outperforms most technology purchases that return 60% over eighteen months. The problem is that boards apply the wrong benchmark and kill good investments too early.

CFO model fix: Replace payback-period analysis with discounted cash flow over a 36-month horizon. Discount rate should reflect the AI program’s risk profile, not the firm’s standard WACC — AI has higher optionality value and should be modeled closer to an R&D investment than a SaaS license.

1.2 The Attribution Problem

Productivity gains from AI are frequently diffuse. A customer success rep who uses an AI assistant closes tickets 30% faster — but that improvement is rarely captured in a single system as an attributable AI outcome. It shows up, if at all, as a CSAT uptick, a reduced headcount request, or an analyst observation six quarters later.

This is the measurement gap McKinsey identified in their 2025 State of AI report: 88% of organizations report regular AI use in at least one function, but only 39% can link any EBIT impact to AI at the enterprise level. The problem is not that AI is not creating value — it is that the measurement infrastructure does not exist to capture it.

CFO model fix: Establish a pre-deployment baseline with a control group methodology wherever possible. This is standard A/B test discipline applied to internal operations. Without a baseline, any savings claimed are unfalsifiable — which makes them useless in a board conversation.

1.3 Displacement vs. Augmentation: Two Different ROI Models

AI value arrives in two fundamentally different forms, and conflating them produces misleading business cases.

Displacement AI removes workflow steps entirely: automated invoice processing, contract review triage, code generation that replaces manual writing. The ROI model is straightforward — hours eliminated × loaded FTE cost. The risk is overstatement: displacement rarely reaches the theoretical maximum because workflows have more interdependencies than process maps show, and quality review still requires human oversight.

Augmentation AI makes humans faster without removing steps: a developer who writes code 55% faster is still writing code; a lawyer who reads contracts 40% faster still reads contracts. The ROI model is capacity expansion, not headcount elimination. A team of ten doing the work of fifteen with the same quality is real value, but it only converts to financial return if the organization actually deploys that freed capacity to revenue-generating or cost-reducing work.

CFO model fix: Build separate models for each type. Displacement AI: use a realized-savings model (hours × cost) with a realization rate assumption of 60–75% (not 100%) to account for overhead, ramp time, and quality assurance. Augmentation AI: use a capacity redeployment model that projects what the freed capacity will produce — and then track whether leadership actually redirects it.

1.4 Network Effects and the Data Flywheel

AI systems improve as data accumulates. A customer service AI trained on six months of tickets performs measurably better than it did at launch. A code review AI that has seen a firm’s full codebase catches more issues than one deployed on a greenfield project. This means a six-month evaluation significantly understates a system’s three-year value.

Klarna’s AI customer service evolution illustrates this directly: at launch it handled the equivalent of 700 FTE; by Q3 2025 that figure had grown to 853 FTE-equivalent capacity at the same infrastructure cost, as the system learned from accumulated interactions. The per-unit economics improved over time, not despite maturity but because of it.

CFO model fix: Include a data flywheel factor in multi-year projections. A conservative model applies a 5–10% annual efficiency improvement on AI-handled volume in years two and three. A moderate model uses 10–15%. Do not project flywheel benefits beyond year three without evidence specific to the deployment.

1.5 Option Value: Platform Investments vs. Point Solutions

AI infrastructure investments — a vector database, an enterprise LLM API contract, a unified observability stack — do not produce a single defined return. They create a platform from which future use cases can be deployed at near-zero marginal infrastructure cost. A firm that builds a solid retrieval-augmented generation (RAG) infrastructure for one use case can deploy a second, third, and fourth use case at 20–30% of the first use case’s implementation cost.

This option value is systematically excluded from point-solution ROI models, which makes platform investments look expensive relative to single-application purchases.

CFO model fix: For platform investments, add a real options component to the model. Estimate the cost of the next two to three anticipated use cases if deployed on existing infrastructure vs. built from scratch. Apply a probability weight (50–70%) to each. This approach is standard in capital allocation for R&D platforms and is auditor-defensible.

2. The Three-Horizon ROI Framework for AI

The single most common failure in AI board presentations is building a business case entirely around Horizon 1 returns. This makes the investment look small, low-conviction, and easily deferrable. The three-horizon framework forces completeness.

Horizon 1 (0–12 months): Cost Reduction and Efficiency Gains

What it contains: FTE-equivalent productivity gains, call deflection, error rate reduction, manual process automation, license cost consolidation. These are measurable, near-term, and defensible.

Typical weight in board deck: 40–50% of projected three-year value, but it is the primary evidence that the program is working. Horizon 1 metrics are your proof-of-concept evidence.

Key risk: Over-weighting Horizon 1 creates a false precision. A business case built only on measurable near-term savings will understate value by at least 40–60% relative to full three-year NPV. It also creates a perverse incentive: teams optimize for measurable near-term wins rather than transformative deployments.

Measurement cadence: Monthly. Track adoption rate, tasks automated, hours saved, error rates. These are your leading indicators.

Horizon 2 (12–36 months): Revenue Enhancement and Competitive Moat

What it contains: Conversion rate improvement from AI-assisted sales, time-to-close reduction, customer lifetime value improvement from AI-driven personalization, new product features enabled by AI capabilities, competitive pricing power from operating cost advantages.

Typical weight in board deck: 35–45% of projected three-year value. This is where most of the financial magnitude sits, but it requires cleaner attribution methodology because AI is one of several variables driving revenue outcomes.

Key risk: Overclaiming attribution. If your sales AI is deployed alongside a headcount expansion and a new pricing strategy, attributing revenue growth cleanly to AI requires an A/B test design at the territory or account-type level — not a post-hoc correlation.

Measurement cadence: Quarterly. Revenue per sales rep assisted vs. not-assisted, customer retention rate in AI-personalized segments vs. control, NPS in AI-handled customer journeys.

Horizon 3 (36+ months): Transformative Business Model Change

What it contains: New revenue streams enabled by AI capabilities (products that could not exist without AI), structural competitive advantages from proprietary data assets, business model transformation (moving from services to software, from reactive to predictive operations), and workforce model redesign.

Typical weight in board deck: 10–20%, and it should be presented as scenario analysis, not a point estimate. A 5% probability of a transformative outcome has real expected value, and boards need to see it even if it cannot be precisely quantified.

Key risk: Either ignoring Horizon 3 entirely (leaving value on the table in the board conversation) or presenting it as if it were certain (destroying credibility). The right approach is explicit scenario framing with probability weights.

Measurement cadence: Annual. Leading indicators are new product revenue share, proprietary data asset growth, competitive win rates against AI-native competitors.

How to Weight Horizons in a Board Presentation

Present the three-year NPV as the headline number. Decompose it explicitly by horizon. Show the Horizon 1 number as evidence the program is working, Horizon 2 as the primary value driver, and Horizon 3 as the option value that justifies platform investment over point-solution procurement.

A common board-ready framing: “The conservative three-year NPV assuming only Horizon 1 and Horizon 2 returns is $X. Horizon 3 optionality adds a further $Y in expected value under our base scenario. Even the conservative case clears our hurdle rate.”

3. Cost Reduction ROI: Formulas and Benchmarks

3.1 FTE-Equivalent Productivity Gain

Formula:

Annual Value = (Hours Saved Per FTE Per Year) × (# FTEs Using AI) × (Loaded FTE Cost) × (Realization Rate)

Variables:

  • Loaded FTE cost = salary + benefits + overhead (typically 1.25–1.40× base salary)
  • Realization rate = 60–75% for augmentation AI; 80–90% for displacement AI with fully automated handoffs

Benchmarks:

  • GitHub Copilot (Microsoft 2024 controlled study, n=95): developers completed tasks 55% faster; average task time dropped from 2 hours 41 minutes to 1 hour 11 minutes. At ANZ Bank (n=not disclosed), a six-week study showed 42% faster task completion with variation by skill level: beginners 52%, intermediates 42%, senior developers 40%.
  • Forrester 2025 (Microsoft 365 Copilot): 108 additional productive hours per user per year — a figure corresponding to approximately 5% of annual available working time per knowledge worker.
  • IBM developer productivity data: enterprise studies show 25–40% productivity improvement on well-scoped tasks, declining to 10–15% on ambiguous or novel problem types.

Example calculation: An engineering team of 100 developers at $180,000 loaded cost uses AI coding assistance. Productivity gain: 40% on roughly 30% of tasks (constrained to AI-compatible work). Hours saved: 0.40 × 0.30 × 2,000 hours/year = 240 hours/developer/year. Value: 100 × 240 × $90/hour × 0.70 realization rate = $1.51M annual value.

3.2 Call Deflection Value

Formula:

Annual Value = (Deflected Contacts Per Year) × (Cost Per Handled Contact) × (Quality Adjustment)

Variables:

  • Cost per handled contact in enterprise service centers: $8–$25 depending on channel and complexity
  • Quality adjustment: 0.85–0.95 to account for escalations and re-contacts from imperfect deflection

Benchmarks:

  • Klarna (2025): AI handles volume equivalent to 853 FTE at a per-transaction cost of $0.19, down from $0.32 in Q1 2023 — a 40% reduction in cost per transaction over two years. Annual cost avoidance: approximately $60M by Q3 2025.
  • Enterprise service desk benchmarks: well-implemented AI deflection achieves 25–40% ticket deflection within six months of deployment for Tier 1 queries.

Caveats for Klarna benchmark: The savings represent cost avoidance during growth (avoided hires, not layoffs), and Klarna began reintroducing human agents in mid-2025 due to quality issues on complex interactions. The Klarna case demonstrates both the ceiling and the floor of AI-only customer service strategies.

3.3 Error Rate Reduction Value

Formula:

Annual Value = (Errors Prevented Per Year) × (Average Cost Per Error)

Variables:

  • Average cost per error depends heavily on domain: a compliance error in financial services can cost $50K–$500K+ in remediation; a data entry error in order management averages $300–$800 in rework cost; a software defect caught in code review saves approximately $1,500–$7,600 versus catching it in production (NIST data).

Benchmarks:

  • AI-assisted code review: 20–35% reduction in defect escape rate to production in enterprise deployments with structured CI/CD integration.
  • AI-assisted contract review: law firms and legal ops teams report 15–30% reduction in missed clause errors on routine contract types.
  • AI-assisted data entry validation: 40–60% reduction in downstream error correction workload in structured data workflows.

3.4 Time-to-Market Acceleration Value

Formula:

Annual Value = (Revenue Per Day In-Market) × (Days Accelerated) × (# Products or Releases)

This metric is most applicable in software and pharmaceutical contexts where days to market have direct revenue implications. In software, for a product generating $10M ARR, each month of acceleration is worth approximately $830K in NPV-adjusted revenue. McKinsey analysis suggests AI-assisted software development can accelerate release cycles by 15–25% for well-instrumented teams.

4. Revenue Impact Measurement

4.1 Conversion Rate Lift Attribution

The cleanest method for attributing AI-driven revenue improvement is a prospective A/B test. Assign accounts, territories, or deal types randomly to AI-assisted and non-AI-assisted conditions. Measure conversion rate and average contract value over a sixty to ninety-day window.

Methodology for auditor scrutiny:

  • Document random assignment procedure before deployment
  • Ensure sample sizes are sufficient for statistical significance (minimum 200 deals per arm for a 5% lift to reach 80% power)
  • Control for deal size, industry vertical, and rep tenure as covariates
  • Report intent-to-treat results (assigned to AI condition) alongside as-treated results (actually used AI tool)

Benchmarks: AI-assisted sales tools show conversion rate improvements of 5–15% in controlled enterprise studies, with higher lift in early-stage qualification (AI surfaces intent signals) and mid-funnel follow-up cadence (AI personalization).

4.2 Customer Lifetime Value Improvement

AI personalization affects CLV through two mechanisms: retention improvement (reduced churn) and expansion revenue (higher product attach rates).

Formula:

CLV Delta = (Churn Rate Reduction × Average Revenue Per Account × Margin) + (Expansion Rate Improvement × Average Expansion ACV)

A 1-point improvement in net revenue retention (e.g., 108% to 109% NRR) at a $100M ARR company is worth $1M annually, compounding. AI-driven retention programs show 1–3-point NRR improvements in documented enterprise deployments.

4.3 Time-to-Close Reduction in Sales

AI tools that surface relevant case studies, auto-draft proposals, and flag at-risk deals based on engagement signals reduce average sales cycle length. A 10% reduction in sales cycle length in a business with 90-day average cycles releases 9 days of working capital per deal and allows the same sales force to handle proportionally more pipeline.

Model: (Deals Won Per Year × Days Accelerated × (ACV / 365)) × (WACC / 365) = working capital value of acceleration. At $500K ACV, 10 days faster close, 100 deals/year, 10% WACC: $500K × (10/365) × 0.10 × 100 = $137K working capital value — small, but additive to the conversion rate and CLV story.

4.4 Option Value Calculation for AI-Enabled Features

For AI investments that enable product features that could not exist without AI capabilities (predictive pricing, anomaly detection, automated report generation), value the optionality using a real options framework.

Simplified approach: Estimate the revenue potential of the AI-enabled feature × probability of achieving product-market fit × discounted at 20–30% (venture-style hurdle for novel product features). This is less precise than DCF but captures value that pure-cost models exclude entirely.

5. Total Cost of AI Ownership

5.1 The 3x Implementation Rule

Forrester’s analysis of enterprise Microsoft 365 Copilot deployments established the foundational ratio for enterprise AI TCO: $1 in annual license cost corresponds to approximately $3 in total implementation and change management cost over the first two years.

This ratio holds across multiple enterprise AI deployment studies and is the single most commonly violated assumption in AI business cases. A $5M annual model API + tooling cost implies $15M in implementation overhead — not $5M.

TCO Component Breakdown:

Component Typical % of Total 2-Year Cost
Software licenses / API costs 25–30%
Implementation (integration, prompt engineering, testing) 30–35%
Change management (training, process redesign, communication) 15–20%
Governance and compliance (policy, monitoring, audit) 8–12%
Ongoing model management and retraining 8–10%
Shadow AI remediation 3–7%

5.2 Shadow AI Remediation Cost

IBM’s 2025 Cost of a Data Breach report found that breaches involving shadow AI cost organizations $4.63M on average — $670K more than standard incidents. Shadow AI incidents represent 20% of all enterprise data breaches in 2025.

A firm deploying enterprise AI without a shadow AI audit and governance program is not avoiding shadow AI costs — it is deferring them. The median remediation timeline after a shadow AI-related incident is 94 days. CFOs building an AI TCO model should include a shadow AI governance program in year-one costs, estimated at $150K–$500K depending on organization size, as insurance against a $4.63M expected loss.

5.3 Change Management as a First-Class Cost

The most common TCO underestimate is change management. AI tools that are not adopted do not generate returns. Adoption requires:

  • End-user training (typically 8–16 hours per knowledge worker, plus refreshers as models update)
  • Process redesign (identifying which workflows benefit from AI vs. which create AI-human friction)
  • Manager enablement (supervisors who do not understand AI outputs cannot validate them)
  • Communication programs (employees who fear job displacement resist using AI)

An enterprise deploying AI to 5,000 knowledge workers at $180K loaded cost, requiring 12 hours of training + process redesign, faces a direct change management cost of $1.5M+ before accounting for the indirect productivity dip during the transition period (typically 4–8 weeks at 10–15% reduced throughput).

5.4 Technical Debt from Rushed Implementations

A secondary cost that rarely appears in initial business cases is the technical debt created by fast-follow AI deployments that lack architectural discipline. Common patterns:

  • Point integrations hardcoded to specific model versions that break on API deprecation
  • Prompt engineering that is undocumented and cannot be maintained after the original team moves on
  • Evaluation frameworks that were never built, leaving the organization unable to detect model quality regressions
  • Data pipelines built for the initial use case that become bottlenecks when use cases scale

Gartner estimates that enterprises with undisciplined AI implementations will spend 2–3× the initial build cost in remediation and re-architecture within 24 months. Building proper MLOps and observability infrastructure at deployment adds 15–20% to initial implementation cost but eliminates this deferred expense.

5.5 Full TCO Formula

Two-Year AI TCO = (Annual License/API Cost × 2)
               + (Implementation Cost = License × 2.5–3.0)
               + (Change Management = $1,500–$3,000 per end user)
               + (Shadow AI Governance = $150K–$500K flat)
               + (Ongoing Model Management = 8–10% of License per year)
               + (Technical Debt Reserve = 15–20% of Implementation Cost)

Apply this formula before finalizing any AI business case. A project that shows positive ROI on license + implementation but negative NPV when TCO is fully loaded is not a sound investment. A project that is positive on full TCO is genuinely attractive.

6. Board-Ready AI Investment Scorecard

The following one-page template is designed for a CFO who has five minutes and needs to answer one question: is this investment on track, and what does the board need to decide?

AI Investment Scorecard — [Program Name] — [Quarter]

Investment Summary

Metric Value
Total invested to date $X.XM
Projected 3-year NPV (base case) $X.XM
Projected 3-year NPV (conservative, 60% of base) $X.XM
Current payback period estimate X.X years
Internal rate of return (3-year) XX%

Leading Indicators (Monthly)

Indicator Target Actual Trend
Active user adoption rate 75% by M6 XX% ↑ / → / ↓
Tasks automated per week X,XXX X,XXX —
Error rate (AI-handled workflow) <X% X.X% —
Time-per-task vs. baseline –30% –XX% —

Lagging Indicators (Quarterly)

Indicator Target Actual Trend
Cost savings realized (cumulative) $X.XM $X.XM —
Revenue attributed (A/B methodology) $X.XM $X.XM —
FTE redeployment to H2/H3 activities XX FTE XX FTE —

Horizon Value Attribution (3-Year NPV Decomposition)

  • Horizon 1 (efficiency, 0–12 months): $X.XM — REALIZED: $X.XM
  • Horizon 2 (revenue, 12–36 months): $X.XM — IN PROGRESS
  • Horizon 3 (optionality, 36+ months): $X.XM — SCENARIO ESTIMATE

Risk Register

Risk Probability Financial Impact Mitigation
Regulatory (EU AI Act, state AI laws) Medium $X.XM Governance program in place
Model dependency (single provider) Medium $X.XM Multi-provider architecture roadmap
Adoption shortfall Medium –$X.XM NPV Change management program
Shadow AI incident Low-Medium $4.6M avg. Shadow AI audit completed

Board Decision Required: [None — on track] / [Additional investment to accelerate H2] / [Scope change required]

7. Benchmarks and Peer Comparisons

7.1 Using Benchmarks Correctly

Peer benchmarks provide a market context for what is achievable; they do not substitute for organization-specific measurement. A benchmark that says “AI leaders achieve 40% more cost savings than laggards” tells a CFO what the upside looks like if the program is executed well — it does not tell the CFO what their specific deployment will return.

The correct use of benchmarks in a board presentation: “Peer data from BCG’s 2025 study of 1,250 executives suggests that future-built AI companies achieve 1.6× EBIT margin relative to laggards. Our model conservatively assumes one-third of that differential over three years, reflecting our lower starting AI maturity. Even at that conservative assumption, the three-year NPV is $X.”

7.2 Validated Benchmarks (as of June 2026)

Deloitte State of GenAI (2025, n=1,854 executives, Europe and Middle East)

  • Median ROI payback period: 2–4 years
  • Only 6% of deployments achieve payback within 12 months
  • 78% of enterprises report positive ROI within two years
  • 85% increased AI investment in 2025; 91% plan to increase again
  • Headline conclusion: AI ROI is real, but the payback horizon is fundamentally different from standard technology investments

BCG AI Value Gap Report (2025, n=1,250 senior executives, 25+ sectors)

  • Future-built AI companies (top 5% globally) achieve 1.7× revenue growth, 3.6× three-year total shareholder return, and 1.6× EBIT margin vs. laggards
  • AI leaders expect 2× revenue increase and 40% greater cost reductions in AI-deployed areas vs. laggards
  • Agentic AI already accounts for 17% of total AI value in 2025; projected to reach 29% by 2028

McKinsey (GenAI Economic Potential baseline, 2023, updated 2025)

  • $2.6–$4.4 trillion annual value across 63 use cases; extended estimate including broader knowledge worker productivity: $6.1–$7.9 trillion annually
  • Only 39% of organizations can link EBIT impact to AI at enterprise level; of those, 80%+ report sub-5% enterprise-wide effect
  • Largest value pools: customer operations, marketing and sales, software engineering, R&D

GitHub Copilot (Microsoft controlled study, 2024, n=95 developers)

  • 55% faster task completion (2h41m → 1h11m)
  • Success rate improvement: 70% → 78%
  • ANZ Bank six-week study: 42% faster task completion overall; beginner developers 52% faster
  • 88% of surveyed developers report feeling more productive; 87% report reduced cognitive load on repetitive tasks

Klarna AI Customer Service (2025, public disclosures)

  • AI equivalent to 853 FTE capacity (up from 700 at launch)
  • Cost per transaction: $0.19 in Q1 2025 vs. $0.32 in Q1 2023 (40% reduction)
  • Annual cost avoidance: approximately $60M by Q3 2025
  • Response time 82% faster than pre-AI
  • Caveat: Klarna reintroduced human agents in mid-2025 for complex interactions; pure-AI customer service has a quality ceiling

IBM 2025 Cost of Data Breach Report

  • Shadow AI incidents: 20% of all breaches
  • Shadow AI breach cost premium: +$670K vs. standard breach ($4.63M average)
  • 69% of organizations suspect employees use prohibited public GenAI tools

7.3 Benchmarks to Avoid

The following figures circulate widely in AI business cases and should not be used. They are either fabricated, debunked, or stripped of context to the point of being misleading:

  • “95% of AI projects fail” — attributed to a NANDA/MIT 2025 study. No such study was published. The figure is fabricated. Its use will destroy credibility with any audience that has done due diligence.
  • “95% of pilots fail to scale” — same fabricated source.
  • “23x productivity improvement” — unverifiable, routinely misattributed, and not replicable in any published study.
  • “420% ROI” — has appeared in vendor materials with no credible primary source.
  • Any McKinsey figure attributed as a productivity multiplier for individual workers without citing the specific use case and methodology. McKinsey’s $2.6–$4.4T figure is an aggregate economic potential across the global economy, not an enterprise-level projection.

8. Common Failure Modes and How to Avoid Them

Measuring Only What Is Easy to Measure

Teams default to measuring Horizon 1 metrics (tasks automated, hours saved) because they are trackable in existing systems. Revenue attribution and option value calculations require additional instrumentation and A/B design work. The result: AI programs that are creating substantial value look underperforming because the measurement system cannot see Horizon 2 and Horizon 3 returns. Build the measurement infrastructure for all three horizons before deployment, not after.

Treating Adoption as a Lagging Indicator

Adoption rate is a leading indicator that predicts whether any financial return will materialize. A tool with 40% active adoption at month six will not produce the returns modeled at 80% adoption. Most AI programs discover adoption shortfalls too late — when the six-month savings review comes back underperforming. Track adoption weekly in the first three months and intervene before the curve flattens.

Confusing Vendor ROI Studies with Independent Evidence

Vendor-sponsored TEI (Total Economic Impact) studies — Forrester, IDC, and others commissioned by the vendor — are useful for understanding the mechanics of value capture, but their headline ROI figures are based on top-quartile customer interviews, not median deployments. Use vendor studies for formula structure and benchmark direction. Discount the headline numbers by 40–50% for your conservative planning case.

Underestimating Change Management as a Cost and a Risk

The technology is rarely the constraint on AI ROI. The change management is. A deployment to 5,000 users where 60% actively use the tool produces 60% of the modeled return — and 60% of the cost. Change management is both a line item in TCO and a risk factor in the NPV sensitivity analysis.

Sources