Calculating ROI for AI initiatives: the framework Finance teams actually can use
- Jean Latiere
- 4 days ago
- 6 min read
What you'll learn this week: How to calculate real ROI for AI projects using the three-layer framework that separates winners from failures.
Your AI project costs $50,000 per month. Is that expensive or cheap?
This question stops most finance teams cold. Traditional cost analysis offers no useful answer because it treats AI spend as a line item rather than an investment with measurable returns. A $50,000 monthly bill represents extraordinary value if the project generates $500,000 in savings. It represents complete waste if the project delivers nothing.
The difference between these outcomes is not luck.
According to NANDA's 2025 State of AI in Business report, 95% of enterprise AI implementations deliver zero ROI. The remaining 5% extract millions in value. Note: this failure rate applies to custom enterprise systems, not productivity tools like ChatGPT where adoption is high and returns are modest but positive.
The gap between winners and failures comes down to one factor: whether organisations measure the right things.
The unit economics foundation
The first step toward meaningful ROI is abandoning aggregate cost views entirely. Knowing that your organisation spent $63,000 on AI last month tells you nothing about value. Knowing that each AI-processed invoice costs $0.10 while saving $0.50 in manual processing tells you everything.
The core question shifts from "how much did we spend?" to "how much does one useful unit cost?" A unit might be a customer conversation, a processed document, or a resolved ticket. This enables direct comparison: if your AI assistant costs $0.05 per conversation and generates $0.10 in value improvement, the 2:1 ratio signals positive ROI. If the ratio inverts, you have actionable information.
The three-layer ROI framework
Calculating AI ROI requires examining costs at three distinct layers. Each layer answers a different question, and skipping any layer produces an incomplete picture.
Layer 1: Infrastructure unit costs
The foundation layer measures raw computational costs. For most enterprises, this means inference costs: the per-token or per-call charges incurred every time the model processes a request. Training costs matter for organisations building custom models, but inference accounts for 80-90% of total AI lifetime costs, with GPT-5's projected inference bill reaching 15x its original training investment.
For language models accessed via API, infrastructure costs translate to cost per thousand tokens. Claude 3.5 Haiku costs $0.25 per million input tokens. Claude 3.5 Sonnet costs $3.00 per million. If both produce acceptable outputs for a given task, the 12:1 cost difference directly impacts ROI. A support bot using Sonnet exclusively might cost $0.025 per conversation; the same bot routing simple queries to Haiku might cost $0.008.
Layer 2: Solution unit costs (the harness)
The second layer aggregates all technical components required to deliver one unit of work. The "harness" refers to everything surrounding the model: the compute environment, orchestration logic, data retrieval systems, and connectivity. Model inference is only part of the cost.
A complete cost-per-conversation calculation might include: model inference costs ($0.015), compute costs for Lambda functions or EC2 containers running the orchestration ($0.003), API Gateway requests ($0.001), and vector database queries for retrieval-augmented generation ($0.001). The solution unit cost becomes $0.020 per conversation, the true cost of running one transaction through your AI system.
This layer reveals costs that infrastructure metrics miss. Data egress charges between regions can be higher than model costs for high-volume applications. Retry logic can multiply token consumption by 10x during error conditions. Storage costs for conversation history accumulate faster than expected. Solution unit costs expose these factors before they become surprises.
One critical caveat: unit economics stabilise only after volume absorbs fixed costs. Platform engineering, initial integration, and model fine-tuning represent upfront investments that must be amortised across transactions. Early-stage AI systems often appear unprofitable because these fixed costs dominate. The $0.020 per conversation figure assumes sufficient throughput to spread integration overhead.
Layer 3: Value-based ROI
The third layer connects costs to business outcomes. This is where most implementations fail, not from inability to calculate, but from failure to define what value means.
Value takes multiple forms.
Cost displacement replaces expensive manual processes. An offshore support agent earning $500/month fully loaded, handling 10 tickets per hour across 160 monthly hours, resolves 1,600 tickets at $0.31 each. With management overhead, the realistic cost reaches $1-2 per ticket. A support bot at $0.020 per conversation saves $0.98-1.98 per automated interaction. At 10,000 conversations monthly, annual savings reach $118,000-238,000.
Revenue generation creates new income streams. A personalised recommendation engine costing $0.03 per recommendation that increases conversion rates by 2% on $50 average orders generates $1.00 per recommendation. The 33:1 return justifies the investment directly.
Retention improvement extends customer lifetime value. An AI-powered onboarding assistant that reduces 90-day churn from 15% to 12% on customers worth $2,000 annually creates $60 additional value per customer. If the assistant costs $5 per customer to run, the return is 12:1.
Premium monetisation packages AI capabilities as paid features. A document analysis tool sold at $49/month with solution costs of $8/month generates $41 margin per subscriber. The AI becomes a profit centre rather than a cost centre.
These calculations require tracking outcomes, not just costs. A support bot ROI depends on resolution rates and escalation frequency. A recommendation engine ROI depends on actual conversion improvements. Without outcome measurement, value-based ROI remains theoretical.
Risk-adjusted ROI accounts for quality. A recommendation engine with $1.00 potential value per output but 80% accuracy delivers $0.80 realised value. Escalation rates, hallucination incidents, and rework cycles all reduce effective value. The formula becomes: value per output × success rate. This keeps ROI calculations honest and treats accuracy as a financial variable, not just a technical one.
ROI over time
Static ROI ratios mislead without a time dimension. Finance teams need to know: is this month-one ROI, steady-state ROI, or twelve-month ROI?
AI systems follow a predictable ramp.
Month one typically shows negative ROI: integration costs dominate, accuracy is low, and retries are high.
Month three approaches cost parity as prompts improve, routing optimises, and edge cases get handled.
Month six onward delivers positive ROI as learning effects compound and volume absorbs fixed costs.
This pattern explains why organisations should tolerate early losses rationally. The question is not "is ROI positive today?" but "is the trajectory toward breakeven on track?" Systems showing no improvement after 8-12 weeks warrant scrutiny. Systems showing steady weekly gains in deflection, accuracy, or cost-per-unit justify continued investment.
What the data shows
The Wharton 2025 GenAI Adoption Report surveyed approximately 800 US enterprise decision-makers.
72% of enterprises now formally measure GenAI ROI, focusing on productivity gains and incremental profit. Organisations without ROI frameworks increasingly cannot justify continued investment.
75% of leaders report positive returns, primarily from productivity tools like ChatGPT and Copilot. Enterprise-grade custom implementations, where the 95% failure rate applies, require the three-layer framework to succeed.
Approximately one-third of GenAI budgets now fund internal R&D: organisations achieving sustained ROI treat AI as a capability to develop rather than a product to purchase.
Why most implementations fail
The NANDA research identifies the root cause of the 95% failure rate: most GenAI systems do not retain feedback, adapt to context, or improve over time. These static systems function as tools rather than learning systems. Tools deliver one-time efficiency gains. Learning systems compound value.
The differentiator is learning capability. Learning means specifically: persistent memory across sessions, feedback loops that improve outputs, prompt or model adaptation based on usage, and workflow optimisation from observed patterns. Systems with these capabilities cross the GenAI Divide into positive ROI territory. Systems that reset with each interaction remain stuck with zero returns.
Leading indicators signal ROI collapse before it happens. Cost per unit trending upward faster than volume growth indicates efficiency degradation. Declining deflection rates suggest users are abandoning the system. Increasing context sizes or retry rates without outcome improvement reveal hidden waste. Manual overrides reappearing mean the system has lost user trust. Track these metrics weekly, not quarterly.
A practical decision framework
Before approving any AI initiative, finance teams should require answers to four questions.
First, what is the infrastructure unit cost? If the team cannot specify cost per token or GPU-hour, the project lacks fundamental visibility.
Second, what is the solution unit cost? If the team cannot aggregate all components into a cost-per-transaction figure, hidden costs will erode returns.
Third, what is the value per outcome? If the team cannot quantify what each successful output saves or generates, ROI calculation has no basis.
Fourth, does the system learn? If the implementation resets with each interaction, expect the 95% failure pattern.
Projects that answer all four questions with specific numbers warrant investment.
Ownership determines accountability. Layer 1 optimisation belongs to Engineering and Platform teams. Layer 2 management sits with Product and FinOps jointly. Layer 3 value realisation is the Business owner's responsibility, tied to P&L. Without clear ownership at each layer, "everyone agrees, nobody owns it" becomes the default failure mode.
Moving forward
The transition from AI experimentation to AI accountability defines 2026. The framework itself is not complex: infrastructure costs aggregate into solution costs, solution costs compare against business value, value adjusted for quality and time produces ROI. The complexity lies in implementation: tagging workloads correctly, capturing all cost components, defining value metrics, tracking outcomes consistently, and assigning clear ownership at each layer.
The 5% of organisations achieving substantial returns share one characteristic. They treat AI investments with the same financial rigour applied to any other capital allocation. Cost per unit plus value per unit equals ROI transparency. Everything else is speculation.
Model your own AI initiatives with our ROI Calculator → Input unit economics across three layers, scenario model volume changes, and track time-to-breakeven.
Check our AI ROI Calculator: https://airoicalculator.optimnow.io/ or check the github project and adjust the model according to your specific needs https://github.com/OptimNow/ai-roi-calculator/
.png)