In this article8
AI ROI is the value an AI system creates minus what it costs to build and run, measured against how the process performed before. To calculate it, you need a baseline (volume, time per item, error rate), an honest estimate of how much the AI takes over, and the full cost picture: build, model usage, hosting, maintenance and the human review that remains. Most AI projects that fail to show ROI don't fail on the math. They fail because nobody measured the baseline first.
The short version:
- MIT found 95% of organizations saw no measurable P&L impact from GenAI pilots. Wharton found three in four leaders see positive returns. The difference is mostly in how you measure.
- Broad time savings are modest: 2.2% of all US work hours in Q2 2026, according to the St. Louis Fed.
- Real returns come from specific high-volume processes, not from giving everyone a chatbot.
- Measure your baseline before you build: volume per week, minutes per item, error rate.
What is the ROI of AI? The research disagrees
Search for AI ROI studies and you'll find numbers pointing in every direction:
| Source | Finding | What it measures |
|---|---|---|
| MIT NANDA, 2025 | 95% of organizations get no measurable P&L impact from GenAI pilots | Pilots, profit and loss effect |
| IBM CEO study, 2025 | Only 25% of AI initiatives delivered expected ROI | CEO estimates |
| Wharton, 2025 | Three in four enterprise leaders see positive GenAI returns | Self-reported, 72% measure formally |
| McKinsey, 2026 | 37% attribute some EBIT impact to AI, but only 6% are high performers | Survey of 1,719 respondents |
| St. Louis Fed, 2026 | Time saved across all work hours rose from 1.6% to 2.2% | All US workers |
These numbers don't really contradict each other. Together they say the average is low and a small group gets clear results. The MIT report names a "learning gap" as the main cause: tools that don't fit the company's processes and data. The model is rarely the problem.
The St. Louis Fed data is a useful reality check. The share of work hours in which AI helped rose to 6.3%, but time saved across all hours was 2.2%. A business case built on "everyone gets 30% more productive with ChatGPT" is built on sand. A case built on one high-volume process is much stronger.
What costs go into an AI ROI calculation?
One-off costs
- Discovery and analysis. Which process, which data, which integrations, what the baseline is.
- Build. The automation or agent itself, integrations with your systems (ERP, CRM, accounting, Microsoft 365 or Google Workspace), an admin view, logging.
- Data cleanup. Deduplicating customers, updating documentation, standardizing fields. Often forgotten.
- Rollout. Training, testing with your team, updating work instructions.
Running costs
- Model usage. You pay per token, roughly per amount of text in and out. Some examples, pricing as of September 2026 per million tokens: Claude Sonnet 5 costs $2 input and $10 output, Claude Opus 5.5 $4 and $20, GPT-6 Sol $2 and $10, GPT-6 Luna $0.10 and $0.50. For most mid-market processes, model costs are small compared to the build.
- Licenses. For per-seat tools: ChatGPT Business and Claude Team are $20 per user per month billed annually; Microsoft 365 Copilot enterprise is $30.
- Hosting and monitoring. Servers, database, logging, alerts.
- Maintenance. Models change, vendor APIs change, your processes change. Budget for ongoing care.
Hidden costs
- Human review. If someone checks every output, that time is a cost.
- Errors. What does a misbooked invoice or a wrong answer to a customer cost you?
- Runaway usage. An agent stuck in a loop can burn through tokens unnoticed. Set hard limits on usage and spend.
For build costs in general, see custom software development cost. For agents specifically: AI agent development cost.
How to measure AI ROI: the calculation model
The formula is simple. The work is in honest inputs.
- Measure your baseline. How often does the process run per month? How many minutes per run? What is the fully loaded hourly cost? How often does it go wrong and what does an error cost?
- Estimate the automation rate. What share can the AI handle fully, what share with review, what share stays manual?
- Calculate new monthly cost. Remaining manual time plus review time plus running AI costs.
- Calculate monthly savings. Baseline minus new cost.
- Calculate payback. One-off costs divided by monthly savings.
Worked example (illustrative)
Take a company that manually processes 1,200 incoming invoices a month. These are assumptions to show the model, not client figures.
| Item | Before | After |
|---|---|---|
| Invoices per month | 1,200 | 1,200 |
| Fully automated | 0% | 75% |
| With review (2 min) | 0% | 20% |
| Manual (6 min) | 100% | 5% |
| Hours per month | 120 | 14 |
| Staff cost at $55 per hour | $6,600 | $770 |
| Running AI costs (model, hosting, maintenance) | $0 | $900 |
| Total per month | $6,600 | $1,670 |
Savings: $4,930 per month. If the build including discovery costs $33,000, payback is just under seven months. For comparison, Ardent Partners puts the average cost of processing an invoice at $9.90, and $2.67 at best-in-class organizations. More on this specific process in AI invoice processing automation.
Two caveats. First, saved hours only turn into money if you redeploy them or avoid new hires. With tight labor markets, that last one is often the real argument. Second, run a pessimistic scenario too, say 50% automated instead of 75%. If it still pays off, your case is robust.
Why so many AI projects miss their ROI
- No baseline. You can't prove afterwards what it delivered.
- The wrong use case. A process that runs three times a month will never earn back a build.
- A pilot without integrations. A demo that produces nice text, but someone still retypes the result into the accounting system.
- Costs that scale with usage. McKinsey found 20% of organizations say AI costs have limited their use.
The full picture is in why AI projects fail.
What is a good ROI for an AI project?
There's no fixed number, but for process automation in a small or mid-sized business, payback within twelve months is a reasonable bar. If an honest model doesn't get there on paper, the process is probably too small or too rare. Also count non-financial gains: faster cycle times, fewer errors, less dependence on the one or two people who know the process. Harder to put in dollars, but real.
Frequently asked questions
What is a typical ROI for AI in business?
On average it is still modest: MIT found 95% of organizations saw no measurable P&L effect from GenAI pilots, while Wharton reports three in four leaders see positive returns. Companies that get results focus AI on specific high-volume processes and measure the baseline first. For process automation, payback within twelve months is a reasonable target.
How do you measure AI ROI?
Measure the current cost of the process: volume per month, minutes per item, hourly cost and error cost. Estimate what share the AI takes over, subtract remaining manual time and running AI costs, and divide the one-off costs by the monthly savings. That gives you the payback period.
Is AI making companies any profit?
For a minority, clearly yes. McKinsey's 2026 survey found 37% of respondents attribute at least some EBIT impact to AI, but only 6% qualify as AI high performers. The gap is mostly in focus, integration and measurement, not in the models.
What are the running costs of an AI automation?
Model usage per token, hosting, monitoring and maintenance. Maintenance is often underestimated, because models, vendor APIs and your own processes all change. Also set usage and spend limits so a bug in an agent can't quietly inflate your bill.
Next steps
Want to know which of your processes has the strongest business case? Our free AI scan helps you map it out. For how we build and what it costs, starting with a Discovery at EUR 3,500, see our pricing.
Sources
- Fortune: MIT report: 95% of generative AI pilots at companies are failing, August 18, 2025
- IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles, May 6, 2025
- Wharton: 2025 AI Adoption Report, October 28, 2025
- McKinsey: The state of AI in 2026, August 25, 2026
- FRED Blog (St. Louis Fed): Does generative AI save time at work?, August 27, 2026
- Ardent Partners: Accounts Payable Metrics that Matter in 2026 (via Medius), 2026
- Claude Platform Docs: Models overview, accessed September 26, 2026
- OpenAI: Introducing GPT-6 Sol and Luna, September 22, 2026
- Microsoft 365 Copilot for enterprise pricing, accessed September 26, 2026

Builds AI agents, automations and custom software for businesses at Airflows.



