
Key takeaways
- Measure the task before you automate it. Time per item, volume, error rate and the hourly cost of the people doing it.
- Count all the costs: build, licences, model usage, maintenance and the staff time to review outputs.
- MIT research reported in 2025 found most generative AI pilots stalled, while the strongest returns came from back-office automation.
- Set a payback target and a review date up front, and be willing to stop if the numbers don’t show up.
Why most AI projects never prove their value
Plenty of AI projects don’t fail technically. They fail because nobody can say whether they worked. In 2025, MIT’s NANDA initiative published The GenAI Divide: State of AI in Business 2025, based on 150 interviews with leaders, a survey of 350 employees and an analysis of 300 public deployments. As Computing reported, it found 95% of corporate generative AI pilots stalled and never reached scaled adoption.
Treat that figure with some care. Coverage doesn’t explain exactly how “success” was defined, and The Stack called the report preliminary and some headlines overhyped. The more useful finding is where returns did show up: back-office work such as process automation, reduced outsourcing and operational efficiency, even though more than half of budgets went to sales and marketing.
Step 1: Measure the task as it is today
Pick one process and write down five numbers before anyone builds anything:
- Minutes per item. How long one invoice, enquiry or report takes a person.
- Volume. How many a day, week or month.
- Hourly cost. Salary plus overheads for the people doing it.
- Error or rework rate. How often it has to be fixed or redone.
- Delay cost. What waiting costs you, such as slow quotes losing deals.
Monthly hours = minutes per item × items per month ÷ 60. Multiply by the hourly cost and you have what the task costs now. Our free automation savings calculator does this sum for you. If you’re not sure which task to measure, start with our guide on which processes to automate first.
Step 2: Be realistic about what the automation will handle
Automation rarely covers 100% of a task. Some items will be odd, some will need a judgement call, and some outputs will need a quick human check. Estimate a share, say 60% or 70%, based on how many items follow the normal pattern, and test that estimate in a pilot.
Don’t forget the review time. If a person spends 30 seconds checking each AI-drafted reply, that time comes off the savings.
Step 3: Count every cost
| Cost | One-off or ongoing | Often missed? |
|---|---|---|
| Design and build | One-off | Rarely |
| Software licences and integrations | Ongoing | Sometimes |
| AI model usage (per request or token) | Ongoing, grows with volume | Often |
| Monitoring, fixes and updates | Ongoing | Often |
| Staff time for review and exceptions | Ongoing | Very often |
| Training and change management | Mostly one-off | Often |
For AI agents specifically, our AI agent development cost guide breaks down build and running costs in more detail.
Step 4: Work out payback, then measure again
Payback in months = one-off costs ÷ (monthly savings − monthly running costs). Here’s a made-up example to show the maths. A team spends 10 minutes on each of 1,200 enquiries a month: 200 hours. At $30 an hour that’s $6,000 a month. If automation handles 60% of them, that’s about $3,600 saved. Take off $900 a month for running costs and review time, and you’re left with $2,700. A $10,000 build would pay back in under four months.
Write the target down and pick a review date, usually 60 to 90 days after launch. Then measure the same five numbers again. If the gain isn’t there, find out why before you scale. Sometimes the fix is small; sometimes the right call is to stop.
Our view: a clear, modest win on one process beats an ambitious project nobody can measure. That’s how we scope work in our AI automation services and AI strategy consulting. If you’d like help putting numbers on an idea, book a strategy call.
Frequently asked questions
How do you calculate ROI for AI automation?
What is a good payback period for an automation project?
What costs do people forget when estimating AI ROI?
Did MIT find that 95% of AI projects fail?
Sources
- Computing: MIT report finds 95% of corporate generative AI pilots fail to deliver returns
- The Stack: What CIOs can learn from MIT’s AI report, beyond the doomsday overhype (25 Aug 2025)