AI ROI Calculator for Businesses

The model, in full

The calculator shows every assumption it uses. The work ranges, leak weights, and sources below come from the same module that runs the tool.

In the benchmark used here, the average firm converted 41% of AI time savings into measurable business value. This calculator starts with task studies and a labeled admin estimate, then estimates where some of that time may be lost.

Why this calculator refuses to give you an ROI percentage

Four studies produced four different headlines. MIT's NANDA project said 95% of generative AI pilots returned nothing. Wharton said 75% of firms reported a positive return. A Microsoft-commissioned IDC study reported $3.70 back for each dollar invested. McKinsey found that 39% of organizations saw any profit effect.

These studies did not measure the same thing. They used different samples, questions, and definitions of a return. Choosing one headline would hide that problem.

So I start with what the studies measured at the task level, then estimate how much of that gain could survive the way your firm runs the work. A single percentage would imply more certainty than the evidence gives us.

The gains that are real

Controlled studies show that AI can make some tasks faster. They do not show that the whole business becomes faster. The model starts with the task, because that is where the evidence is strongest.

Two rows need care. Structured production covers both scoped tasks that became much faster and complex work where experienced people became slower. Move that slider down when the work needs deep context or judgment. The admin range is an estimate, not a controlled finding.

The eight delivery presets are starting points, not findings. Change the sliders until the mix looks like a normal week in your firm.

Where the value leaks

A task gain still has to pass through review, approval, scheduling, pricing, and the firm's knowledge. The studies support those mechanisms. I chose provisional weights for those problems, the studies did not measure how much loss each one caused. The weights divide the estimated loss so you have a place to start checking, not a diagnosis to accept without testing.

Validation (22% of the leak model)

A fast draft still needs checking. BetterUp Labs and Stanford found that 40% of desk workers received AI output that needed about two hours of rework. Validation receives 22% because review can touch every AI-assisted deliverable. That weight is an allocation, not a measured share. Decide which outputs need a full review and which need a spot-check. That work is part of Proofwork.

Approval chain (24% of the leak model)

Faster production can move the delay into review. Faros measured teams merging 98% more pull requests while review time rose 91%. Approval carries the heaviest weight, 24%, because every deliverable may still pass through that queue. The allocation is model judgment, not a percentage Faros measured. The Production Gap helps find the new delay.

Reallocation (22% of the leak model)

Suppose a three-hour draft now takes 40 minutes. The firm still has to decide what receives the time it recovered. Freed hours do not choose their next job. The Optimum Partners benchmark found that much of the measured time gain did not reach the business. Reallocation receives 22% of the loss because unused capacity can erase a gain even when the task became faster. The exact weight is an allocation. Name where the hours will go before they arrive.

The Four Numbers asks the early question: what share of freed capacity is booked to something named? The answer should point to more delivery, better delivery, or a changed offer.

Pricing (18% of the leak model)

When a firm sells hours, faster work can mean fewer hours billed. The ACC and Everlaw found that about 60% of in-house counsel saw no savings from their law firms' AI use, while 13% reported fewer billable hours. Pricing receives 18% because time-billed work can hand the gain to the client. This is an allocation, not a measured share of firm-wide loss. Check fixed-fee, retainer, and outcome work separately.

Context (14% of the leak model)

AI cannot use knowledge it cannot reach. Client history, past work, and house rules often sit across inboxes, shared drives, and people's heads. That missing context can push people toward shadow AI. Context receives the smallest weight, 14%, because this model treats it mainly as a limit on usable output. That allocation is model judgment, not a measured causal share.

How the math works

Each answer scores 0, 0.5, or 1. The model multiplies the answer by its leak weight, then adds the five results.

Pass-through starts at 0.27 and moves toward 0.71 as the score improves. The measured average is 41%. The top 7% in the same benchmark reached 71%.

The floor is also a judgment call in this model. 0.27 is not a measured anchor. It estimates a firm with none of the five practices in place.

The answer is always a range, never a single point. The midpoint carries 6 points on either side, bounded by 15 and 80. The annual figure uses 46 working weeks and stays labeled capacity value, not profit. Freed hours become margin only when the firm gives them a useful job.

What the timeline depends on

The range I have observed is 6 to 18 months, sometimes longer. This is not a forecast. Deloitte's survey of 1,854 executives also found that only 6% of organizations saw ROI inside a year.

Three conditions separate the shorter and longer cases: a named owner, learning time on the calendar, and a short policy that says what people may safely test. The calculator uses them as planning signals, not as proof that a result will arrive in a given month.

What to do with your result

Start with the first leak the result names. Write down one change, one owner, and the business number that should move. Use The Delivery Model Ladder to see how large the change needs to be.

• Validation: define what gets checked and by whom.
• Approval: find the queue that grew after production got faster.
• Reallocation: give freed hours a named use.
• Pricing: check whether the firm or the client keeps the gain.
• Context: put approved firm knowledge where the tool can reach it.

Then track cost to acquire, cost to deliver, retention, or price. Use margin and revenue per person to confirm that the change reached the business. This works whether or not an AI workflow is involved, and it is how a firm moves toward AI-native delivery. That is how you check whether the change helped the business.

Sources

External sources used by the model, in citation order.

1. BCG consultants RCT (N=758), Organization Science 2026
2. Noy & Zhang writing-tasks RCT (N=453), Science 2023
3. Brynjolfsson et al. support-agent study (N=5,172), QJE 2025
4. GitHub Copilot productivity RCT
5. METR developer RCT (2025)
6. BetterUp Labs / Stanford workslop study (September 2025)
7. Faros AI software-engineering benchmark
8. Optimum Partners benchmark (N=255), May 2026
9. ACC / Everlaw survey of approximately 650 in-house counsel
10. MIT NANDA, The GenAI Divide report
11. Deloitte's AI ROI survey (N=1,854 executives)