An AI ROI Calculator: What to Count and What to Leave Out

An AI ROI calculator needs four inputs per workflow: the full cost, a baseline from before AI, the change in a number the business already tracks, and the value of that change. Leave out self-reported hours, vendor benchmarks and capacity nobody redeployed. What remains is smaller than most calculators show, and it holds up.

What should an AI ROI calculator include?

A useful AI ROI calculator works one workflow at a time. For each, it holds the full cost, a baseline measure from before AI, the same measure afterwards, and a rule for turning the difference into money. It also records who owns each number, so the result can be checked rather than argued about.

The value rule is the input most calculators skip. Without it, two people can take the same data and produce returns that differ tenfold. Agree it with whoever signs off spending before the first number goes in. If the result will go to a board, the guide to an AI business case shows how to present it.

Which costs belong in the calculator?

Five lines belong in every calculation: licences and usage, setup, the time people spend reviewing output, outside help, and upkeep. Price each per workflow, and split one-off costs from monthly ones. Software is usually the smallest line, which is why calculators that start and finish with licences look so good.

1. Licences and usage. Seats, usage charges and add-on tools. According to Anthropic's pricing page in October 2026, a standard Claude Team seat cost USD 20 a month billed annually.
2. Setup. Writing down the method, loading the knowledge the workflow needs, connecting systems.
3. Review time. Minutes per output, times outputs per month, times the reviewer's hourly cost.
4. Outside help. Consultants, contractors or agency fees tied to this workflow.
5. Upkeep. Keeping the knowledge current and fixing the workflow when tools change.

Review time needs one rule. If your result is measured end to end, from request to approved work, review is already inside that number. Adding it again as a cost counts it twice. Count it separately only when you measure drafting time alone.

What should you leave out of the calculator?

Leave out anything that cannot be traced to a record: hours people say they saved, percentage gains quoted by vendors, capacity that was freed but not used, and wins counted in two workflows at once. Also leave out costs you would have paid anyway, such as a subscription bought for other reasons.

Self-reported hours. People misjudge their own speed with AI. In METR's randomised trial, developers took 19% longer with AI tools, yet still believed afterwards that they had been 20% faster.

Vendor benchmarks. A "40% productivity gain" from a product page describes someone else's work, measured someone else's way.

Unused capacity. Hours freed but not redeployed are real, but they are capacity, not return.

Double counting. If a research workflow feeds a proposal workflow, credit the win to one of them.

Sunk costs. A licence the team already held for other work is not a cost of this workflow.

How do you turn a change in a business number into dollars?

Use a value rule per measure. A lower cost per lead is worth the saving times lead volume. More pipeline is worth the extra pipeline times your historic win rate and margin. Faster work is worth money only when it removes spend, such as contractor fees or a hire, or produces more of something measured.

Typical value rules:

• Cost per lead: old cost minus new cost, times leads per month.
• Pipeline: added pipeline, times historic win rate, times gross margin.
• Time to approved work: contractor or overtime spend avoided, or extra output sold.
• Error rate: fewer errors, times the average cost to fix one.

Time is where calculators go wrong most often. A faster workflow frees hours, and hours look like money, so they get multiplied by salary. That number only becomes real when the hours go somewhere measured. The article on why time saved is not ROI covers how to tell the difference.

What does a worked example look like?

Halcyra, a fictional 40-person services firm, writes 30 proposals a month. Each took six hours from brief to approved, and overflow went to a contractor. With AI drafting from past wins, end-to-end time fell to three and a half hours, and the contractor was no longer needed. All figures here are illustrative.

On those numbers, year-one ROI is about 98%: net benefit of AUD 17,840 divided by AUD 18,160. Setup pays back in roughly five months, because the monthly benefit of AUD 3,000 runs well ahead of the monthly running cost of about AUD 513.

Now count it the flattering way. The 30 proposals freed 75 hours a month. Valued at AUD 90 an hour, that is AUD 81,000 a year, and the calculator shows about 346%. However, only the contractor saving is backed by an invoice. The rest is capacity until someone points it at a measured job.

How do you stop the calculator flattering the result?

Run two cases, not one: an expected case and a low case that assumes the smaller result. Fund the workflow on the low case. Name an owner for every number, compare against the real alternative, such as a hire or doing nothing, and recheck the result at 90 days with fresh system data.

This matters because leaders are under pressure to show a number. According to Microsoft and LinkedIn's 2024 Work Trend Index, 59% of leaders worry about quantifying AI's productivity gains. Under that pressure, generous assumptions creep in, and the first audit wipes out the credibility of every later figure.

The wider picture is sobering, too. McKinsey's 2025 State of AI survey found 88% of organisations use AI in at least one function. Yet only 39% report any earnings impact at the enterprise level. A calculator that admits the low case is more believable than one that promises the average.

The calculator is far easier to keep honest when the workflow records its own results. Skein builds workflows that report measures such as pipeline, cost per lead and time to first draft; the outcomes page lists what is tracked. That suits a business that wants its ROI fed by system data rather than estimates. A team with one or two light AI uses may find a spreadsheet and a quarterly check is plenty.

Frequently asked questions

What is an AI ROI calculator?

It is a worksheet that estimates the return on one AI workflow. It takes the full cost, a baseline measure from before AI, the same measure afterwards and a rule for turning the change into money. Good calculators work per workflow and record who owns each number, so the result can be checked.

What costs should an AI ROI calculator include?

Include licences and usage, setup, the time people spend reviewing output, outside help and upkeep. Split one-off costs from monthly ones. Leave out costs the business would have paid anyway, and do not count review time twice if your result is already measured from request to approved work.

Should hours saved go in an AI ROI calculator?

Only when the freed hours go into something measured, such as more sales calls, extra output sold or contractor spend avoided. Self-reported hours are unreliable and freed capacity is not a return by itself. Value the result of the redeployed time, not the time itself, and keep estimates out of the total.

What is a realistic payback period for AI?

It depends on the workflow. A frequent job with a clear saving, such as contractor spend avoided, can pay back setup within months. Broader projects take longer because setup is larger and results are slower to show. Work out payback on the low case, not the hopeful one, before committing more budget.

Are online AI ROI calculators accurate?

Most are generous, because they multiply hours saved by salary and leave out review time, upkeep and setup. Some come from vendors with an interest in a high number. Use them for a rough sense of scale, then build your own from system data, a baseline and an agreed value rule.

Related reading