AI ROI

The return on AI is knowable. Most of the numbers you've been shown aren't.

AI ROI is net value over total cost of ownership. The formula is the easy part — the discipline is capturing a baseline before you launch, loading the true cost, and being willing to publish the number when it disappoints.

What AI ROI means

AI ROI is the net financial return from an AI system measured against everything it costs to build, run, and govern over its useful life.

ROI = (value created − total cost of ownership) ÷ total cost of ownership

Always report it alongside payback period. A ratio without a clock is half an answer — a 200% return over three years and a 200% return in five months are different decisions.

It differs from traditional software ROI in two ways that matter. The costs are recurring and usage-based rather than a fixed licence — spend scales with traffic, not seats. And the value depends on the system's hit rate on a real task and on what the misses cost, not on whether the feature shipped. Software either works or doesn't. An AI system works a percentage of the time, and the economics live in that percentage. There's a fuller definition in our glossary entry for AI ROI.

The baseline is the whole thing

Almost every unreliable AI ROI number fails at the same place, and it isn't the maths. It's that nobody recorded what the metric was before launch. Without a pre-deployment baseline, a measurement window, and a named system of record, you cannot attribute what happened afterwards — you can only tell a story about it. And plenty of things move a conversion rate in the same quarter you launch an agent.

A baseline is three specific things, captured before anything goes live:

The metric, defined precisely enough that two people would compute it the same way — not "conversion" but "inbound form fills that become closed-won within 60 days."
The window — a period long enough to survive seasonality, recorded with its start and end dates.
The source of record — the one system the number is pulled from, so nobody relitigates it later.

This is unglamorous and it is the entire difference between a number you can take to a board and a number you can only put on a slide. It also has a cost: a real baseline will sometimes tell you the thing didn't work. That's the point. A measurement system that can only return good news isn't a measurement system.

Where the value actually comes from

Value created splits into three streams. Count them separately — they have very different evidentiary standards.

Labour reclaimed

Hours no longer spent on work that never needed human judgment — valued at loaded cost, not headline salary, and only counted where the time is genuinely redeployed rather than notionally freed.

Revenue moved

Pipeline or conversion the system actually shifted: faster response, more of the queue worked, fewer leads going cold. This is the stream most worth having and the hardest to attribute honestly.

Loss avoided

Churn that did not happen, errors not made, cost-to-serve that came down. Real, but the easiest place to talk yourself into a number, because the counterfactual is invisible.

Total cost of ownership is not the licence fee

The most common way an AI business case goes wrong is booking the platform fee as “the cost.” The real bill looks like this:

Platform and consumption — usage-based inference that scales with traffic, not with seats.
Build: integration, data cleanup, and the unglamorous foundation work that dominates the bill.
Humans in the loop — reviewing, correcting, and escalating. This never goes to zero.
Ongoing management: monitoring, retuning, and catching drift before it costs you.
The cost of the misses — what a wrong answer costs, times how often the system is wrong.

The last one is the one nobody models, and on a customer-facing system it can dominate everything above it. The model is the cheapest part of the bill.

A word on benchmarks

You'll be offered a lot of multiples — 3×, 4×, 8× — usually with no baseline, no window, and no source attached. Treat a multiple without those three things as decoration rather than evidence. When a vendor quotes you a number, the useful questions are: what was the baseline, over what window, and from which system of record?If they can't answer all three, you've learned something more useful than the number.

We'd rather run a baseline on your data than quote someone else's figure at you — the honest way to set expectations is per use case, against your own starting point. There's more on that in how to set honest AI-agent ROI benchmarks by use case.

Work through it on your own numbers

Calculator

Model your AI ROI

Put your own volume, conversion, and labour cost in and see the shape of the return.

Open the calculator

Deep dive

Measuring Agentforce ROI in 2026

The formula worked end to end on Salesforce Agentforce, with the cost lines named.

Read it

Frequently asked

What is AI ROI?

AI ROI is the net financial return from an AI system measured against everything it costs to build, run, and govern over its useful life: ROI = (value created − total cost of ownership) ÷ total cost of ownership. It is expressed as a ratio or percentage, and it is only meaningful next to a payback period — a 200% return that takes three years is a different decision from a 200% return that pays back in five months.

How is AI ROI different from traditional software ROI?

Two ways. First, the costs are recurring and usage-based — per-token inference, monitoring, error correction — rather than a fixed licence you amortise. Spend scales with how much you use it, not with how many seats you bought. Second, the value depends on the model's hit rate on a real task and on what the misses cost, not on whether the feature shipped. Traditional software either works or does not. An AI system works a percentage of the time, and the economics live in that percentage.

Why do most AI ROI numbers not hold up?

Because there was no baseline. If you did not record what the metric was before launch — with a defined measurement window and a source of record — you cannot attribute what happened after. Most reported AI ROI is a post-hoc story fitted to a number that moved for several reasons at once. The second failure is scoping the cost as the platform fee, then discovering the real spend was integration, data cleanup, and humans reviewing output.

What is a good ROI for an AI deployment?

Anyone quoting you a universal benchmark is selling something. The honest answer is that it depends on the use case, the baseline you started from, and what your misses cost — and that a credible number always comes with the baseline attached. Ask any vendor quoting a multiple what the pre-deployment baseline was, over what window, and from which system of record. If they cannot answer all three, the multiple is decoration.

Should AI ROI be measured per use case or across the whole programme?

Per use case, then rolled up. Programme-wide averages hide the thing you most need to know: which use cases pay and which are subsidised by the ones that do. Most AI portfolios have a small number of genuine winners funding a long tail that would be killed if measured on its own.

When should you measure — at pilot or in production?

Production, on real traffic. A pilot measures the system on inputs it was designed for, which is exactly where an AI system looks best. The long tail the pilot never hit is where the cost and the risk concentrate, so a pilot ROI is an upper bound rather than a forecast.

Ready when you are

Worth a
conversation?

Tell us one number you'd like AI to move. We'll show you how we'd do it, what it's worth, and how we'd tie our fee to getting you there.

Book a call hello@skysync.nyc