Category:
AI ROI
Published date:
Key Takeaways
95% of enterprise AI pilots produce no measurable P&L impact. The failure is almost never model choice. It is that nobody instrumented anything.
Both halves of the ROI fraction are usually broken. Shadow AI hides 20% to 40% of real spend, and most companies never attribute outcomes to the tools that produced them.
Adoption is the leading indicator. Spend is the lagging one. Dormant licences are negative ROI dressed up as investment.
Cost per outcome is the metric that works: cost per resolved ticket, per shipped commit, per booked lead, against the manual baseline. It turns ROI from an argument into arithmetic.
Guickly measures adoption, attribution and marginal utility in one place, sanctioned and shadow. It connects every AI dollar to what it actually returned.
AI ROI is the business value attributable to AI divided by the fully-loaded cost of producing it, including shadow AI. The formula is (value attributable to AI minus total AI cost) divided by total AI cost. Most companies cannot compute it because the denominator misses 20% to 40% of real spend and the numerator is never measured at all.
Your company will spend more on AI this year than last, probably by a lot. Ask your leadership team what last year's spend returned and you will get a number nobody can defend.
That is not a local failure. 95% of enterprise AI pilots produce no measurable P&L impact, according to MIT's 2025 research. The 5% that do have one thing in common, and it is not which model they picked. They measure properly. Almost nobody else does.
The AI ROI formula
Simple arithmetic. Both terms are usually wrong.
The denominator is understated. Shadow AI hides 20% to 40% of real spend outside the budget finance can see. If you are dividing by the sanctioned bill alone, your ROI is flattering and false.
The numerator does not exist. Most companies never attribute outcomes to the tools that produced them. You cannot compute a ratio when one term is understated and the other was never measured.
Fix the denominator first. It is the easier half, and an inflated ROI number is more dangerous than no number at all.
Why do most companies fail to measure AI ROI?
Because they measure the bill. The bill tells you what you spent. It tells you nothing about what you got. So the board conversation stalls on the only number anyone can produce, and that number is an input, not a return.
The data backs the pattern. McKinsey's State of AI found only about 6% of companies attribute more than 5% of EBIT to AI, and only 39% report any enterprise-level EBIT impact at all.
It is not that AI does not work. It is that most organisations cannot see whether it is working. Spend is the lagging indicator. The things that predict ROI sit upstream, and almost nobody instruments them.
The three layers you have to measure
A defensible AI ROI number rests on three measurements, in order. Skip one and the number collapses.
Layer | The question it answers | The metric | What it tells you |
|---|---|---|---|
1. Adoption | Is what we bought actually being used? | Active users by department, weighted by headcount | Leading indicator. Dormant licences are negative ROI dressed as investment. |
2. Attribution | Which dollar produced which outcome? | Cost per outcome vs manual baseline | Turns ROI from argument into arithmetic |
3. Marginal utility | Is the next dollar still worth it? | Value per additional dollar of inference | Tells you where to stop spending and reallocate |
Layer 1: Adoption. This is the leading indicator, and the one boards never ask for. Buying seats is not adoption. Dormant licences produce zero return at full cost. Before you ask what AI returned, ask how fluent your organisation is with it, by department. AI fluency is the leading indicator of AI ROI. Spend is the lagging one.
Layer 2: Attribution. Aggregate AI spend tells you nothing actionable. Spend mapped to a workflow, a team and an outcome tells you everything. The unit you want is cost per result: cost per resolved ticket, per shipped commit, per booked lead. Once AI is expressed as a cost per outcome, you can compare it to the manual baseline and the ROI becomes arithmetic.
Layer 3: Marginal utility. Foundation Capital's Jaya Gupta calls this marginal token utility, the business value created by each additional dollar of inference. ROI is not a single annual figure. It is a curve. Some workflows return 4x. Some return nothing past a certain volume. Leadership's job is to find where the next dollar stops paying and move it. You cannot do that without the first two layers feeding it live.
Adoption tells you it is being used. Attribution tells you what it produced. Marginal utility tells you whether to keep going. The bill tells you none of it. Seeing all three in one place, sanctioned and shadow, is the problem Guickly was built to solve.
What good AI ROI actually looks like: benchmark numbers
Guickly modelled a 1,000-person enterprise across its support, engineering and go-to-market functions. The results below are our own analysis, not vendor claims.
Metric | Before AI | After AI | Change |
|---|---|---|---|
Cost per support ticket | $6.10 | $1.42 | −77% |
Gross margin | baseline | +4.2 points | +4.2 pts |
Blended return per AI dollar | — | 3.8x | — |
Share of real AI spend invisible to finance | 20% to 40% | — | the denominator problem |
None of those numbers exist without the three layers underneath them. The companies in the 95% bought the same tools. They just never built the instrumentation to know what came back.
The pattern repeats at scale. IBM credited its AskHR agents and related automation with $3.5B in cumulative productivity, but only because it knew which functions to automate and which to expand. Goldman Sachs reported over 20% developer productivity from its firmwide assistant rollout, a number it could only state because it measured adoption across 12,000 engineers first.
What is the best single metric for AI ROI?
Cost per outcome. Cost per resolved ticket, per shipped commit, per booked lead, measured against the manual baseline.
It works because it is comparable across very different workflows, it survives contact with a CFO, and it turns a debate into a calculation. Every other metric is either an input (spend, seats, tokens) or too abstract to act on.
How long before AI ROI shows up?
Expect adoption signal in weeks, attribution in a quarter and margin impact in two to four quarters. If you are looking for P&L movement in month one you will conclude AI does not work, which is the mistake most of the 95% made. Instrument adoption first, because it moves first and predicts everything downstream.
The question for your next board meeting
Do not walk in with the AI spend number. Everyone has that one and it answers nothing.
Walk in with three. How fluent is the organisation. What is each major workflow's cost per outcome against its manual baseline. And where does the next AI dollar stop paying.
If your team cannot produce those, you do not have an AI ROI problem yet. You have an AI measurement problem. And the measurement problem decides which 5% you end up in.
FAQ
How do you measure AI ROI? AI ROI is the business value attributable to AI divided by its fully-loaded cost, sanctioned and shadow. The formula is (value attributable to AI minus total AI cost) divided by total AI cost. Computing it requires three upstream measurements: adoption, attribution of outcomes to spend, and marginal utility of the next dollar. Spend alone is not a measure of ROI.
What is the AI ROI formula? (Value attributable to AI − Total AI cost) ÷ Total AI cost. The common error is understating total cost by excluding shadow AI, which typically hides 20% to 40% of real spend, and having no method to quantify the value term at all.
What is a good AI ROI? In Guickly's modelling of a 1,000-person enterprise, AI returned a blended 3.8x per dollar, cut cost per support ticket from $6.10 to $1.42 and added 4.2 points of gross margin. Treat these as directional. Returns vary enormously by workflow, which is why marginal utility matters more than a single company-wide figure.
Why do most companies fail to get ROI from AI? MIT found 95% of enterprise GenAI pilots produce no measurable P&L impact, and McKinsey found only about 6% of companies attribute more than 5% of EBIT to AI. The failure is rarely model choice. Most organisations never instrument adoption or attribute outcomes, so they cannot tell whether AI is working.
What is the best metric for AI ROI? Cost per outcome: cost per resolved ticket, per shipped commit, per booked lead, compared to the manual baseline. It is comparable across workflows and turns AI ROI from an argument into arithmetic.
Is AI fluency a leading indicator of AI ROI? Yes. Adoption and fluency lead, because dormant licences produce zero return at full cost. If a tool is not being used, no downstream ROI is possible. Spend is a lagging indicator, which is why measuring it alone tells leadership nothing.
What is marginal token utility? A term coined by Foundation Capital's Jaya Gupta for the business value created by each additional dollar of AI inference. It reframes ROI as a curve rather than a single figure. Some workflows keep paying as spend grows, others stop. Leadership's job is to find the inflection and reallocate.
How does shadow AI affect AI ROI? It breaks the denominator. Shadow AI typically accounts for 20% to 40% of real AI spend and sits outside the budget finance can see. Any ROI calculated on the sanctioned bill alone overstates the return. Fixing visibility is a prerequisite for computing ROI honestly.
How long does it take to see AI ROI? Adoption signal in weeks, outcome attribution in about a quarter, margin impact in two to four quarters. Looking for P&L movement in the first month is the mistake that leads teams to abandon working deployments.
![](data:image/svg+xml,<svg display="block" role="presentation" viewBox="0 0 17 17" xmlns="http://www.w3.org/2000/svg"><path d="M 3.335 0 L 13.335 0 L 13.335 1.667 L 3.335 1.667 Z M 0 3.333 L 1.667 3.333 L 1.667 13.333 L 0.001 13.333 L 0.001 3.333 Z M 15 3.333 L 16.667 3.333 L 16.667 13.333 L 15 13.333 Z M 13.333 1.667 L 15 1.667 L 15 3.333 L 13.333 3.333 Z M 1.667 1.667 L 3.333 1.667 L 3.333 3.333 L 1.668 3.333 L 1.668 1.667 Z M 3.333 16.667 L 13.333 16.667 L 13.333 15 L 3.335 15 L 3.335 16.667 Z M 13.333 15 L 15 15 L 15 13.333 L 13.333 13.333 Z M 1.667 15 L 3.333 15 L 3.333 13.333 L 1.668 13.333 L 1.668 15 Z M 7.501 3.333 L 9.168 3.333 L 9.168 9.167 L 7.501 9.167 Z M 9.168 9.167 L 10.834 9.167 L 10.834 10.833 L 9.168 10.833 Z M 10.834 10.833 L 12.501 10.833 L 12.501 12.5 L 10.834 12.5 Z" fill="rgb(43, 54, 52)" height="16.667px" id="PxKarjvpl" width="16.667px"/></svg>)