Executives keep asking for the ROI of AI.
It is the right instinct and the wrong unit of analysis.
AI is not one investment. It is a new input into many workflows. Asking for the ROI of AI is like asking for the ROI of Excel. The answer depends on where it entered the work, what changed around it, and whether anything downstream actually improved.
Excel did not have one ROI. It had different returns in budgeting, forecasting, reporting, scheduling, sales operations, inventory, audit prep, and the thousand unofficial models people built because the official system could not answer the question quickly enough.
AI is similar, but faster and stranger.
It can sit inside research, drafting, coding, analysis, meeting prep, training, customer support, compliance review, software testing, proposal work, knowledge search, and management reporting. In one workflow, it may save meaningful time. In another, it may create confident rework. In another, it may produce a useful first draft but no business value because the old work never stopped.
That is why the question feels so slippery.
The solvable question is not:
What is the ROI of AI?
The solvable question is:
Where is AI creating workflow ROI?
That can be measured.
Usage is not ROI
The first measurement layer is usage.
Are people using the tool? How often? Which teams? Which features? Which tasks?
This is useful. It is not ROI.
Usage tells you where attention is going. It tells you where people are curious, pressured, opportunistic, or desperate enough to try something new. It can reveal demand. It can reveal shadow work. It can reveal that one function is experimenting far more than another.
But a usage dashboard can make an organization feel more advanced than it is.
Someone used AI to draft a memo. Someone summarized a meeting. Someone generated twenty slide titles. Someone asked it to explain a regulation. The chart goes up. The adoption story sounds good.
The work may not have changed.
This is the first measurement trap. Activity is easy to count, so it starts masquerading as progress. The organization begins to optimize for more prompts, more seats, more weekly active users, more generated output. None of those are bad signals. They are just incomplete signals.
The question usage can answer is: are people trying?
It cannot answer: did the business get better?
Task value is still not always ROI
The second layer is task value.
Did AI help with a specific task?
Could someone draft faster, summarize faster, compare documents faster, write code faster, find examples faster, prepare for a meeting faster, or create a first version faster?
This is where people first feel the value. It is also where many organizations stop thinking.
Task value is real. It matters. If a person can complete a recurring task in twenty minutes instead of two hours, something has changed. It would be silly to dismiss that because it has not yet shown up in EBIT.
But task value is not automatically business value.
The saved time may disappear into more versions. The faster draft may still require the same senior review. The meeting summary may be accurate enough to feel useful but not reliable enough to replace attendance. The code may compile but create new maintenance burden. The analysis may be polished but answer the wrong question.
Task acceleration often creates a local feeling of progress before the workflow has absorbed it.
This is why the phrase "AI saved me time" needs a follow-up question.
What happened to the time?
If the answer is "I made more drafts," the organization may have more activity, not more value. If the answer is "we moved the client decision forward," "we reduced rework," "we caught the issue earlier," or "we stopped doing the old manual step," the conversation becomes more interesting.
Task value is the beginning of the ROI investigation, not the end.
Workflow value is where ROI starts
The third layer is workflow value.
This is where the unit of analysis gets serious.
Did the full workflow improve?
Not the prompt. Not the task. The workflow.
For example:
- The proposal cycle went from ten days to six.
- The contract review covered more risks with the same review capacity.
- Support cases resolved faster without lowering customer satisfaction.
- Research memos reached review with fewer basic issues.
- Engineering teams migrated code with fewer regressions.
- Training content moved from ad hoc production to a repeatable update process.
- Internal knowledge search reduced the number of people interrupted for the same answer.
This is where AI ROI starts to become measurable because the workflow has boundaries. It has volume, inputs, handoffs, review points, cost, cycle time, quality standards, and outcomes.
It also has politics.
A workflow does not improve just because one task gets faster. The surrounding work has to change. The review step may need to move earlier. The approval rule may need to change. The handoff may need to disappear. The old status meeting may need to stop. The manager may need a different quality standard. The team may need to decide which outputs are allowed to leave the draft stage.
This is why AI ROI so often underwhelms.
The tool changes faster than the workflow around it.
Public research keeps pointing at this gap. McKinsey's 2025 State of AI report found widespread organizational AI use, but much lower reported enterprise-level EBIT impact. BCG's 2026 AI at Work research found that many regular frontline AI users save significant time, while many receive little guidance on what to do with that time. Gallup found that frequent AI users often report productivity gains, but evidence of fundamental changes to work remains more limited.
The pattern is not mysterious.
Individual productivity can rise while organizational value stays flat.
That is what happens when AI makes tasks faster but the workflow remains intact. The organization gets more output moving through the same old pipes.
Workflow ROI requires the pipes to change.
Business value is where ROI becomes real
The fourth layer is business value.
Did the workflow improvement affect money, risk, capacity, quality, speed, or customer outcomes?
This is the level executives actually care about, even when the conversation starts with tool adoption.
The measures differ by workflow:
- lower delivery cost
- higher margin
- faster revenue cycle
- fewer errors
- reduced risk exposure
- more clients served with the same headcount
- better win rate
- fewer escalations
- faster onboarding
- shorter cycle time
- lower rework
- better decision quality
This is also where AI costs belong.
Token spend, licenses, integration work, review time, security work, model evaluation, and training are not separate from ROI. They are part of the unit economics of the workflow.
The question is not whether the AI was free. It is whether the AI-assisted workflow is better after its full cost is included.
A costly AI run may be cheap if it prevents a bad decision, accelerates a high-value review, or lets a team handle more work without lowering quality. A cheap AI run may be expensive if it creates rework, false confidence, or another layer of process nobody needed.
The price of the tool is rarely the whole story.
The cost of bad workflow measurement is usually larger.
The ladder matters
The ladder is simple:
- Usage: are people using it?
- Task value: did a specific task get easier?
- Workflow value: did the full workflow improve?
- Business value: did money, risk, capacity, quality, speed, or customer outcome change?
Most AI reporting gets stuck at level one or level two.
That is understandable. Usage data is easier to collect. Task stories are easier to tell. "We have 80 percent adoption" sounds cleaner than "we found four workflows where AI reduced rework, six where it created more review burden, and twelve where the result is still unclear."
The second sentence is more useful.
Executives do not need a universal AI ROI number. They need a portfolio view of workflow bets. Some will be noise. Some will create local productivity. A smaller number will create serious enterprise value. The work is to find those faster, standardize them, and stop funding the obvious waste.
This is not a softer ROI discipline. It is a better one.
It forces the organization to name where the value is supposed to appear. It prevents tool adoption from passing as transformation. It makes time savings answerable to a harder question: what changed because the time was saved?
The question to ask instead
The better executive question is not:
What is the ROI of AI?
It is:
Which workflows changed, and how do we know?
That question is harder to answer in a board deck. It is also much harder to fake.
If AI is only producing more drafts, the answer will show up. If it is saving time but not changing capacity, the answer will show up. If it is creating rework, the answer will show up. If it is quietly becoming infrastructure inside a critical workflow, the answer will show up there too.
AI ROI is not missing.
It is usually measured in the wrong place.