ARTICLE · RESEARCH AND BENCHMARKS
Why AI ROI stalls
at the workflow
The tools are in place. The licences are paid for. People are using them. The return is still difficult to find.
Thrivin · Research and Benchmarks · September 2026 · 6 min read
Most AI investments are not failing at the model. They are failing at the work around it.
An executive team approves the pilot, the pilot succeeds, and the rollout follows. A year later the finance review asks what changed. Cycle times look similar. Headcount looks similar. The people using the tool report that it helps, and they are right.
Both things are true at once, and the reason is structural: the tool was introduced into a workflow that was never redesigned to accept it.
A faster step inside an unchanged workflow produces a faster step, not a better return.
SECTION 01
Adoption is not the constraint
Adoption is the metric that is easy to collect, so it becomes the metric that is reported. Seats activated. Prompts run. Weekly active users. None of them describe whether the business is getting more for what it spends.
The questions that do are harder and older than AI. What does this workflow cost today? How long does it take end to end? How often is the work redone? Who touches it, and at what rate?
Organizations that cannot answer those four questions before deployment have no baseline afterwards — and without a baseline, return is an assertion.
EXHIBIT 1 — ADOPTION IS NOT RETURN
Report active use of an AI tool in at least one core workflow.
Redesigned the workflow the tool now sits inside.
Reassigned the capacity the tool released to named work.
Illustrative figures. To be replaced with the 2026 State of the WorkSystem™ survey values. Source: Thrivin
SECTION 02
Where the return actually stops
Each of the four stalls below sits outside the tool, which is why a successful pilot does not predict a return at scale.
EXHIBIT 2 — THE FOUR STALLS
Handoffs the tool does not touch. A model drafts the summary in seconds. It then waits four days for an approval that no one has redesigned.
Inputs assembled by hand. Someone still gathers the data, checks the fields and pastes it in. The work moved upstream rather than leaving.
Verification nobody planned for. Output that must be checked line by line transfers effort from producing to reviewing, often to a more expensive person.
Released capacity with no destination. Six hours a week are freed and quietly absorbed. Saved time that is not reassigned never appears in the result.
Source: Thrivin
SECTION 03
Measure the workflow, not the tool
A tool metric tells you the tool is being used. A workflow metric tells you whether the investment paid.
EXHIBIT 3 — WHAT ORGANIZATIONS MEASURE (% OF RESPONDENTS)
| Measure | % |
|---|---|
| Tool adoption | 81 |
| Workflow redesigned around it | 34 |
| Baseline measured before rollout | 29 |
| Released capacity reassigned | 22 |
Illustrative figures. To be replaced with the 2026 State of the WorkSystem™ survey values. Source: Thrivin
Four measures are enough to start: total cost to run the workflow, elapsed time from request to delivered outcome, rework rate, and the distribution of effort across the people and systems involved.
Take them before the change. Take them again ninety days after. Report the difference against the baseline rather than against the business case, and separate the two kinds of gain — money that leaves the cost line, and capacity that becomes available.
Released capacity is not a saving until it is reassigned to work that produces value.
Count implementation, training, supervision, integration and verification in the cost. A tool that shifts effort into checking its own output has changed where the work happens, not how much there is.
“We had adoption in eleven weeks. It took us another two quarters to admit the approval chain had never changed, and that was where the money was.”
Chief Operating Officer · Professional Services
SECTION 04
Sequence the change
Most programmes run steps three and four and skip the rest. The skipped steps are the ones that make the return visible.
FIVE STEPS, IN ORDER
Define success. State the outcome the workflow owes the business, in the terms the business already uses.
Establish the baseline. Current cost, cycle time, rework rate and who touches the work. Without it there is no return to report.
Remove and simplify. Eliminate steps before automating them. Automating a step that should not exist makes it permanent.
Place the capability. Decide what belongs to people, to AI, to automation and to external specialists — in that order of judgement.
Reassign the capacity. Name the work the released hours will do. Report the return against the baseline, not against the forecast.
The model is rarely the limiting factor. The work around it usually is.
Start with one workflow that matters, measure what it costs today, and change its design before adding anything else to it.
Take the WorkSystem Pulse™ — free, three minutes, and your WorkSystem Executive Intelligence Brief™ arrives immediately.