The hours saved are not yet the return

Time saved is evidence that a task changed. It is not automatically a financial return. To make a credible AI ROI claim, trace four links: people use the workflow in real work; their behavior changes; an operational result improves; and that result creates or protects financial value. State where the evidence stops.

Multiplying estimated hours by salaries skips most of that chain. A founder needs enough evidence to decide whether to expand the workflow, change it, or stop paying for a result that never leaves the demo.

The ten-hour question

An anonymized US eyewear wholesaler estimated that a tailored ERP and AI-assisted reporting workflow returned about ten hours each week. This is a meaningful client estimate. It tells us the old reporting work consumed real capacity and the new workflow reduced it.

The estimate was not an instrumented time study, and it does not tell us what the business gained from that capacity.

Did the team reduce overtime? Produce reports earlier? Spend more time investigating exceptions? Handle more orders without another hire? Make a faster purchasing decision? Each answer creates a different value case. “Ten hours” is the first observable change, not a license to invent the rest.

Another anonymized project points to a different measurement problem. A US presentation-training business used an AI coaching product to open a new revenue line. Counting labor saved would miss the point. The relevant chain runs through customer uptake, coaching quality, delivery cost, retention and gross margin from a service that did not exist before.

The same technology label can sit on two entirely different economics.

This is why use-case selection should begin with the queue or outcome that needs to change, not a generic promise that AI will save time.

Build an evidence chain before an ROI model

Use four stages and do not skip one because the number is convenient.

Usage is necessary for an employee tool and still not enough. A person can send many prompts while doing the same work twice: once with AI and once manually to check it. A pilot can reduce drafting time while adding so much review that the complete workflow is slower.

OpenAI’s deployment guidance distinguishes product activity from impact in much the same way. Its measurement guide recommends connecting bottom-up efficiency with outcomes such as revenue capacity, customer satisfaction and cycle time. Its evidence guidance is more pointed: attending a demonstration or trying a tool once shows exposure, while repeat use and a changed team process show adoption.

Name the value you expect

An AI project does not need to create every type of value. Name one primary mechanism and a few guardrails.

Capacity

The workflow lets the same team complete more valuable work. Measure completed cases, backlog, elapsed time and what the returned attention is used for. Capacity has value when demand exists.

Quality

The workflow makes outputs more complete, consistent or accurate. Measure against a defined rubric, not whether the prose sounds polished. Include errors the system introduces and disagreements between reviewers.

Revenue

The workflow raises conversion, speeds time to revenue, protects retention, or enables a new offer. Use gross margin rather than headline revenue, and separate a new capability from customer demand for it.

Avoided cost or loss

The workflow reduces overtime, contractor spend, rework, missed follow-up, leakage, or another observable cost. “Avoided hire” is credible only when workload, hiring intent and available capacity support it.

Risk and resilience

The workflow improves traceability, consistency, recovery time or early detection. Expected-loss estimates can be useful for material risks, but false precision is dangerous when probabilities are unknown. Often the honest first result is a control measure: every exception is now recorded and reviewed within a defined time.

Preserve the baseline while you still can

Once the new workflow becomes normal, reconstructing the old one is surprisingly hard. Before the pilot, save a small baseline card.

Workflow:
Unit of work: one [report / case / order / request]
Eligible population:
Primary outcome:
Guardrail outcomes:

Baseline period or sample:
Cases observed:
Elapsed time per case:
Human touch time per case:
Accepted without rework:
Exceptions or harmful errors:
Work in progress / backlog:
Cost per completed acceptable case:

Evidence source:
Owner:
Known seasonal or selection effects:

The unit matters. “Hours saved per employee” invites estimates across unrelated tasks. “Minutes of review per accepted claim file” can be observed. “Cost per successful customer request” includes both efficiency and quality.

Keep the denominator visible. Ten corrected outputs out of 100 attempts means something different from ten out of 10,000. An average can also hide an expensive tail, so record high-consequence exceptions separately.

Measure the complete task, including review and recovery

AI systems can move effort rather than remove it. Drafting becomes faster; fact-checking grows. Classification improves; investigating uncertain cases takes longer. An agent completes routine requests; one incorrect action consumes a day of repair.

For each completed unit, account for:

• model and tool cost;
• employee setup and prompting;
• review and correction;
• escalation and exception handling;
• failed attempts and retries;
• integration and monitoring;
• training and workflow support; and
• incidents or downstream rework.

Then measure cost per acceptable completed unit, not cost per model call. A cheaper model that requires more retries may be more expensive. A slower workflow that produces a substantially better decision may be worth it.

For a concrete way to test acceptable completion, review effort, dangerous failures, and release boundaries together, use the end-to-end AI workflow evaluation method.

NIST’s AI Risk Management Framework recommends documenting test sets, metrics and evaluation methods. That discipline is not only for model safety. It makes the business case auditable: someone else can see what was tested, on which cases, under which definition of success.

Treat published productivity numbers as examples, not your forecast

In a field study of 5,179 customer-support agents, Brynjolfsson, Li and Raymond found that a generative AI assistant increased issues resolved per hour by nearly 14% on average in one customer-support deployment. The gains were much larger for less experienced workers and small or negative for the most experienced group.

That variation is the lesson. The study is strong evidence that AI can affect production work under specific conditions. It is not a benchmark to paste into a proposal for wholesale reporting, clinical intake or sales operations.

Your baseline, workflow, users, task mix, quality threshold and adoption determine your result. A pilot that includes only enthusiastic volunteers may overstate broad rollout performance. A test on easy cases may hide the exception cost. Comparing the month before and after launch may confuse the system effect with seasonality, staffing or a policy change.

Perfect experimental design is rarely economical for an SME, but a few choices improve honesty:

• compare the same definition of an acceptable completed task;
• use representative cases, not only showcase examples;
• preserve case mix and role differences where possible;
• record other changes occurring during the test;
• separate measured data from employee estimates; and
• report a range when attribution or future demand is uncertain.

Convert operations into money only where the link is real

Once an operational result is established, translate the portion that has an economic consequence.

For capacity work:

Value from added capacity
= additional acceptable units completed
× contribution margin per unit
× portion reasonably attributable to the workflow

For a removed recurring cost:

Avoided cost
= observed reduction in overtime, contractor spend, rework, or another cash expense

For a new AI-supported offer:

Contribution
= collected revenue
− variable delivery, model, support, payment and review costs
− incremental refunds or service recovery

Compare the result with the full cost over the same period:

Net value
= attributable financial value
− build, integration, migration, training, operation, review, maintenance and change costs

Do not turn all returned time into salary savings unless payroll actually changes. The stronger case may be that a constrained specialist serves more customers, a report arrives before a decision, or the team absorbs growth without degrading service.

The OECD notes that intangible investments such as training and adjustments to business processes and software complicate productivity measurement. Those investments are easy to omit from the denominator precisely because they do not appear on the model bill.

Use the result to make a decision, not decorate a launch

Agree on possible decisions before the evidence arrives:

• Continue when the workflow is used, quality holds, and the operational signal is moving in the expected direction.
• Change when the model can perform the task but integration, review or adoption prevents the result.
• Narrow when value is concentrated in a role, case type or stage of the workflow.
• Stop when the complete cost exceeds the value, failures are disproportionate, or a simpler intervention wins.

A negative result can still expose the real constraint: data quality, ownership, an unnecessary approval or lack of demand. Do not rescue a weak pilot by changing the outcome after launch.

State where the evidence ends

The most credible ROI sentence is sometimes modest:

The team used the workflow on 140 eligible cases. Median preparation time fell, reviewer effort did not increase, and the error rate remained within the agreed threshold. We have not yet shown an effect on conversion.

That sentence gives a leader something to decide with. It also says what to measure next.

Time saved can be a real outcome when it reduces overtime or clears a backlog. It can be a path to value when the team uses the capacity for more productive work. It can also vanish between a survey and the income statement. Stop the claim at the last link the evidence supports.