What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
To find out whether AI is delivering value at work, measure a specific workflow against a credible baseline—not a universal productivity percentage. Track time and throughput alongside quality, errors, rework, adoption, costs, risks, and the outcome that matters to the business or its customers. Time saved is potential capacity, not automatically a cash saving or return on investment.
Start with a defined workflow and a decision
Choose one bounded task or workflow, such as drafting a particular type of document or handling a defined category of customer-support request. Record who does the work, which AI system and version they use, and the intended outcome. Avoid combining unrelated tasks into a single average: a tool may help with one kind of work and have little effect—or create new burdens—in another.
As an Amazon Associate I earn from qualifying purchases.
Before collecting data, state the decision the evaluation must inform. For example: should the team expand access, change the workflow, run a longer trial, or stop using the system for this task? Define what success and failure would look like in terms the organization actually values.
NIST notes that “how a given component is measured and evaluated can change based on the context in which the AI system operates.” Its AI measurement and evaluation guidance emphasizes context-relevant evaluation rather than a single metric that fits every system and use.
#1 Best Overall
Establish a useful baseline
Measure how the workflow performs before AI, using the same definitions you intend to use during the evaluation. Depending on the task, a baseline can include:
- Work completed per person or per unit of time.
- Time from starting a task to completion, including waiting where relevant.
- Output quality assessed against a stable rubric or review process.
- Error rates, corrections, escalations, and rework.
- Relevant customer or worker outcomes, such as service experience or overtime.
Record the observation period and the mix of tasks, users, and workload. Note seasonality and other changes—such as staffing, policy, or software updates—that could affect results. Without this context, a before-and-after difference may reflect something other than AI.
Compare AI-supported work with a credible counterfactual
The key question is what would have happened to the same kind of work without AI. If feasible, randomly assign access or introduce the tool in a randomized, phased rollout. That makes it easier to separate the tool’s effect from differences in workload, staffing, or timing.
Recommended Free Tools
When random assignment is not practical, use a comparison group or a time-series design, and explain its limitations. Keep controlled task tests distinct from ordinary field performance: a short, structured test can show what people can do under test conditions, but it does not establish what happens amid daily work, interruptions, and real operating constraints.
Document the evaluation method, measures, test conditions, uncertainty, and benchmarks. NIST’s voluntary AI RMF Measure function and Measure playbook offer guidance on context-sensitive measurement, limitations, and ongoing monitoring. NIST says the AI RMF is being revised, so consult its current framework materials when applying them.
Use a balanced set of measures
Speed alone can reward rushed, lower-quality work; usage alone shows that a tool was used, not that it helped. Choose measures that capture both what changed and whether that change mattered.
| What to measure | Example question |
|---|---|
| Throughput and time | Did the team complete more work, finish tasks faster, or reduce waiting? |
| Quality | Did outputs meet the same review standard, and did quality change? |
| Errors and rework | Were there fewer corrections, escalations, or repeated steps—or more? |
| Adoption and actual use | Who used the system, how often, and for which tasks? Did workers avoid it or use it differently than intended? |
| Worker or customer outcomes | Did service experience, workload, overtime, or another use-case outcome improve? |
| Costs and risks | What did setup, training, integration, operation, and oversight require, and what risks arose? |
Use measures that fit the task, and segment results where differences could matter: by task type, role, experience, or other relevant user groups. An overall average can conceal who benefits, who sees no change, and where quality or risk worsens.
Interpret published results as evidence about particular settings
Studies show that positive effects are possible, but their findings are not interchangeable forecasts for a different workplace. Their tasks, participants, tools, and outcome measures differ.
| Study | What it found | What the result applies to |
|---|---|---|
| Noy and Zhang, 2023 | In a preregistered online experiment, participants using ChatGPT had 40% lower average time and 18% higher output quality. | 453 college-educated professionals completing incentivized, occupation-specific writing tasks, with or without ChatGPT; these results do not establish the expected effect for other tasks or workplaces. Science paper |
| Brynjolfsson, Li, and Raymond, 2023 | The study reported 14% more issues resolved per hour on average; the effect was 34% for novice and lower-skilled workers, with minimal impact for experienced and highly skilled workers. | 5,179 customer-support agents after staggered introduction of a conversational AI assistant. The NBER page lists a 2025 published version in the Quarterly Journal of Economics. NBER paper page |
| Dillon, Jaffe, Immorlica, and Stanton, 2025 | In the second half of the experiment, 80% of treated workers who used the tool spent two fewer hours per week on email and reduced work outside regular hours. Researchers did not detect changes in task quantity or composition from individual-level access alone. | A six-month randomized field experiment covering 66 firms and 7,137 knowledge workers. The NBER page records a November 2025 revision; the AEA lists the study as forthcoming in American Economic Review: Insights. NBER paper page |
The findings illustrate why a local evaluation needs more than a headline time-saving figure. In the field experiment, reported email reductions did not come with detected changes in the quantity or composition of tasks from individual-level AI access alone. A time saving may improve a worker’s day without changing organizational output; whether that is valuable depends on the decision and outcome being evaluated.
Translate time saved into an outcome, not an assumed cash saving
If AI reduces time per task, find out what happened to the released capacity. It might support additional output, better quality, shorter waits, less overtime, or a less burdensome workload. If none of these outcomes is measured, the evaluation has shown a time difference—not yet a business benefit.
Build the cost side of the case as well. NIST’s industrial AI evaluation procedure calls for considering baseline risk, installation and operating costs, risks of operating the system, estimated value, and a risk-based investment analysis using business metrics. NIST’s overview of the procedure is a useful reminder that an apparent productivity gain should be assessed alongside the resources and risks needed to achieve it.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errors- Include implementation, training, integration, operating, and oversight costs when they apply.
- Track whether released time resulted in a valued outcome rather than converting every saved minute into a cash figure.
- Compare benefits with costs and risks for the stated decision, not against an invented universal ROI threshold or payback period.
Monitor risks and reassess as conditions change
Measure accuracy and reliability, as well as context-relevant concerns such as privacy, security, and bias. Document what the chosen metrics miss, give workers a way to report problems, and review whether the system changes the work in unexpected ways. Reassess when the model, workflow, user population, or operating context changes.
NIST’s ARIA program describes three evaluation levels: model testing, red-teaming, and field testing. Its ARIA Pilot Evaluation Report and ARIA program overview provide context for evaluating AI risks and impacts beyond a single performance score.
Report what the evaluation can—and cannot—show
For the decision-maker, report the measured effect, comparison method, uncertainty, tasks and people covered, adoption, costs, risks, and how any released capacity was used. State which outcomes improved, which did not, and what the evaluation cannot conclude. That makes the result useful without overstating what one workflow, time period, or group of workers can tell you about AI at work more broadly.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
