
A useful AI pilot begins with one task and a clear human
decision.
Quick answer: Start with one repeatable,
low-consequence task. Measure how long the current process takes, define
what an acceptable result looks like, use only approved information,
keep a qualified person responsible for review, and count the entire
AI-assisted workflow—not just the time it takes to generate a draft. If
review, correction and coordination outweigh the work removed, adjust
the pilot or stop it.
Many workplace AI experiments begin with a broad instruction: use the
tool and find ways to be more productive.
That sounds flexible. In practice, it can create another assignment.
Employees must decide where the tool fits, learn how to prompt it, check
the output, correct mistakes and explain the process to someone else.
The organization may count the speed of generation while the team
absorbs the supervision work.
A more useful starting point is narrower: remove one planning task,
then prove that the full workflow is better.
This is not an argument against AI. It is a way to keep the tool
accountable to the work.
AI
productivity depends on the task and the context
Research supports a measured approach rather than a universal
promise.
A 2025 randomized field experiment across 66 firms and 7,137
knowledge workers found that, among treated workers who used an
integrated generative AI tool, time spent on email fell by about two
hours per week and work outside regular hours also declined. The
researchers did not detect a broader change in the quantity or
composition of tasks from individual access alone. That is useful
evidence, but it concerns a particular tool, set of applications and
six-month study—not every AI workflow. (NBER Working Paper
33795)
An earlier field study of 5,179 customer-support agents found an
average productivity improvement of about 14 percent, with larger gains
among newer and lower-skilled workers and little effect among the most
experienced workers. Again, the result was specific to a defined task
and workplace setting. (NBER Working Paper
31161)
The practical lesson is not that AI always saves time. It is that
outcomes vary by task, worker, workflow and operating environment.
The NIST Generative
AI Profile treats AI as a socio-technical system: the risks and
benefits depend on the technology, the people using it and the context
around it. NIST’s voluntary AI Risk
Management Framework Playbook organizes that work around governing,
mapping, measuring and managing. A small workplace pilot can follow the
same logic without turning into a large governance exercise.
Start with one
task, not a general AI mandate
The first pilot should be easy to describe in one sentence.
For example:
Use an approved AI tool to turn a weekly list of tasks and fixed
meetings into a first-pass planning outline for human review.
That is different from “use AI to plan your week.” The narrower
version identifies the input, the output and the person who decides
whether the result is useful.
Choose a task with five
qualities
A suitable first task is:
- Repeatable. It occurs often enough to compare
several examples. - Measurable. The team can observe time, corrections,
exceptions and accepted results. - Low consequence and reversible. A mistake can be
caught before it changes a record, creates a commitment or affects a
person. - Safe to perform with approved data. Employees know
what information may enter the tool and what must stay out. - Reviewable. A qualified person can recognize a
satisfactory result and has the authority to reject it.
Avoid beginning with sensitive employment decisions, health
information, legal conclusions, confidential client data,
security-critical actions or any task where the reviewer cannot
independently evaluate the output. The organization’s legal, privacy,
security and records requirements still apply.

Run a one-task AI pilot
in seven steps
1. Describe the current task
Write down what starts the task, who performs it, what “done” means,
where the result goes and what usually causes rework.
Do not improve the description yet. You need an honest baseline.
For five to ten representative examples, record:
- active time from start to accepted result;
- waiting or handoff time when it affects completion;
- common corrections;
- unusual cases;
- who checks the work; and
- what happens when the process fails.
2. State what the pilot will
not do
A clear boundary protects attention and judgment.
For a weekly planning pilot, the tool may organize approved tasks
around fixed meetings. It should not infer a person’s health, energy,
mood or family obligations. It should not decide performance priorities,
send messages, modify calendars or make commitments without review.
Those non-goals matter as much as the desired output.
3.
Define an acceptable result before choosing a prompt
Create a short review checklist. A weekly planning outline might need
to:
- include every fixed meeting;
- include the three stated priorities;
- identify obvious time conflicts;
- leave unallocated space for normal variability;
- avoid inventing deadlines or dependencies; and
- label missing information instead of guessing.
If the team cannot define a satisfactory result, it will not be able
to measure the pilot fairly.
4. Set data and review
boundaries
Use only an organization-approved tool and approved inputs. Remove
unnecessary personal, confidential or regulated information. Name the
reviewer and the situations that require escalation or a return to the
manual process.
Human review should be real responsibility, not a box added after the
fact. The reviewer needs enough time, context and authority to reject an
output.
5. Run the old and
new workflows side by side
For the first few examples, complete the task both ways. This
provides a current comparison and keeps a manual fallback available.
For a nonclinical weekly planning task, a constrained prompt could
be:
Using only the task list, deadlines and fixed meetings below, create
a first-pass weekly outline. Preserve every stated deadline. Do not add
priorities, personal information or assumptions. Flag conflicts and
missing details. Leave at least one open block each day. Return a short
outline for human review; do not send, schedule or change anything.
The manual alternative is simple: draw three columns labeled
must do, could do and can
wait; place fixed meetings first; then assign a small number of
priorities while leaving open space. The pilot should beat this
practical alternative, not an imaginary inefficient process.
6. Measure the whole workflow

Count preparation, review and correction—not only the speed of
the first draft.
Generation time is only one part of the work. A useful measure
is:
Net capacity gained = baseline active time − AI-assisted
preparation, prompting, review, correction, coordination and recovery
time
Track at least:
- total active time to an accepted result;
- first-pass acceptance rate;
- number and type of corrections;
- missed exceptions or invented information;
- reviewer effort;
- downstream rework;
- employee-reported friction; and
- whether the task shifted into evenings or other recovery time.
Do not convert a small pilot into a guaranteed financial return. The
goal is to learn whether this task became measurably better in this
setting.
7. Make a real decision
At the end of the pilot, choose one of three outcomes:
- Stop: The tool added work, created unacceptable
risk or failed to improve the result. - Adjust: The task may be suitable, but the boundary,
inputs, prompt, review standard or ownership needs to change. - Scale carefully: The workflow produced consistently
acceptable results with less total effort, manageable exceptions and
clear accountability.
Scaling should not mean applying the same process to every task. It
means repeating the evaluation with the next bounded use case.
Protect the capacity the
pilot creates
Time saved is not automatically recovery time. An organization can
remove twenty minutes from a task and immediately fill it with more
output, more meetings or another tool to supervise.
If AI genuinely creates capacity, decide where that capacity should
go. It might support focused work, learning, quality improvement, a more
realistic workload or a protected transition at the end of the day.
This is where AI adoption and workforce wellbeing meet.
The World Health Organization recommends organizational interventions
that address working conditions and psychosocial risks rather than
relying only on individual coping skills. (WHO:
Mental health at work) The NIOSH Total Worker
Health approach likewise includes technology, work organization,
leadership and policies in a broader view of worker wellbeing.
Wellness should not be used to help employees tolerate a poorly
designed AI workflow. A breathing exercise may support a difficult
moment; it cannot correct unclear priorities, unmanaged review work or
an unreasonable workload.
For related reading, see AI
Workplace Productivity: Why More Code Is Not the Same as Better
Performance, How
to Get Your Energy Back When You Work a Desk Job and How
Breathwork Can Support Stress, Focus and Workplace Performance.
A manager’s one-task pilot
checklist
Before starting, confirm that:
- the task and non-goals fit on one page;
- the current process has a baseline;
- the input data is approved;
- the output has acceptance criteria;
- a qualified reviewer is named;
- the manual fallback still works;
- employees can report friction without being treated as
resistant; - total effort, corrections and downstream rework will be measured;
and - the pilot has an end date and a stop, adjust or scale decision.
The best first AI pilot is not the most impressive one. It is the one
that teaches the team whether the tool removes useful work while
preserving judgment, quality and reasonable boundaries.

A pilot is complete when the team can make an evidence-based
stop, adjust or scale decision.
Frequently asked questions
What is the
best first task for an AI pilot at work?
Choose a repeatable, measurable, low-consequence task with approved
inputs, a clear standard and a qualified human reviewer. A first-pass
planning outline, document classification draft or meeting-summary draft
may be suitable if organizational policies allow it. The exact choice
depends on the workflow and the consequences of error.
How long should a
workplace AI pilot run?
Run it long enough to include several representative examples and
normal exceptions, but keep the scope bounded. For a weekly task, four
to six cycles may be more informative than a single demonstration. Set
the end date and decision criteria before starting.
How do we
measure whether AI really saves time?
Compare total active time to an accepted result. Include preparing
inputs, prompting, reviewing, correcting, coordinating, handling
exceptions and recovering from failures. Also record quality and
downstream rework. A faster first draft is not a time saving if
verification takes longer than the work it replaced.
Should every AI
output receive human review?
Review should match the task’s risk, reversibility, audience and
consequences. A low-risk draft still needs a responsible person during a
pilot. Higher-consequence decisions require stronger controls and may
not be suitable for generative AI. Follow applicable organizational,
legal, privacy and security requirements.
What if employees
prefer the manual process?
Treat that response as evidence. Ask what creates friction and
compare the complete workflows. The right outcome may be to improve the
boundary, use simpler automation, keep the manual method or stop.
Adoption is not success if the task becomes harder or less reliable.
A restrained next step
for organizations
If technology change is adding friction to your team’s workload,
begin with the work design: what should be removed, what requires human
judgment and where recovery is being lost.
Origins Unity supports organizations with practical programs for
stress, energy, focus, movement and whole-person workplace wellbeing. Explore
Origins Unity corporate wellness programs to discuss a program built
around your workforce and constraints.
This article provides general educational information. It is not
legal, cybersecurity, privacy, employment or medical advice.
Organizations should apply their own policies and consult qualified
specialists where appropriate.

Author bio
David Reveles is the founder of Origins Unity and a
corporate wellness consultant. He brings more than 25 years of
experience in fitness, martial arts and movement coaching, along with
more than a decade running projects and teams. His verified credentials
include NASM Certified Personal Trainer, Oxygen Advantage® Instructor,
Breathwork Facilitator, Certified Workplace Wellness, CAPM, Lean Six
Sigma Black Belt and ITIL. Through Origins Unity, he designs and
facilitates practical wellbeing programs for organizations and
individuals.
