AI for Small Business study guide · Lesson 8 of 8
A 30-Day Implementation Playbook
The 30-Day Pilot · Your Small-Business AI Operating System
This is the lesson's study guide from the FlashFeed app, plus its key facts and a self-test.
What you'll be able to do
- Plan a measured 30-day pilot with a baseline, acceptance criteria, parallel run, and evidence-based expansion decision.
- Recognize the policies, approval gates, and audit logs that make AI use dependable.
Pilot
Choose one workflow for a 30-day pilot. A pilot is a limited implementation with a defined owner, scope, period, controls, and success criteria. It exists to create evidence, not to prove that AI is exciting.
In the first week, build a task inventory and choose a frequent, structured, reviewable, low-risk workflow. Record a baseline: current time, cost, error rate, turnaround, backlog, or customer outcome. Define acceptance criteria and the data the pilot may use.
In the second week, create the brief, examples, and checklist. Run a parallel test beside the current process. Record where the AI succeeds, where it fails, what context is missing, and how long review takes.
In the third week, use the workflow with real but limited work. Keep human approval. Track acceptance rate, correction time, errors, and exceptions. Do not expand scope in the middle merely because one result is impressive.
In the fourth week, compare results with the baseline. Include subscription cost, setup time, review time, time saved, avoided cost, and outcome changes in the return on investment calculation. Decide to stop, adjust, or expand.
Operating System
Turn a successful pilot into an owned capability. Write a short AI use policy covering approved tools, prohibited data, review rules, external use, incidents, and responsibility.
Maintain a prompt library with tested instructions, examples, source materials, criteria, owner, and review date. Add an approval gate before consequential actions, an escalation rule for uncertainty and exceptions, an audit log for accountability, and a kill switch for rapid shutdown.
Review the operating system periodically. Check access, vendor terms, source documents, failures, value, and whether the workflow still solves the right problem. Retire workflows that no longer earn their complexity.
For continued learning, prioritize educators who show full workflows, discuss limitations, and distinguish evidence from sponsorship. Ethan Mollick's One Useful Thing offers practical, research-aware guidance. NIST, the SBA, SBDCs, SCORE, and regulator resources provide durable guardrails. OpenAI Academy and other major vendors provide useful product-specific demonstrations; treat those as training for a tool, not neutral rankings of the market.
A weekly scorecard
Use a one-page scorecard throughout the pilot. Record volume processed, total time, AI cost, review time, accepted outputs, corrected outputs, serious errors, exceptions, and an outcome that matters to the business. Notes should explain unusual weeks.
Calculate value conservatively. Time saved is valuable only when it frees capacity, reduces delay, improves quality, or replaces a real cost. Ten minutes saved on work no one needed is not a return. Include setup and maintenance time, not just the subscription.
At the end, choose one of three decisions. Stop when the workflow does not create enough value or the risk is unacceptable. Adjust when the idea is sound but instructions, data, controls, or scope need work. Expand when results repeatedly meet criteria and the owner can support the operating burden.
Choosing people to learn from
Thought leaders can add practical examples, but use a filter. Prefer people who show inputs and outputs, identify failure cases, update old advice, cite research, and disclose sponsorships. Be skeptical of channels built around daily tool rankings, guaranteed income, or dramatic statistics without methodology.
For practical general guidance, Ethan Mollick is a strong starting point. For business education, use SBA, SCORE, and Small Business Development Center programs. For risk, use NIST, FTC, EEOC, and relevant state regulators. For specific products, use the vendor's current documentation and academy resources.
The goal is not to become an AI news expert. It is to build one useful, controlled workflow and develop the judgment to repeat the process.
Key concepts
Task inventory
- A list used before selecting an AI tool.
- It focuses attention on recurring work and bottlenecks.
- It records frequency, time, structure, and error consequence for candidate workflows.
Baseline
- A reference point established before a change.
- It prevents teams from relying only on impressions of improvement.
- It records current time, cost, errors, or outcomes for comparison with a pilot.
Pilot
- A limited implementation designed to produce evidence.
- It reduces exposure while a workflow is tested and refined.
- It uses a defined scope, owner, period, and success criteria before wider rollout.
Parallel run
- A transition technique that preserves an existing process temporarily.
- It makes side-by-side quality and timing comparisons possible.
- The new AI-assisted workflow operates alongside the manual workflow before replacement.
Acceptance criteria
- A quality tool defined before reviewing results.
- It makes approval less dependent on mood or impressive wording.
- It lists the conditions an output must satisfy to be considered usable.
Return on investment
- A comparison between value received and resources spent.
- For a pilot, its inputs can include subscriptions, setup, review time, savings, and avoided cost.
- It helps decide whether to stop, adjust, or expand an AI workflow.
AI use policy
- A governance document sized to the business.
- It gives staff common boundaries before a problem occurs.
- It names approved tools, prohibited data, review rules, external-use expectations, and incident reporting.
Prompt library
- A reusable knowledge asset rather than a single conversation.
- It helps a team repeat successful AI-assisted work.
- It stores tested instructions, examples, source materials, criteria, an owner, and a review date.
Approval gate
- A control placed before consequential action.
- It limits autonomy without preventing drafting or analysis.
- It requires a person to confirm an action before money moves, a message sends, or a record changes.
Audit log
- A record that supports troubleshooting and accountability.
- It helps reconstruct what a workflow did and why review failed.
- It records inputs, outputs, actions, timestamps, errors, and approvals.
Escalation rule
- A boundary for uncertainty, exceptions, or authority.
- It prevents a system from bluffing its way through unusual cases.
- It defines when automation stops and a case transfers to a responsible person.
Kill switch
- An operational safeguard for a malfunctioning workflow.
- It should be quick to use and known to responsible staff.
- It immediately stops automated actions while an incident is investigated.
Key facts
One Pilot Beats Ten Trials
A focused 30-day pilot creates evidence. Tool hopping creates opinions without a stable process or baseline.
Measure Before and After
Record current time, cost, error rate, turnaround, and customer impact before changing the workflow. Without a baseline, improvement becomes guesswork.
Track Usable Output
A useful pilot metric is the share of AI outputs accepted after review, along with the time and type of corrections required.

Keep the First Policy Short
A beginner AI policy can fit on one page: approved tools, prohibited data, review requirements, external-use rules, incident reporting, and an owner.
Every Workflow Needs an Owner
Someone must maintain instructions, permissions, source material, tests, and performance—even when the workflow runs automatically.
Review the System, Not Just the Output
Periodically review tool access, vendor terms, failed outputs, time savings, source documents, prompts, and whether the workflow still solves the right problem.
Self-test
Answer each question from memory, then check your work against the answer key.
1. When should a small business expand an AI pilot?
- AAfter buying an annual plan
- BAs soon as the output looks impressive
- CWhen competitors mention the same tool
- DWhen measured results meet criteria and risks are controlled
2. What best describes a 30-day pilot?
- AA permanent rollout across every team at once
- BA demo meant to show that AI is exciting
- CA time-boxed implementation run under clear ownership and success criteria
- DAn open-ended trial with no fixed end date
3. Which measurement is NOT part of recording a baseline before the pilot?
- AThe final expansion decision
- BCurrent time and cost
- CError rate and turnaround
- DBacklog or customer outcome
4. In the second week, how should the AI workflow be tested against the current process?
- ABy expanding scope before any review
- BBy running a parallel test beside the current process
- CBy replacing the current process entirely
- DBy skipping the current process to save time
5. When building the first-week task inventory, which workflow should you choose for the pilot?
- AWhichever task competitors are automating
- BA rare, one-off task with high stakes
- CA frequent, structured, reviewable, low-risk workflow
- DThe most creative and unpredictable task available
6. In a 30-day AI pilot, when should acceptance criteria be defined?
- AAfter results are already compared to competitors
- BDuring the annual budget review
- COnly after the pilot ends
- DIn the first week before the pilot runs
7. What role do acceptance criteria play in a 30-day AI pilot?
- AScheduling the parallel run's start date
- BDeciding in advance what results count as success
- CChoosing which employee owns the workflow
- DTracking how much the subscription costs
8. What turns an AI experiment into a dependable business capability?
- AA documented workflow that someone owns and reviews
- BRemoving people from every decision
- CA famous prompt
- DMore subscriptions
9. Which item does NOT belong in a short one-page AI use policy?
- AExternal-use rules and an owner
- BA guaranteed monthly revenue target
- CReview rules and incident reporting
- DApproved tools and prohibited data
10. What is the purpose of a kill switch in an AI operating system?
- ATo log every action for accountability
- BTo require human approval before actions
- CTo rapidly shut the workflow down
- DTo store tested prompts and examples
11. An escalation rule tells the workflow what to do in which situation?
- AWhen outputs are routine and low-risk
- BWhen a prompt needs to be rewritten
- CWhen the subscription is due for renewal
- DWhen uncertainty or an exception arises
12. Which control in the operating system exists primarily for accountability?
- AThe audit log
- BThe prompt library
- CThe kill switch
- DThe approval gate
Go deeper
Answer key
- 1. D When measured results meet criteria and risks are controlledExpansion should follow evidence and control, not novelty or sunk cost.
- 2. C A time-boxed implementation run under clear ownership and success criteriaA pilot is a limited, controlled implementation that exists to create evidence, not to prove AI is exciting or to launch everywhere immediately.
- 3. A The final expansion decisionA baseline captures current time, cost, error rate, turnaround, backlog, and customer outcome. The expansion decision comes at the end, after comparing results to that baseline.
- 4. B By running a parallel test beside the current processA parallel run works the AI beside the existing process so you can record where it succeeds, fails, and how long review takes without risking live work.
- 5. C A frequent, structured, reviewable, low-risk workflowThe pilot should target a frequent, structured, reviewable, low-risk workflow so results are measurable and mistakes are recoverable.
- 6. D In the first week before the pilot runsAcceptance criteria should be defined in the first week, alongside the baseline and the data the pilot may use.
- 7. B Deciding in advance what results count as successAcceptance criteria set the standard results must meet, so the later expansion decision is evidence-based, not a guess.
- 8. A A documented workflow that someone owns and reviewsDependability comes from the operating system around the model, not the model alone.
- 9. B A guaranteed monthly revenue targetA beginner AI policy covers approved tools, prohibited data, review rules, external use, incidents, and responsibility. A revenue target is not part of a use policy.
- 10. C To rapidly shut the workflow downA kill switch exists for rapid shutdown. The audit log handles accountability, the prompt library stores prompts, and the approval gate handles human sign-off.
- 11. D When uncertainty or an exception arisesAn escalation rule routes uncertainty and exceptions to a person, rather than letting the AI proceed unchecked on cases it cannot handle.
- 12. A The audit logAn audit log records what happened for accountability. The approval gate blocks consequential actions, the library stores prompts, and the kill switch enables shutdown.
Sources
- OpenAI Academy — Small Business Resource Hub — academy.openai.com/en/public/resources/openai-academy-small-business-resource-hub-2026-06-03
- NIST — AI Risk Management Framework — airc.nist.gov/airmf-resources/airmf/
- U.S. Small Business Administration — AI for small business — www.sba.gov/business-guide/manage-your-business/ai-small-business
Image credits
- Keep the First Policy Short: AI-generated (gpt-image-1.5)