Which AI Opportunity Should a 20-Person Company Investigate First?

Most 20-person companies don’t need six AI pilots. They need one defensible reason to start with one.

Here is my worked answer, followed by the worksheet that produces it.

The worked answer: start with document or inbox triage

For the representative company in this worksheet, investigate document or inbox triage first. It scores 86 out of 100 because it is frequent, mostly reversible, based on authorized information, easy to keep in draft mode, and suitable for a measurable human-reviewed test.

That is a situated choice, not a universal best AI use case. If your company has stable historical demand and a real forecasting owner, the same worksheet makes forecasting the winner. If the first action is irreversible or high consequence, a high score cannot rescue it from a veto.

Microsoft’s business-envisioning guidance starts with the problem, opportunity, business objective, success measure, and accountable owner, then compares business, experience, and technology viability. The Microsoft guidance is written for ISVs, but its discipline fits a small-company shortlist.

When I taught product managers who moved from writing specs to building and shipping, the recurring failure was often that nobody could say what “done” meant. That is my bounded F-pms observation from Marius Manolachi’s locked entity facts, not a measured rate. For opportunity selection, “done” means a baseline, a reviewable output, an owner, and a stop rule.

If you are building a wider backlog, use this page as a spoke to the AI opportunity pillar and compare it with the broader small-business AI prioritization guide.

What should the scorecard measure?

Score the work, not the tool. Use seven dimensions with weights that total 100.

Dimension	Weight	Score 1	Score 3	Score 5	Evidence to collect
Business impact	20	No named business result	Helps one team, result is indirect	Directly improves a named objective	Objective, stakeholder, and baseline measure
Task frequency	15	Less than weekly	Several times per week	Multiple times per day or a persistent queue	Ten-business-day sample or system log
Reversibility	15	Error is hard to undo	Human review or partial rollback exists	Read-only, draft, or easy undo	Action boundary and rollback path
Data readiness	15	Data is unavailable or unauthorized	Data exists but needs access or cleanup	Authorized, findable, consistent sample exists	Source, permissions, format, and gaps
AI fit	15	Deterministic rules are plainly better	AI helps one bounded step	Language, retrieval, classification, or prediction is central and testable	Task decomposition and non-AI alternative
Owner capacity	10	No accountable owner	Owner exists but review time is unclear	Named owner has time and the needed skills	Owner, review hours, and escalation path
Measurement ease	10	No credible baseline	Proxy is available	Baseline, threshold, and cadence are explicit	Unit of analysis and success rule

The categories translate Microsoft’s strategic fit and BXT questions into a compact worksheet. Microsoft explicitly covers business value and change timeframe, user personas and change resistance, operational risk and safeguards, and AI/LLM fit. OECD’s SME paper adds a useful constraint: data, skills, connectivity, and finance are adoption enablers, so a promising idea without data or an owner is not ready just because the model can do the task. Frequency is an operational measure grounded in the recent-work evidence a small company can collect, with the San Francisco Fed’s reported use cases and capacity barriers explaining why recurring work deserves attention. Reversibility is an operational form of Microsoft’s change and safeguard questions and NIST’s Manage function, which supports documenting a response and disengaging when outcomes do not match intended use.

How do you apply the worksheet?

Use this sequence for every candidate:

1. Write the current workflow and the result it should improve.
2. Collect the evidence named in the table. Do not award a 4 or 5 from enthusiasm alone.
3. Score each dimension from 1 to 5. Explain every score in one sentence.
4. Calculate weighted score = sum(score × weight) ÷ 5.
5. Apply the vetoes below. A veto is not a penalty. It removes the candidate from the first-investigation shortlist.
6. If two candidates are within three points, choose the more reversible one. If still tied, choose lower risk exposure, then the cheaper and faster investigation.

Risk vetoes

Veto question	If yes, what happens?
Would a wrong result affect employment, credit, legal rights, health, safety, payments, or a material customer entitlement without qualified human review?	Veto until the qualified reviewer and approval point are explicit.
Is the input data unauthorized, sensitive, or subject to unclear privacy or intellectual-property rules?	Veto until the data boundary and approved handling are documented.
Can the system send, publish, delete, purchase, promise, or change a record without a reversible approval step?	Veto the autonomous action. Narrow the investigation to read-only output or drafts.
Is there no accountable owner, baseline, or stop condition?	Veto the pilot. Prepare the missing evidence first.

This follows the shape of NIST’s AI Risk Management Framework: Govern the responsibility and policy, Map the context and harms, Measure performance and risk, then Manage the response and the decision to continue or disengage. NIST’s AI RMF Core is voluntary and not a checklist, but it is a useful guardrail against treating risk as a small deduction from a large score.

How do six representative opportunities rank?

The following is a worked example for a 20-person services company with a shared inbox, recurring operational documents, a modest knowledge base, no dedicated data science team, and an owner who can review a pilot for a few hours per week. These are assumptions, not client data or measured rates.

Opportunity	Impact	Frequency	Reversible	Data	AI fit	Owner	Measure	Score	Veto result
Internal knowledge retrieval	3	4	5	3	4	4	3	74	Pass with citations and read-only access
Document or inbox triage	4	5	4	4	5	4	4	86	Pass if routing and drafts stay human-approved
Customer-service assistance	4	5	4	4	5	3	4	84	Pass for suggestions; narrow high-impact cases
Sales research	3	4	5	3	4	4	3	74	Pass with approved sources and no automatic outreach
Marketing production	3	4	5	4	5	5	3	82	Pass for drafts; review claims and publication
Forecasting	5	2	2	3	2	2	2	55	Veto in this context because the owner and review path are weak

The winning arithmetic is transparent: (4×20 + 5×15 + 4×15 + 4×15 + 5×15 + 4×10 + 4×10) ÷ 5 = 86.

The candidate set is not arbitrary. The San Francisco Fed reports small-business AI applications across productivity, marketing, written communications, visuals, customer service, analytics and forecasting, while also reporting barriers involving policy, cost, staff time, training, implementation knowledge, accuracy, and intellectual property. Its 2026 brief is qualitative context, not a ranking of what your company should do. A separate Intuit QuickBooks survey of more than 2,200 US businesses with up to 100 employees similarly reports marketing, customer service, administrative work, data processing, and bookkeeping as common uses. That survey is descriptive and commissioned by Intuit, so I use it as a cross-check, not as independent proof of priority.

When should a different company choose forecasting?

Change the context, then recalculate. Consider a 20-person subscription business with several years of stable demand history, a finance or operations owner who reviews weekly forecasts, and a forecast used only to prepare a human planning decision.

Opportunity	Impact	Frequency	Reversible	Data	AI fit	Owner	Measure	Score	Result
Document or inbox triage	3	4	4	3	5	2	3	70	Still viable, but owner capacity is weak
Customer-service assistance	4	5	4	4	5	3	4	84	Viable for reviewed suggestions
Forecasting	5	5	3	5	4	4	5	89	First investigation candidate

Forecasting wins here because frequency, data readiness, owner capacity, measurement, and reversibility changed. It is still advisory. It does not place an order, change a price, or commit cash. If the team removes that approval boundary, the forecasting candidate hits a veto again.

What should the first investigation brief contain?

For the typical context, keep the first test narrow: classify incoming documents or inbox items, suggest urgency and owner, and draft a next step. Do not let it send, delete, pay, publish, or update a system of record.

Brief field	Filled investigation choice
Hypothesis	AI-assisted triage can reduce handling time without increasing critical misroutes.
Baseline metric	Median minutes from receipt to correct queue, measured on 40 recent historical items. Record current routing accuracy and critical misroutes as secondary measures.
Sample	40 items, stratified across the four largest categories, with the category and correct queue labeled by the current owner.
Mode	Shadow mode first. Produce a classification, urgency, owner suggestion, and draft next step without writing to systems or contacting anyone.
Human approval point	The named operations owner approves the queue, urgency, and any draft before routing, external communication, deletion, payment, or commitment.
Success threshold	At least 25% lower median handling time and at least 95% correct routing, with zero unreviewed high-consequence actions. This is a proposed threshold, not a measured result.
Cost fields	Model and tool cost per item, fixed monthly cost, reviewer minutes per item, correction minutes, and total cost per accepted item.
Latency fields	Model p50 and p95 latency, end-to-end time to a reviewable result, queue wait, and human approval time.
Stop condition	Stop immediately for unauthorized data transfer, an unapproved external action, or a high-consequence misroute. Stop and redesign after two consecutive review batches below 95% routing accuracy or when the full review path fails to reach the 25% time reduction.
Decision after sample	Continue to a bounded live shadow test, narrow the categories, or stop. Do not expand because the demo looked good.

The brief turns Microsoft’s success-measurement and accountability questions into an operational test. It also gives NIST’s Measure and Manage functions something concrete to inspect. The sample is deliberately small enough for a 20-person team to label and review, but it is not large enough to prove long-term reliability.

What should you remember before copying the winner?

The winner is not “AI triage.” The winner is the candidate that best fits a stated context after value, recurrence, reversibility, data, AI fit, owner capacity, measurement, and risk have been checked.

This worksheet has important limits. Its scores are illustrative assumptions, not measurements from a real company. It does not estimate ROI, model accuracy, legal compliance, or adoption. The surveys describe reported uses and barriers, not outcomes for your team. Change the weights when your strategy, data, owner, or risk tolerance changes.

Start with the candidate that gives your team the fastest trustworthy learning loop. For the representative company above, that is document or inbox triage in shadow mode. For the alternate company, it is forecasting with a human approval boundary. The first investigation is a decision about evidence, not a permanent commitment to a technology.

Continue with a related field note

• Should We Investigate an AI Opportunity Before Training the Team?