Is Your Small Business Ready for AI Automation? A 25-Point Checklist

An owner can be highly capable with AI and still have a business that is unready for automation. The opposite is also true: a team with little prompt experience may have an excellent candidate because the workflow is stable, repetitive, well measured, and already governed by clear rules. Readiness should be assessed for a named job such as classifying inbound requests, drafting a quote summary, or extracting fields from vendor documents—not for AI in general.

Different workflows in the same company deserve different autonomy. Drafting an internal meeting summary is reversible and low consequence. Confirming a price, sending a marketing text, changing a reservation, approving a refund, or giving safety advice can create direct harm. The controls, evidence, and review required should rise with consequence and uncertainty. A green score for one use case does not transfer automatically to another.

The 25-point checklist is intentionally operational. Score each statement yes only when you can show the evidence, not when the team believes it probably happens. Partial means the practice exists inconsistently or cannot be demonstrated. No means the capability is absent. The objective is not a perfect score. It is to find the lowest-cost prerequisite that could invalidate or unlock the test.

A model cannot compensate for an undefined operating process. Start by drawing the workflow from trigger to outcome, including the quiet work that happens in spreadsheets, inboxes, phone calls, and employee memory. Use actual examples to uncover variants. If every employee handles the same request differently, decide which variation is legitimate and which is drift before automating it.

Stable does not mean unchanging. It means the team knows the current states, owners, and approved paths and can intentionally update them. The best early candidate has frequent volume and a recognizable structure, while exceptions are uncommon enough to route to people. A rare process with unique judgment may be intellectually interesting but commercially weak.

AI performance is constrained by the information it receives. A beautifully written assistant will still fail if customer status is stale, two systems disagree on availability, policy lives in an old PDF, or staff enter the same field five ways. Build an inventory of inputs, source owners, update frequency, access permissions, retention, and quality problems. Identify which source wins when records conflict.

Do not make every document available just because retrieval technology can index it. Limit the source set to information required for the task. Separate approved public facts, internal operating guidance, customer records, and sensitive data. Determine whether the vendor retains prompts or outputs, uses them for training, supports deletion, encrypts data, and lets you control access. Free tools may have different terms and controls from enterprise products.

Some decisions should not depend on generated language. Service-area eligibility, approved price tables, business hours, refund limits, required disclosures, opt-outs, and account permissions are usually better represented as deterministic rules. AI is stronger at classification, extraction, summarization, drafting, and ranking where inputs vary and a plausible output can be reviewed or bounded.

Write a responsibility matrix before choosing the model. For each step, state whether a rule, AI, or person acts; what evidence it receives; the confidence or condition required; what it writes; and who can reverse it. Treat low confidence, conflicting data, missing information, and customer frustration as explicit states. Human-in-the-loop means the person has context, authority, and time—not a review button nobody monitors.

A workflow is not production-ready because it succeeds on the happy path. Test what happens when an API times out, the system of record is unavailable, a required field is blank, the model output is malformed, a customer asks for a person, or a vendor changes behavior. The safe fallback should reduce scope, preserve a traceable record, and alert someone. It should not silently drop the lead or continue using stale information.

Operational ownership includes monitoring, incident response, content updates, access review, cost review, and vendor changes. Decide which failures page an owner immediately, which create a daily queue, and which belong in weekly analysis. Keep a manual path and a rollback. If the team cannot operate the system when Dane or the implementer is unavailable, the design has created dependency rather than leverage.

A pilot needs a question that can be answered. Replace improve efficiency with a measurable hypothesis such as reducing median time from missed-call capture to qualified human contact without increasing complaints or staff rework. Establish a baseline from the same workflow, define the eligible cohort, and decide how long or how many cases are needed. If seasonality is strong, a short before-and-after test may mislead.

Count full costs and downstream outcomes. Model tokens are often the smallest component compared with integration, cleanup, review, monitoring, training, exception handling, and vendor minimums. Revenue must be collected and attributable; booking value is not profit. Guardrails such as errors, complaints, opt-outs, customer abandonment, and employee rework can invalidate an apparent gain.

Score yes as one, partial as one-half, and no as zero. A total above 20 suggests the workflow may be ready for a limited test, not full autonomy. A score from 14 to 20 usually means the opportunity is plausible but prerequisite work should be completed first. Below 14, focus on process, data, ownership, or measurement. The absolute score matters less than any no in a safety-critical category.

Use hard gates regardless of total. Do not launch if consequential actions lack deterministic controls, sensitive data use is unreviewed, no human owns exceptions, failures disappear silently, or no outcome can be measured. A low-risk internal drafting assistant can tolerate more manual friction than an external system speaking for the business. Scale readiness standards with customer impact.

The readiness exercise should reduce vendor shopping. Once the workflow, sources, rules, controls, and measures are visible, many products can be evaluated against the same brief. That preserves negotiating power and prevents a vendor's demo from defining your process. The durable asset is the operating specification and evidence, not the first model selected.