How to start with AI and automation in a small business

Choose the task before choosing the tool

List the repeated tasks your team handles: sorting enquiries, finding approved business information, drafting replies, copying form details or preparing a handover. Pick one with a clear beginning, an observable result and someone who can judge whether the result is useful.

Write down the current process with the person doing it. Where does the information arrive? Which facts are checked? What decision happens next? Where is the result stored? What happens when something is missing? This often reveals a straightforward improvement before AI is involved.

A useful first project has stable source information and a manageable consequence if a draft is wrong. If the business cannot explain the correct process, first clarify that process. Automation can repeat an unclear rule as efficiently as a useful one.

Separate rules, assistance and permission to act

Ordinary automation follows defined rules. AI can help interpret variable text or draft material, but the business still needs to decide what may happen with the output. A summary, a suggested answer and an action that changes a customer record are different levels of responsibility.

Fictional pilot: preparing a tour-enquiry handover

This is a fictional planning example, not a deployed client system. A tour business receives varied enquiries about dates, party size and meeting points. Staff repeatedly read the messages, identify missing details and prepare a response.

A small pilot could extract the requested date and party size into a draft handover, list missing information and suggest a reply from approved tour notes. A person checks the output and sends the message. The booking provider remains the source of availability; the assistant does not invent available spaces or confirm a reservation.

The team can compare the reviewed handover with its normal process. It should record corrections, missed facts and the time needed for review, rather than count a draft as a success simply because text was generated. If the draft creates more correction work than it removes, simplify the task or stop the pilot.

Agree a pilot brief and acceptance tests

The proposal should identify the input, expected output, business owner, approved information and connected systems. Define the pilot’s limits and review points. A demonstration with carefully prepared inputs is useful for discussing an idea, but it is not evidence that the system handles the range of everyday business messages.

Use representative examples and known expected outcomes. Include missing dates, contradictory details, a duplicate message, an unsupported request and a supplier outage. The number of examples should reflect the workflow’s variety and consequence, rather than an arbitrary claim that a small sample proves reliability.

NIST’s AI Risk Management Framework is a voluntary resource for considering AI risks during design, use and evaluation. For a small business, the practical question is specific: what failure could occur in this task, who would notice, and what would the business do next?

• Correctness: which facts or decisions must be accurate, and who checks them?
• Exceptions: what is referred to a person instead of guessed?
• Actions: what may be drafted, saved, sent or changed, and by whom?
• Failure: how does the team continue if the AI or a connected service is unavailable?
• Recovery: how are duplicates, incorrect records or unintended actions identified and corrected?

Control information, access and operating costs

Identify the information the task needs and avoid connecting unrelated business records just because the tool supports it. Agree where data is processed or stored, who can see it, how long it is retained and what can be exported or deleted. These are questions to resolve with the chosen provider and the business’s requirements, not assumptions to leave until launch.

Give the workflow only the access required for the agreed job. An enquiry assistant does not automatically need permission to change prices, move money or access every customer record. Decide which actions need a person’s approval and retain an understandable record of what was approved or changed.

OWASP describes prompt injection as a way that input can steer a language model in unintended directions, including through external content. Do not rely on a reassuring prompt alone to protect business actions. The implementation needs permission limits and checks around connected tools, with human review where the consequences warrant it.

Ask for the recurring cost model: tool subscriptions, usage charges, hosting, integration services and support. Agree how usage is monitored, who receives alerts and what happens at the agreed spending limit. A low-cost demonstration is not a complete estimate for a workflow used every day.

Measure the result and assign ongoing ownership

Compare the pilot with the current process on the same kinds of tasks. Useful measures can include review time, correction rate, successful completion, handover clarity and how often a person must intervene. Keep failed and unusual cases in the review. Do not treat an isolated impressive answer as a business outcome.

If the pilot is worth continuing, name the person responsible for its information, access and performance. Approved service details change, tools are updated and connected systems can fail. Support should state who checks those changes, responds to incidents and authorizes adjustments.

Keep a handover record of accounts, workflows, approved sources, prompts or configuration, known limits and recovery steps. Clarify which parts you can export or transfer. Retain a usable manual route so the business is not forced to continue with a system it no longer trusts or needs.

Choose a realistic starting engagement

A workflow review, small automation pilot or practical coaching session can fit a business that has a repeated task and a staff member ready to take ownership. A knowledge assistant may fit when approved information is available and answers can be checked. The first useful result may be a clearer process rather than a fully automatic system.

An unattended system that can make important commitments is a larger engagement. It needs specific implementation evidence, authority boundaries, verification and ongoing operation. Caribbean AI Vision is shown in this portfolio as a scripted prototype; it should not be treated as proof of a deployed autonomous workflow or real-world camera accuracy.

Bring an anonymized example of the task, the current steps, approved business information and the person who handles exceptions to a consultation. Ask for a proposed pilot, its limits, acceptance tests, running costs and stop conditions. That gives you a concrete basis for deciding whether AI belongs in the process at all.