The short answer
Start with one repetitive task that has a clear input, a checkable output, and a person responsible for the result. Document the current process, test with invented customer details, and run a small supervised pilot. Expand only when the work is more reliable or measurably easier.
What makes AI worth introducing?
A useful AI project removes a recurring bottleneck without creating a larger review burden. Look for frequent work, information the business already has, and an output someone can verify. Inquiry summaries, request classification, and draft follow-ups are candidates to assess—not automatic recommendations for every business.
The buying question is whether the whole process improves. A tool that drafts a reply still needs accurate context, permission to use that information, and a person responsible for the next step. Those dependencies should be part of the scope and price.
Choose the problem before the tool
Write down where your team repeats work or loses track of a customer. A useful starting problem is specific: “We read every inquiry and manually copy the job type into our customer system.” “We need AI” is not specific enough to evaluate.
Our recommendation is to choose a task the person doing it can explain in a few steps. Ask them what information they need, what a good result looks like, and what usually goes wrong. If the team disagrees about who owns the task, resolve that before automating it.
A first AI project needs a clear job, access to the right information, and an owner for exceptions. The implementation should connect those three things before adding more capabilities.
Decide whether the task needs AI
Some tasks need a fixed rule, some benefit from interpreting language, and some should stay with a person. Use the simplest approach that handles the job reliably.
| Task | First approach to consider | What a person still owns |
|---|---|---|
| Send an appointment reminder | Scheduled automation using confirmed booking details | Timing, wording, and communication preferences |
| Summarize an incoming inquiry | AI prepares a draft from the actual message | Checking names, requests, and missing details |
| Route a standard service request | A form or fixed rules; add AI only if free-text interpretation is needed | Service boundaries and exception routing |
| Answer a complicated complaint | A person leads; AI may organize the history | The response, commitments, and resolution |
| Approve a price or refund | Existing approval process | Financial decisions and promises to the customer |
These are design recommendations, not a ranking of products. A reminder does not become more useful just because it uses an AI model.
Use five readiness questions
- Is the input available? Identify the inbox, call record, form, or document the workflow starts from. Do not depend on information the system cannot access.
- Can you check the output? Define the fields or actions that must be correct. A convincing summary is not enough if the callback number is wrong.
- Is there a named owner? Choose the person who reviews exceptions and maintains the instructions.
- Can you undo or stop it? Begin with drafts or suggestions. Know how to pause the workflow and return to the manual process.
- Is the information appropriate for the tool? Review what will be sent, who can access it, how it is retained, and whether it is used for training. Use invented examples until you have approved the data handling.
If any answer is unclear, include that preparation in the project scope before selecting a tool.
Run a small, supervised pilot
Choose a limited scope, such as drafting summaries for one inquiry inbox. Keep customer-facing actions under human approval while you learn where the system fails. Include incomplete messages, spelling mistakes, duplicate inquiries, and requests outside your service area in your tests.
For each run, keep the input, the expected result, the actual result, and any correction. Use synthetic details in shared test records. If the workflow contacts customers or changes a booking, test the failure path as carefully as the successful path.
Agree on a baseline, the pilot scope, and the review period before work starts. The period should contain enough comparable tasks to judge the result; a seasonal or low-volume business may need longer.
NIST’s voluntary AI Risk Management Framework treats risk management as part of designing, using, and evaluating AI. Our small-business translation is to assign ownership, test realistic failures, and keep a way to intervene.
Measure the work that remains
Record the time spent doing the task before the pilot. During the pilot, include reviewing drafts, fixing mistakes, and maintaining the workflow. Count incorrect outputs even when a person catches them before they reach a customer.
Track setup and maintenance separately. Time saved is capacity released, not automatically cash saved.
Illustrative example: a task that takes eight minutes manually and three minutes to review and correct releases five minutes per task. Across 60 comparable tasks, that is five hours before setup and maintenance. These numbers are an example, not OneCompass client results.
Also track the share of outputs needing correction and the time unresolved requests wait for a person. Decide what would make you pause the pilot before you start. Do not expand a workflow just because the demo looked good.
Choose the next step
If the pilot produces usable outputs, fewer corrections, and a manageable review workload, expand one part at a time. If it fails, identify whether the problem is missing information, unclear rules, the tool, or a process that needs a person.
For call handling, use the AI receptionist evaluation checklist. If you are unsure where inquiries go missing, begin with the lead-response audit.
Our Bakeris case study shows the wider operational context: inquiries, office follow-up, customer history, and reporting need to work together. This guide does not report an AI-only performance lift from that engagement.
How OneCompass helps
Choose an AI project worth doing.
OneCompass helps identify the task, connect it to the tools your team already uses, and set up the review and follow-up around it. The starting point is your actual inquiry or office process.
Our work with Bakeris connects inquiries, office follow-up, customer history, and reporting. Those connections are the work around the AI that makes the service useful.
See the Bakeris engagementBy Brando, founder of OneCompass. Examples are labeled where illustrative, with client work and sources linked separately. To suggest a correction, email OneCompass with the page and section.
Reference: Brando, “Where to start with AI in a service business,” OneCompass, October 7, 2026. Permanent link.