Start with a recurring task that has clear inputs, a result someone can check, and mistakes you can correct. Measure the work, choose the smallest useful workflow, and keep a person responsible for the outcome.
AI automation for a small business can begin with sorting incoming requests, preparing a project handoff or turning notes into a draft. The first decision is which piece of work deserves to be automated. A complicated process with unclear ownership will still be complicated after you connect an AI model to it.
This guide gives you a way to choose a first project, estimate the time it might recover and write a pilot brief. The worksheet is available without an email address.
01 / Choose a task you can describe
Write down three tasks someone on your team repeats. Be specific about where each task starts and ends. “Automate sales” is too broad. “Read a new inquiry and prepare an internal summary for the right person” is a workflow you can inspect.
For each candidate, record what arrives, what someone does with it, where the result belongs and who checks it. Spend a few days observing the actual work. Count the repeated requests and time a representative sample, including awkward cases.
When a new business inquiry arrives, extract the requested service and deadline, prepare an internal summary, and suggest an owner. A person reviews the summary before the customer receives a response.
That is an illustrative workflow, not a claim about a completed Geronco client project. It has a clear trigger, a limited output and a place for a person to intervene.
Prefer work where delays are visible and the same steps happen repeatedly. If the process changes every time, document the decisions first. You may discover that a clearer form or a shared checklist solves much of the problem.
02 / Compare practical starting points
The examples below describe narrow starting scopes. They are options to evaluate against your own process, rather than a recommendation to automate every category.
| Recurring work | First useful scope | Human checkpoint |
|---|---|---|
| Incoming inquiries | Summarize the request and suggest a destination. | Confirm routing and the reply. |
| Meeting follow-up | Turn approved notes into a draft task list. | Confirm commitments, owners and dates. |
| Project handoffs | Check required fields and assemble an internal brief. | Resolve missing or conflicting information. |
| Recurring reporting | Collect approved figures and flag missing inputs. | Reconcile totals and approve the explanation. |
| Support requests | Suggest a category and an answer from approved material. | Check the draft before sending. |
A useful boundary is to automate preparation before automating commitments. Drafting an internal summary is easier to inspect than allowing a system to promise a delivery date, change a price or approve a refund.
For the first pilot, keep medical, hiring, credit and other consequential decisions with an appropriately qualified person. Choose an operational task where you can review the output and correct a mistake before it affects someone.
03 / Decide where AI earns its place
Ordinary automation follows defined rules: when a form arrives, create a task; when a date passes, show a reminder. AI can help with an interpretation step, such as categorizing a free-text message or turning notes into a draft.
Use a fixed rule when the answer can be expressed reliably as a condition. If a customer selects “website” from a form, routing by that field may be enough. If the customer writes a paragraph describing a mixed request, an AI-assisted classification step may be worth testing.
Use a rule
The input is structured and the correct action is defined. Copy a field, check a required value, calculate a total or route a selected category.
Test an AI step
The input needs interpretation. Summarize notes, suggest a category or draft a response from an approved source.
Keep exact calculations and final business permissions in ordinary code or application rules. Ask the model to produce a proposal, then check that proposal before another system acts on it.
Anthropic’s engineering guidance recommends starting with the simplest workable solution and adding complexity when it is needed. That supports a small workflow with one useful AI step before considering a more autonomous agent. Read the engineering guidance.
04 / Use a five-question scorecard
Compare candidates with the same questions. These questions are a planning aid, not a validated scoring model. A difficult consequence can outweigh an otherwise promising score.
- Does the work repeat? Record its actual frequency. Weekly volume matters more than how impressive a demo looks.
- Are the inputs available? Confirm that the required information exists and that you have permission to use it.
- Can someone check the output? Define a correct result and identify the person who can judge it.
- Can you recover from a mistake? Keep the original request, make changes traceable and provide a manual fallback.
- Does someone own the workflow? Name the person who handles exceptions and maintains the process.
Choose a task that passes all five checks for the first pilot. Where a check fails, fix the missing input, owner or recovery path before building. Compare passing candidates using the time worksheet below.
Download the scorecard & pilot worksheet05 / Count the time that stays
Automation still takes review and maintenance. Estimate the weekly time recovered after those tasks, then replace your estimates with observations during the pilot.
(current work − review − exception rework − upkeep) ÷ 60
Current work is weekly task volume multiplied by manual minutes per task. Review time is the same volume multiplied by review minutes. Exception rework is weekly volume multiplied by the exception rate and the extra minutes per exception. Add upkeep once per week.
Weekly time worksheet
Enter estimates to compare one candidate. Values stay in this page.
Planning estimate. Excludes build time, tool costs and the value or cost of errors. Exception minutes are additional to normal review.
For illustration only: 40 tasks at 6 manual minutes each take 240 minutes per week. If every task needs 2 minutes of review, 10% need 5 extra minutes of correction, and upkeep takes 20 minutes, the estimate is 120 minutes recovered per week. Those numbers are invented to demonstrate the arithmetic; they are not customer results.
Include setup work and subscription or usage costs when deciding whether the project is worthwhile. Time released also needs a purpose: serving customers, reducing a backlog or spending less of the evening on administration. Hours recovered do not automatically become revenue.
06 / Run a bounded pilot
Write a one-page brief before connecting tools. It should describe the task, the exact output, the review owner and what happens when the workflow cannot complete a request.
Your pilot brief
- Trigger
- What event starts the workflow?
- Inputs
- Which approved fields or documents can it read?
- Output
- What must the draft contain, and where does it go?
- Review
- Who approves it before an external action?
- Exceptions
- What must stop or route to a person?
- Success
- What time and quality measures decide whether to continue?
- Fallback
- How does the team return to its existing process?
Test against examples that represent the work: a complete request, missing information, a duplicate and an ambiguous message. Check access failures and a model response that does not match the expected format. Use approved test data and avoid moving private customer information into a new tool without reviewing that use.
Begin with drafts or internal suggestions. Record each output, the correction it needed and the time spent reviewing it. Keep the original request available. If a workflow writes to another system, use a stable identifier to avoid creating a second record when the same event is retried.
Agree on the review period and success criteria before the pilot starts. Choose them around your volume and the consequences of an error. Expand only when the observed quality, time recovered and maintenance burden justify it.
07 / Choose how to build it
After the brief is clear, compare ways to deliver it. Check an existing app’s settings first. A built-in rule may cover the entire job. An integration can connect systems when the workflow spans more than one app. A custom build becomes relevant when the data model, permissions or review process need behavior existing tools cannot provide.
Ask who will maintain each option, how failures become visible and whether you can export the records. Review access permissions and the total ongoing cost. A demonstration that handles one clean example is a starting point for testing.
For context on William’s approach, explore AI strategy and implementation and selected work across Cacele and WebElevated. The pilot brief gives that discussion a concrete starting point.
Common questions
What should a small business automate first?
Start with a frequent, well-defined task whose output a person can check and whose mistakes can be corrected. Draft preparation, internal request routing and structured handoffs are candidates to evaluate.
Does every automation need AI?
No. Fixed conditions often work well for structured data. Test AI when a step requires interpreting language, summarizing information or drafting material for review.
How much does AI automation cost?
The total depends on the workflow, tools, usage, setup and maintenance. Write the scope first and compare the ongoing cost against measured time and quality improvements.
How long should the first pilot take?
Use a review window that captures enough representative work to test normal and awkward cases. A low-volume workflow needs a longer observation window than a task repeated many times a day.
Sources & further reading
The scorecard, worksheet and numerical example above are original planning aids prepared for this guide. They are not research findings or a completed client case study.
- Anthropic: Building effective agents — engineering guidance on workflows, agents and starting with simple patterns. Published December 2024; its tooling context has since evolved.
- Microsoft: Small business automation — background on use-case selection and manageable initial projects.