Elena has three jobs waiting when she opens the service desk: total yesterday's invoices, answer a customer's unusual question, and decide whether to make an exception to a refund policy.
She could hand all three to an AI assistant. But each job asks for a different kind of help.
A spreadsheet can total the invoices. AI can draft an answer to the question. The refund exception belongs to someone with authority to weigh the circumstances and make the decision.
Elena and Mesa Equipment Service are fictional. We will use this small business throughout the opening chapters to show how a workflow develops.
Your result from this chapter: a task map showing where to use ordinary software, AI assistance, human judgment, or a combination.
What “AI” means in this guide
AI is a broad field. Businesses already use systems that classify transactions, forecast demand, recommend products, or detect patterns. This guide begins with generative AI: systems that produce text and other outputs in response to instructions and information.
A prompt is the instruction and context you give the system. A model is the system that generates or evaluates the response. The app around it may add search, file access, memory, and tools.
Those distinctions matter. A chat assistant that can discuss your calendar may have no access to it. Another application may have permission to read appointments. A third may be able to create or cancel them. The familiar chat window does not tell you which powers are available.
Before relying on a capability, check the actual application and account.
Ordinary automation follows a defined process
Consider this rule:
When an approved order is marked paid, create a packing task.
If the order status and the rule are clear, a standard integration may handle the job. The task does not need an AI-generated interpretation of what “paid” means.
Other good candidates include calculating tax from an approved rate, sorting records by date, reminding someone about a deadline, and checking whether a required field is empty.
Ordinary software still needs correct rules, inputs, permissions, and error handling. Its advantage in these examples is that the desired operation can be specified directly.
When a spreadsheet formula solves the problem clearly, use it.
AI assistance helps with varied information
Customer messages rarely arrive in a standard form. One person asks, “What's the inspection charge?” Another writes three paragraphs about a broken machine and mentions the same question near the end.
AI can help identify the question and draft an answer from the approved policy. It can also turn rough notes into a structured summary, suggest ways to explain a service, or compare the wording of two proposals.
The output still needs checking. A fluent response may omit an exception, blend two documents, invent a detail, or misunderstand the request. NIST's generative-AI risk profile includes confidently presented false content among the risks organizations should address. NIST Generative AI Profile.
Use AI where its ability to work with varied language helps, and where the workflow provides a way to verify the result.
People supply judgment and authority
A refund exception may involve more than policy text. What was promised? Did the business make a mistake? What remedy is fair? Who is allowed to approve it?
AI can organize the facts and suggest questions. A responsible person must decide how those facts apply and authorize the outcome under the business's rules.
The same principle matters in hiring, disciplinary action, financial commitments, and other decisions that materially affect people. The issue is not whether AI can produce a recommendation. It is whether the process has adequate evidence, appropriate review, and someone accountable for the decision.
Do not turn “the assistant recommended it” into the explanation for a business action.
Compare the approaches
| Task | A sensible starting point | What to check |
|---|---|---|
| Total a column of approved invoice amounts | Spreadsheet or accounting software | Correct records, formulas, currency, and date range |
| Draft a reply from a current service policy | AI-assisted draft with human review | Facts, exceptions, tone, and missing information |
| Send the same approved reminder on a known date | Ordinary automation | Correct recipient, consent where applicable, and schedule |
| Summarize a long supplier proposal | AI-assisted summary | Important terms against the original document |
| Approve a refund exception | Authorized person, optionally assisted by an evidence summary | Policy, circumstances, approval limits, and record of the decision |
| Update an order after an approved decision | Controlled system action | Permission, correct record, confirmation, and duplicate handling |
These are starting choices, not universal rules. A complicated calculation may need specialist review. A familiar email may be best handled with a saved template. A useful process can combine several approaches.
Break the job into steps before choosing tools
“Handle customer service” is too broad for a first project. Break it down:
- Read the request.
- Identify the question.
- Find the relevant policy and order information.
- Prepare a proposed answer.
- Check facts and commitments.
- Send the approved message.
- Record the outcome.
At Mesa Equipment Service, Elena starts by using AI for steps two and four. She supplies the policy, checks the reply, and sends it herself. Customer records remain in the business's normal system.
This makes the first experiment manageable. She can see whether drafting improves before giving the application access to more information or actions.
Understand the jump from answering to acting
An AI agent is commonly used to describe a system that uses a model to choose steps and tools while pursuing a task. The term is used loosely, so ask what the system actually does.
Can it read files? Write to customer records? Send messages? Spend money? Does the application enforce an approval, or does a prompt merely ask the model to seek permission?
More permissions create more ways for an error to affect the business. OWASP identifies excessive agency—too much functionality, permission, or autonomy—as a risk for applications using language models. OWASP LLM application risks.
Keep your first project easy to observe. Add actions only when you can control, test, and recover from them.
Try a ten-minute task map
Choose one recurring job and write its individual steps. Beside each step, answer:
- Is there a clear rule or calculation?
- Does this step require interpreting varied information?
- Does it create a commitment or affect someone materially?
- How would we know if it went wrong?
Then assign a starting method: ordinary software, AI draft, human decision, or controlled action after approval.
If nobody can describe a good result or detect a bad one, improve the process before adding AI. If information is missing, fix the information problem. If the task is already handled well by a simple tool, spend your attention elsewhere.
Your next step: use this map to choose one small project in Chapter 2.