You will be able to pick the steps in a workflow where a language model adds value and where it does not.
Use an AI step only where a person has written free text that a rule can't handle reliably. Everywhere else in a workflow, a rule, formula or template does the job faster, at no extra cost and without mistakes.
Once people learn that automation tools can call a language model, they tend to want one for every step. The over-eager version is a single AI flow that reads the enquiry, decides what to do, writes the reply, updates the sheet, calculates the fees and books the trial lesson. It looks efficient, but in practice it is slow and costs money on every run. It also makes mistakes where a simple rule would never have failed. An AI step is a tool with a particular shape, and you add one only where it earns its place.
A model is good at reading messy human text and turning it into something structured or readable. Four jobs come up again and again:
Sorting: putting a free-text message into a category. Take an invented message from a parent: "my son is struggling with A-math and his prelims are in two months, do you have weekend classes?" A model could sort it as "secondary, exam help, schedule question". A keyword-matching rule would miss many of the ways people phrase the same thing. Extracting: pulling details out of free text into named fields that later steps can use. The same message contains a subject, a level, a timeline and a scheduling preference, and a model could put each one into its own field. Summarising: condensing long text into a few lines for someone. For example, a long renovation email thread could become 3 lines for the site supervisor. Drafting: writing a first version of a reply from the facts in the record and an example of the tone, which a person then edits.
These four jobs share two things. The input varies too much for fixed rules, and a small error can be caught by a person before it matters. If you took the course Working with AI assistants: prompting that gets usable output, you have already seen all four in a chat window. In a workflow, though, nobody watches each answer as it is produced.
Models are poor at a different set of jobs, and inside an automation those mistakes are easy to miss. Lesson 7.3 of Working with AI assistants, "Checking numbers, quotes, sources and code", explains why these errors happen.
The first is exact arithmetic. A model can add up invoice lines wrongly and still sound certain. Any number that must be right should come from a formula or a built-in math step, never from the model.
The second is current facts. A model does not know today's class fees, this term's timetable or whether a slot is still free, but it will produce something plausible anyway. These facts must come from your sheet or your systems and be passed into the AI step. Don't rely on the model to recall them.
The third is anything that must be identical every time. If you ask a model the same question twice, the wording can differ, and that's fine for a draft. Reference numbers, legal clauses and date formats need a fixed rule or template.
These errors are more dangerous inside a workflow because the output moves on to the next step without anyone reading it.
An AI step is an action in a workflow that takes mapped inputs plus a prompt, sends them to a language model, and returns output fields for later steps. Automation tools usually offer it in one of 2 forms. The first is a built-in AI action that uses a model the tool provides. The second is a connection to a model provider using your own API key, which is the access code a model provider gives you when you set up an account with it. With your own key, the provider bills you for usage directly. Many tools offer both forms, so check which ones yours supports and how each is billed.
In either form, the AI step behaves much like any other action. You map inputs from earlier steps, such as the message field from a form, and add a prompt you write. The step returns output fields that you then map into later steps. Everything in lesson 2.2, "Data mapping: passing fields from one step to the next", applies to AI steps too.
Data sent to an AI step leaves your flow and goes to the model provider. If the record contains personal data, the questions from lesson 3.3, "What a tool can see once you connect your accounts", apply to the model provider as well.
Every AI step costs money on every run, through either the tool's usage charges or the provider's charges. Each one also adds another chance of a wrong answer. A rule costs nothing extra and never has an off day.
For each step in your workflow diagram, ask: could a rule do this? If the step deals with fixed fields, routing, filtering, dates, arithmetic or anything that must be identical every time, the answer is yes. Use a rule, formula, math step or template. If the step needs to sort, extract, summarise or draft from messy free text, the answer is no, and an AI step fits. Pass in the current facts it needs, such as fees, timetables and availability, from your sheet or systems. Where the output matters, have a person check it, for example by approving a drafted reply before it goes out.
Mei Ling, an invented example, handles parent enquiries for lessons. She went through her workflow diagram and checked 5 steps. Routing by the level field became a rule, because the form already asks for level. Spotting test entries was also a rule. A rule could not reliably work out what the parent actually wants from the free-text message, so she made that an AI step. Drafting a reply that answers the parent's specific question also became an AI step, with a person approving it. Calculating the call-back date went to a date formula. She ended up with 2 AI steps out of the 5, and in both the input is free text written by a parent.
Now go through each step of your own workflow diagram and mark it as rule-based or AI-suited. For every AI step, write one sentence explaining why a rule would not do the job.
Mark each step in your workflow diagram as rule-based or AI-suited, and write one sentence justifying each AI step.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).