Pick out the steps in a workflow where a language model is worth using and the steps where it is not.
Put an AI step into a workflow only at a point where someone has typed free text and no rule can deal with it reliably. At every other point, use a rule, formula or template. It runs faster, adds no cost and does not get things wrong.
Once people find out that their automation tool can call a language model, many want a model in every step. Picture one AI flow that reads an enquiry, decides how to handle it, writes the reply, updates the sheet, calculates fees and books the trial lesson. It looks efficient. When it runs, it is slow, it charges you each time, and it gets things wrong at points where a simple rule would have been correct. Treat an AI step as a tool with a particular shape, and add one only where it earns its place.
Language models are strong at taking messy text written by people and making it structured or readable. In an automation, this shows up as four jobs. If you have done the course Working with AI assistants: prompting that gets usable output, you have already met all four in a chat window.
Sorting is placing a free-text message in a category. Take a parent who writes, "my son is struggling with A-math and his prelims are in two months, do you have weekend classes?". A model can file that under "secondary, exam help, schedule question". Rules that match keywords fail on many of the different ways people word the same request.
Extracting is lifting details out of free text and putting them into named fields for later steps to use. The parent's message holds a subject, a level, a timeline and a scheduling preference, and a model can place each of these in a separate field.
Summarising is cutting long text down to a few lines for a particular reader. For example, a long renovation email thread can be reduced to 3 lines for the site supervisor (an invented example). Drafting is producing a first version of a reply from the facts in the record and a sample of the tone you want. A person then edits it.
In all four jobs, the input varies too much for fixed rules to cope with, and a person has room to spot a small error before it causes trouble.
When you use a chat window, you read every answer. Inside a workflow, no one reads each answer as it comes out, and the output passes straight to the next step. Models are poor at three jobs, and inside an automation their mistakes on these 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 total invoice lines incorrectly and still state the result with confidence. Any figure that has to be right should come from a formula or a built-in math step, never from the model.
The second is current facts. The model has no knowledge of today's class fees, this term's timetable or whether a slot is open, yet it will still give an answer that sounds reasonable. Pull those facts from your sheet or systems and feed them into the AI step. Do not count on the model to remember them.
The third is anything that has to come out the same every time. If you put the same question to a model twice, the wording may change. For a draft that does no harm, but reference numbers, legal clauses and date formats need a fixed rule or template.
An AI step is a workflow action that takes mapped inputs and a prompt, passes them to a language model, and hands back output fields for later steps to use. Most automation tools provide one in 2 forms. One is a built-in AI action running on a model the tool supplies. The other is a connection to a model provider through your own API key, the access code a model provider issues when you open an account with it. When you use your own key, the provider bills you directly for what you use. Many tools support both forms, so find out which ones your tool has and how each is charged.
Whichever form you use, the step works like any other action. You map inputs from earlier steps, for example the message field from a form, and then write a prompt. The step returns output fields, and you map those into the steps that follow. All of Lesson 2.2, "Data mapping: passing fields from one step to the next", holds for AI steps as well.
Anything you pass to an AI step leaves your flow and goes to the model provider. If the record includes personal data, ask the questions from Lesson 3.3, "What a tool can see once you connect your accounts", about the model provider too.
Each AI step is charged on every run, through the tool's usage charges or through the provider's. Each one is also one more place where the answer can be wrong. A rule costs nothing extra and never has an off day.
Work through your workflow diagram and ask this of each step. If the step handles fixed fields, routing, filtering, dates, arithmetic or anything that must be identical every time, the answer is yes, and it gets a rule, formula, math step or template. If the step has to sort, extract, summarise or draft from messy free text, the answer is no, and an AI step suits it. Give every AI step the current facts it relies on, 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 is sent.
Mei Ling (an invented example) deals with parent enquiries about lessons. She took her workflow diagram and checked 5 steps. Routing by the level field became a rule, since the form already asks for the level. Spotting test entries also became a rule. Working out what the parent actually wants from the free-text message became an AI step, because no rule could do it reliably. Drafting a reply that answers the parent's specific question became an AI step as well, with a person approving the reply. Calculating the call-back date went to a date formula. She ended up with 2 AI steps out of 5, and both take free text written by a parent as their input.
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.
Go through your workflow diagram and mark each step as rule-based or AI-suited. For each AI step, write one sentence saying why.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).