AI business process automation works best when most of the process is not AI at all. This article answers one question: inside a rule-based workflow, which steps should be handed to an AI Agent action, and which should be decided by an If/Else branch? Our view is firm. Build the process as a workflow of plain rules first, then add an AI Agent step only at the two or three points where a decision depends on reading messy human input, such as a free-text reply or an enquiry that could mean three different things. Everything else stays deterministic, cheap and auditable.
Two kinds of step in any automated process
Every business process is a chain of steps. Some are mechanical: when a form arrives, create a contact; when a deal is won, send the onboarding email; when an invoice is seven days overdue, send a reminder. The input is structured and the right action is always the same. A rule handles these perfectly, every time, at almost no cost.
Other steps need judgement on unstructured input: reading an enquiry and deciding which service it is about, reading a reply and deciding whether the person is interested, pulling a property address out of a rambling email. A rule cannot do these well, because the input varies too much. This is where AI earns its place.
Good AI business process automation is a rule-based spine with a few AI joints. On the platform, the spine is a workflow in Automation, the joints are AI Agent actions placed inside it, and the decisions between them are If/Else branches that read whatever the AI step wrote. This page is about AI as one step in that chain. If you want an agent that holds a whole conversation with a customer by chat, SMS or phone, that is a different tool, covered in AI agents for chat and voice.
The If/Else test: who should decide this step
Before you add any AI, run every decision in your process through one question: could an If/Else branch decide this from a field that already exists? If/Else compares a field against a value with an operator, and it can group several conditions with AND or OR. That covers a surprising amount of real work.
An If/Else branch should decide when the answer lives in structured data. Lead value above a threshold, a tag present or absent, a pipeline stage, a dropdown chosen on a form, a date that has passed. These are exact comparisons, and a branch makes them the same way every time, with a visible condition you can point to when a client asks why they got a message.
An AI Agent action should decide when the answer lives in words. A reply that says "maybe after Easter, can you send pricing first" is not a yes, a no or a question in any field the system holds. The AI step reads it, then writes the result somewhere structured: a tag, a contact field, or an output that the next branch can test. The pattern that works is always the same: AI reads and classifies, If/Else routes.
The common mistake is asking the AI Agent action to decide and act in one go: read the reply, judge interest, send the next email and move the deal. When something goes wrong you cannot tell which part failed. Keep the AI step narrow, have it write one result, and let an If/Else branch take the action from that result.
What most articles on this topic get wrong
Most writing on AI automation describes agents running a process end to end: an agent receives the lead, decides everything, takes every action, and the process "runs itself". It is an attractive picture and a poor design for a firm that cannot afford mistakes.
Each AI decision carries a small chance of being wrong. Chain ten of them and the chance that something in the chain goes wrong becomes large. Rules do not have that problem. A branch that checks whether a tag is present gives the same answer every time.
End-to-end AI processes are also hard to audit. "The agent decided" is not an answer your operations lead, your compliance adviser or your client will accept. The question is never "can AI do this step?" It usually can. The question is "does this step need judgement on unstructured input?" If not, use a branch.
Where an AI Agent step belongs in a typical process
Across service firms, the same handful of AI joints come up again and again.
Classifying inbound messages. An enquiry arrives by email, form or chat. The AI step reads it and sets a field: service type, urgency, new or existing client. An If/Else branch then routes on that field.
Extracting details. A message contains a company size, a postcode or a preferred date in free text. The AI step pulls it into the right contact fields so the rest of the process can use them.
Reading replies. A lead replies to a follow-up. The AI step labels it as interested, not interested, a question or an out-of-office, and the branches take it from there. Conversation history across channels is available to the step without extra setup, so it can read what was said before.
Drafting internal notes. The AI step drafts a handover summary from notes and call transcripts, which a person checks before it is used.
How to set up an AI Agent action and an If/Else branch
You need a workflow with a trigger already chosen, and you need to know which field the AI step will write to. The first block adds the AI Agent action; the second adds the If/Else that reads its result.
- Go to Automation > Workflows.
- Create a new workflow or edit an existing one.
- Add a relevant trigger, such as Form Submitted or Contact Created.
- Click + Add, or select the Please select action placeholder on the workflow canvas.
- Select the AI Agent action from the Workflow AI section.
- Enter an Action Name.
- Select Build Your Own to configure the action from scratch.
- Add your Instructions.
- Click Enhance Prompt if you want to improve the structure of the prompt.
- Choose the Model you want to use.
- Click + Add Tool and select the tools the agent can use.
- Configure each tool, including whether the AI decides field values or whether specific values are set manually.
Then add the branch directly after it:
- Open your workflow, click the + icon and select If / Else from the action list.
- Choose Build My Own to build the logic manually.
- Configure the first branch: select the field to evaluate, choose an operator and set the comparison value.
- Click Add Segment to add more conditions, and use AND or OR to group them.
- Click Add Branch to create additional condition groups, each a separate path.
- Configure the None branch, which is created automatically, by renaming it and adding fallback actions.
The None branch is your safety net. Send anything the AI step could not classify to a person, with an internal notification, rather than letting it fall silently out of the process.
A worked example: reading replies to a quote
Suppose the firm sends about 60 proposals a month. Replies arrive by email and SMS, and today a partner reads each one before anyone knows what to do next. The workflow would start on a reply. An AI Agent action reads the message and its history, then uses an Add Tag tool to apply one of four tags: ready to sign, wants changes, price objection, or unclear. An If/Else branch with four paths routes on those tags. Ready to sign sends the contract and moves the deal. Wants changes creates a task for the partner. Price objection notifies the sales lead. Unclear, plus anything in the None branch, goes to the office manager, who would also spot-check around twenty tagged replies a month by hand.
Notice what the AI does not do. It does not send the contract or move the deal. Those are actions behind a branch, written once and visible to everyone.
A second example: where a branch replaces AI entirely
Imagine the agency plans an AI step to decide which onboarding track a new client needs. Mapping it shows the track depends on only two things: the service bought and the contract value, both already on the opportunity. An If/Else branch with two segments would decide it exactly, at no usage cost. The one AI Agent step worth keeping would be the one that drafts the internal handover note from call transcripts, which a salesperson then edits in a few minutes.
Cost, memory and output settings
Each run of an AI Agent action is billed as usage, based on the model's tokens and any premium tools it calls, so a step placed early in a busy workflow runs far more often than one placed after a filter. Put a cheap If/Else in front of it where you can, so only the contacts that genuinely need reading reach the AI.
Conversation Memory is off by default. Turn it on only for workflows a contact goes through repeatedly, where the agent benefits from what it saw last time. Under Advanced Options you can also set an Output Format: none, text, or a JSON structure that later steps can read. A structured output is the cleanest way to feed a branch.
Troubleshooting
The If/Else always takes the None branch. The AI step probably wrote nothing, or wrote a value that does not match the comparison exactly. Open the Execution Logs tab, select the AI Agent run and check what the tool actually set.
The branch runs before the AI result is there. If/Else evaluates when the contact reaches it. If the value depends on something slower, insert a Wait action before the branch.
Classifications drift over time. Tighten the Instructions with two or three real examples of each label, and keep the list of labels short. Four labels classify far more reliably than twelve.
A template is greyed out. Some AI Agent templates need a specific trigger. Add that trigger to the workflow and the template becomes available.
Frequently asked questions
What is the difference between AI business process automation and ordinary workflow automation?
Ordinary workflow automation follows fixed rules on structured data. AI business process automation adds steps that read unstructured input, such as emails, chats and call transcripts, and turn it into structured decisions. The best designs use both, with AI only where a branch cannot decide.
Should the AI Agent action send messages itself?
It can, because sending an SMS is one of its tools. For anything a client sees, we prefer the AI step to classify and a branch to send a message you wrote and approved. Let the AI send directly only for low-risk replies you have tested.
How many AI steps should one workflow have?
Usually one, occasionally two. If a workflow needs more, the process probably has several separate decisions that deserve their own workflows. For ideas on where to start, see how to use AI for business and AI follow-up.
How do we stop the AI making costly mistakes?
Keep each AI step narrow, have it write to a tag or field a person can see, and route uncertain results through the None branch to a human. Sample twenty decisions a month and check them by hand. For the sales side of the same idea, read the AI sales assistant guide.