Skip to content
IvarAutomation
Back to Insights

AILLMautomationarchitecture

When Do You Actually Need AI in Automation?

Fixed rules where they're enough, AI where judgment is needed. Here's how to decide what fits each part of a workflow and how much an AI agent can handle on its own.

The question is where AI is needed

When a company starts looking at automation, the question of whether to use AI often comes up. It's a good question, and it gets easier to answer if you split the workflow into steps and ask, for each one, what it takes to get it right.

Fixed rules and AI judgments

A step in an automated flow can be handled in two ways.

With fixed rules, the same input always gives the same result. The rules are written in advance, for example that an invoice over €5,000 must be approved by someone before it's paid. Rules like that can be tested in advance and cost almost nothing to run.

An AI model makes a judgment instead. It can read an email and decide whether it's a support request or a sales inquiry, even when the customer phrases it in a way nobody anticipated. The judgment can be wrong, so it needs to be built in with checks. More on that further down.

Most workflows need both. The trick is knowing which fits where.

Two questions for every step

For each step in the flow, two questions decide it.

The first is whether the step can be described with clear rules. Amounts and dates are best handled with ordinary code. If the step requires someone to interpret text or decide what's reasonable in a new situation, AI is usually the better choice.

The second is what happens if the step goes wrong. If the mistake shows up right away and is easy to fix, an AI agent can act on its own. If a mistake could cost money or damage a customer relationship, there needs to be a stop, such as a person approving or a rule blocking the action.

What AI does well

Some tasks can't be handled with fixed rules. That's mainly text written freely, like a customer email where you have to understand what the person wants, or an invoice where every supplier uses their own format. Solving that with rules leads to long lists of special cases that still break when a new variation appears.

It also covers sorting into many categories described in words rather than numbers, and writing summaries and replies.

An AI agent goes a step further. It can run an entire workflow, use the tools and systems it has been given access to and make decisions along the way. How much it may do on its own is decided by the risk in each step.

An example: sales inquiries by email

Here's what a flow for incoming sales inquiries can look like when fixed rules and an AI agent work together.

  1. The email is received by an ordinary integration that stores it and starts the flow.
  2. The agent reads the email and extracts what the customer wants and whether a budget or timeframe is mentioned. The result comes back in a fixed format so the rest of the flow can rely on the fields.
  3. If it's a common question with a set answer, the agent writes the reply itself based on the company's own documents and sends it.
  4. If it's a larger inquiry, the agent adds it to the CRM, drafts a reply and notifies the salesperson, who takes it from there.
  5. Anything that commits the company, such as prices outside the price list or contracts, requires a person to approve.
  6. Every step is logged, so you can see which decision the agent made and on what grounds.

The agent handles most of the flow here. Fixed rules are used where they're enough, and the stops sit where a mistake would cost something.

How the weaknesses are handled

An AI model can make mistakes, but there are proven ways to limit and detect them.

  • The model returns its result in a fixed format with set fields, and the code checks that the fields are filled in and reasonable before anything happens.
  • Amounts and recipients are checked with ordinary rules in the code, whatever the model concluded.
  • The agent only gets the permissions and tools it needs. These are often called guardrails. An agent that answers email doesn't need to be able to make payments.
  • Before the flow goes live, it's run against a set of real examples with known answers (evaluation), and the same tests run again whenever something changes.
  • Every run is traced and logged (observability), so you can see afterwards exactly what the agent did and why.

What this means for cost

A step that can be solved with a rule shouldn't cost a model call. That makes the flow cheaper and easier to debug. Where AI is needed, you pick a model that suits the task. Simple sorting tasks often work fine with a smaller, cheaper model, while harder judgments may need a larger one. Prices vary a lot between models, so it's worth testing what's good enough before deciding.

How to think about your next automation

  • Write down the steps in the flow, one at a time.
  • For each step, ask whether it can be described with rules or whether it needs interpretation and judgment.
  • Think about what happens if the step goes wrong, and decide how much the agent may do on its own there.
  • Decide how you'll know the flow works, with tests before launch and logs afterwards.

Want help working out which parts of your workflows suit fixed rules and where an AI agent would make a difference? Get in touch and we'll go through it together.

Next step

Does any of this sound familiar?

Tell us how things look for you, and together we'll see what can be automated.

Free of charge · No obligation