What are AI automations?

What limited automations for many years?

Traditional automations work well when the rules are clear: if a particular thing happens, take a particular action. If a customer selects “Sales” on a form, the enquiry goes to sales. If they select “Support”, it goes to support. But what happens when the customer writes a message in their own words?

“I bought a machine from you a month ago. It keeps stopping halfway through a job again, and I need someone to get back to me today.”

To handle this message, you need to recognise a fault, understand that it is probably a recurring problem and notice the request for urgent attention. A system based only on rules could look for words such as “fault” or “urgent”. But neither word appears in our example. As you add more possible phrases and scenarios, the rules become more complex, and there will still be cases you haven't anticipated.

Systems that classified text and recognised patterns existed before the current wave of AI. But for many businesses, handling free-form language and changing information flexibly required specialised solutions and considerable effort.

What does AI change about automation?

AI lets you add a step that interprets information without a fixed format: a message, a document, a call transcript or a request phrased in an unfamiliar way.

Alongside using a customer's form selection or keywords, you can ask the system to identify the purpose of the enquiry, extract important details and suggest the next step.

In the machine example, a workflow could be designed to:

Accessing information and taking these actions requires appropriate connections to your business systems and permissions set in advance. Using AI doesn't automatically give it that access.

An AI automation is an automated process that uses AI capabilities in one or more of its stages. Some actions may be entirely conventional, such as saving details in a system, while others involve interpreting information or creating content.

This opens up possibilities for automating tasks that were previously harder to automate: sorting enquiries by their content, comparing documents written in different ways, preparing drafts tailored to a customer or turning meeting notes into a task list.

Can you trust AI automations?

They can be useful when their checks and level of independence are designed to suit the task.

An error in tagging an internal document has different consequences from an error in a quote sent to a customer or a decision affecting safety. The more serious the consequences of an error, the more thorough the checks and oversight need to be.

Before putting a workflow into use, test it on real examples, including unusual cases: missing information, unclear requests, conflicting documents and situations where the system doesn't have an adequate answer. You also need to define when it should stop and hand over to a person.

For business processes where an error could put lives at risk, my recommendation is for a qualified person to review every result before it is used to make a decision or take action. Even that review cannot guarantee zero errors or replace a professional assessment of whether the system is suitable for the task.

Human review can still leave plenty of room to save time

Suppose an employee needs to read several documents, find specific details and prepare a draft report. An automation can prepare the draft and identify the source of each detail, leaving the employee to check and correct it.

The review still takes work, but the employee receives organised material they can assess. The actual time saving needs to be measured, including the time spent checking and correcting the output.

You can also add automated checks between stages: confirm that required fields are present, compare amounts with the source system or run an additional AI check for discrepancies between an original document and its summary.

These checks can improve quality, but a second AI check may miss the same error. It helps to combine them with checks against source information and clear rules, and test whether they actually catch mistakes. Anthropic's guidance on building these systems also emphasises testing, feedback and human checkpoints. Read more