AI workflow automation: what to automate, step by step
AI workflow automation shown on one real workflow: which steps get a rule, an AI step or a human, how it connects to your systems and what to leave out.
- Published
- Reading
- 8 min
- By
- Authoricy
AI workflow automation runs a business process as a fixed flow of rules, with a language model only on the steps a rule cannot handle, such as reading an email. Automate the steps that are frequent, well understood and cheap to get wrong, and keep a person where a mistake costs money or affects someone’s rights.
Most of a good flow is not AI at all. The model often does two or three steps out of ten. The rest are lookups, checks and writes that plain code does the same way every time. Getting that split right is most of the work in AI integration.
Three kinds of step: rule, AI or human
Anthropic draws a useful line in Building effective agents. Workflows run models and tools “through predefined code paths”. Agents decide their own steps. Its advice: find the simplest solution that works. For most business processes that is a workflow.
Inside a workflow, every step is one of three things:
- A rule. Same input, same output. Look up a customer number, compare two amounts, send a template.
- An AI step. The input is language: an email, a PDF, a note from a phone call. The model reads it and returns structured data with a confidence score. It never decides anything that matters on its own.
- A human. Approves, corrects or handles the exception. The case, the source and the suggested answer sit on one screen.
When the input decides which steps come next, you may need an agent. Our guide to AI agents for business covers when that is worth the extra risk.
One workflow taken apart: customer orders by email
This is an illustration, not a client case. A wholesaler with 40 staff gets B2B orders in a shared mailbox. Some come as Excel files or PDFs. Many are a line of text: “Same as last time but 20 boxes of the 6 mm screws.” Two people retype them into the ERP every morning.
Steps 1 to 5 turn the email into a draft order.
| Step | Who does it | Why |
|---|---|---|
| 1. Mail arrives in the orders mailbox | Rule | A webhook or a scheduled check. Nothing to judge. |
| 2. Is this an order, a question or a complaint? | AI step | Free text. The model sorts it. Anything that is not an order goes to the normal inbox. |
| 3. Which customer is it? | Rule | Match the sender’s domain against the customer register. Unknown sender goes to a person. |
| 4. Read the order lines | Rule or AI | An Excel file with the customer’s own template is parsed by code. Free text and PDFs need the model. |
| 5. Match “the 6 mm screws” to an article number | AI step | The model searches this customer’s order history and the catalogue. Below the confidence threshold, a person picks. |
Steps 6 to 10 check, approve and write the order back.
| Step | Who does it | Why |
|---|---|---|
| 6. Check price, stock and credit limit | Rule | The ERP owns price agreements and stock. A model never sets a price. |
| 7. Flag deviations | Rule | Ten times the usual quantity, a new article or a blocked account stops the flow. |
| 8. Approve the order | Human | A person sees the draft next to the original email and confirms or corrects it. |
| 9. Create the order in the ERP | Rule | A write through the ERP’s API, with the approver’s name on it. |
| 10. Confirm to the customer | Rule | A template filled from ERP data, not text the model wrote. |
Two of ten steps always use the model, and a third does when the order is free text or a PDF. One uses a person. The rest is plain code. Every correction in step 8 is logged and used to improve steps 2 and 5. Once corrections are rare, step 8 can become spot checks for repeat orders. Keep it for new customers and large orders.
The model cannot create a customer, change a price or write to a customer. Those limits let the flow prepare the morning’s orders at 06:00, before anyone is in, and leave only the approvals for the team.
How the AI step reaches the systems you already run
The integration is where projects slip. There are four ways in, best first:
- An API. The ERP or CRM accepts reads and writes over a documented interface.
- A webhook. The system announces events, such as a new order or a changed status, so the flow does not have to poll.
- A file. An export or import folder the system reads on a schedule. Slower, but it works with older on-premise software.
- Screen automation (RPA). A robot clicks through the user interface. Use it only when nothing else exists, because a redesigned screen breaks it.
Before you start, check three things. Does your licence include API access? What are the rate limits at peak hour? Can the flow get a user with only the permissions it needs? A flow running as admin is one nobody can audit. More on the connection layer: AI integration services.
What to automate first: six questions
- Volume. How many cases a month? Fifty is rarely worth a build. Fifteen hundred usually is.
- Minutes per case. Volume times minutes is the time you can win back.
- How much is rule-bound? If most steps can be written as rules, the AI part is small and the flow is cheap to run.
- Cost of an error. A misrouted support ticket costs a few minutes. A wrong price on an invoice can cost a customer.
- Personal data. If present, you need a lawful basis, a processor agreement and a set retention period for logs.
- Is the data in a system? The flow must write its result somewhere. A process that ends in someone’s head cannot be automated.
Order intake, support triage and contract review look alike in a pitch deck. Scored on these questions, they rarely do.
What not to automate
- Decisions about a person. Hiring, credit and dismissals. The law limits this.
- Low volume. Ten cases a month do not pay for building, testing and keeping a flow running.
- A process nobody can describe. If three staff do it three ways, fix the process first.
- Steps with no system to write back to. Someone retypes the output anyway.
- Silent, expensive errors. If a mistake would surface months later, for instance in a VAT return, keep a person in the loop or do not automate it.
What it costs to run each month
Running costs have three parts: model calls, hosting and upkeep. For model calls, multiply cases per month by tokens per case and by the provider’s published price per million tokens.
Upkeep is the part people forget. An ERP upgrade changes an API. A customer changes its PDF layout. The provider retires a model version. Budget time for it every month.
A no-code platform suits a simple flow between popular cloud systems. Custom code suits an older system, strict data rules or high volume, where per-task fees add up. Either way, ask who owns the code and the accounts. Our guides on how to choose an AI consulting company and how to choose a custom software development company list the questions to ask.
The build cost depends on the number of systems, the quality of their APIs and how far the flow writes back. Authoricy starts every project with a fixed-price discovery that maps the process step by step and ends in a fixed price for the build. A first integration usually takes 4 to 8 weeks. If you want help mapping your own process, our AI integration service starts there.
AI automation in a European business
Two sets of rules apply to most flows.
GDPR. Article 22 gives people the right not to be subject to a decision based solely on automated processing when it has legal or similarly significant effects. A flow that rejects job applicants or credit applications without a person is the typical case, and it is allowed only under narrow exceptions. The model provider usually processes data on your behalf, which needs a processor agreement.
The AI Act. Article 50 requires that people are told when they interact directly with an AI system, unless it is obvious. An internal order flow is not affected. A chatbot answering customers is. The AI Act also lists high-risk areas in Annex III, among them recruitment and access to essential services. Flows in those areas carry far heavier duties.
One risk is technical. A flow that reads inbound email can be sent instructions by anyone. This is prompt injection. Give the model no tools it does not need, treat its output as data to check, never as a command, and log every call.
Frequently asked questions
What is the difference between AI workflow automation and RPA?
RPA repeats clicks and keystrokes on structured screens and breaks when the input or the layout changes. AI workflow automation adds a model step that reads unstructured input such as emails and PDFs.
Do I need an AI agent or just a workflow with an AI step?
Start with a workflow if you can write the steps down in advance. An agent chooses its own steps, which suits varied tasks but is harder to test and audit. Most business processes run well as a fixed flow.
Is AI workflow automation GDPR compliant?
It can be, if you design for it. You need a lawful basis for the personal data, a processor agreement with the model provider and a clear answer on where the data is processed. Keep a person on any decision with legal or similarly significant effects on someone.
How do I get started with AI automation in my business?
Pick one process with high volume and a low cost of error, and measure how long it takes today. Map it step by step into rule, AI and human, then run a pilot on real cases with a person approving every output.