You’ve got work that repeats every week, someone on the team spends hours on it, and you keep hearing that AI can take it over. As of September 2026, seven in eight of the leading guides to AI automation are written by companies that sell automation platforms, so each one explains the idea through its own tool.
This guide explains AI automation without tying it to one platform, and the automation it walks through can be set up on most of them. It covers how one automation runs from start to finish, where people stay in the loop, the four ways to get one, five platforms worth knowing, how to pick your first, and what goes wrong once it’s live. We’re a UAE-based team that builds AI systems into the tools companies already use, and the main example below is one of ours.
- What is AI automation?
- How does an AI automation work?
- A worked example, step by step
- Human in the loop: where people stay, and how to decide
- AI automation vs RPA and traditional automation
- Examples of AI automation by team
- Four ways to get AI automation
- Five AI automation platforms to start with
- How to choose your first AI automation
- What goes wrong once it runs
- AI automation: common questions
- What to do next
What is AI automation?
AI automation is software that does a task which used to need a person to read, judge or write something, and then acts on the result. Traditional automation follows fixed rules, so it can only handle inputs that look exactly as expected. AI automation adds a model that can read a messy email, a photo of a receipt or a scanned PDF, work out what it means, and pass a clean answer to the next step.
For example, a rule can forward every email with “invoice” in the subject line to the accounts inbox. An AI automation can open the attachment, read the supplier’s name, the amount and the due date, and flag the one invoice whose amount doesn’t match the purchase order. That reading used to be a person’s job.
An AI automation usually uses a large language model (LLM) for that reading and deciding, alongside ordinary software that fetches data and takes the action. Salesforce describes it as machine learning and natural language processing handling routine tasks, and Zapier as AI models embedded into your workflows.
How does an AI automation work?
Every AI automation, from a simple email sorter to a system that runs a department’s paperwork, has the same five parts:
- A trigger
- Context
- A decision
- An action
- A check
The trigger is the event that starts a run. It might be a new email, a file uploaded to a folder, a form submitted, or a time of day.
The context is everything the model needs to decide well. That means the input itself plus whatever the system fetches from elsewhere, such as the customer’s past orders, the company’s expense policy, or the matching purchase order.
The decision is the model’s job. It reads the input and the context, then extracts, classifies, drafts or chooses. Good systems ask the model for a fixed structure, such as a set of named fields, so the next step can check each field.
The action is what the software does with the decision. It creates a record, sends a draft, updates a system or routes the item to a person. The action runs through the other system’s API, so it happens the same way every time.
The check is how you know the run was right. Some checks are rules, such as “the total matches the line items”. Some are a confidence score from the model. Some are a person. The next two sections show all three in one real system.
A worked example, step by step
Travel expenses make a good first example, because every company has them and the work is almost all reading. We built this automation for a company whose travelling staff, about 8 to 10 people, work across the GCC. Before it, each traveller collected receipts for hotels, flights, visas, taxis and day-to-day spending, typed them into a spreadsheet and submitted it after the trip. The accounts team then checked every receipt against the spreadsheet and keyed each one into their accounting system by hand, for every person and every trip.
The same work as an AI automation runs in eight steps:
- Capture the receipt
- Read it
- Categorise it
- Check it against the rules
- Score the confidence
- Route it
- Post it to the accounting system
- Review, correct and learn
1. Capture the receipt
Each traveller signs in once to a lightweight, responsive web page behind a login, which works in their phone’s browser. After a taxi ride, they photograph the receipt. For a hotel, they upload the PDF. For something with no receipt, they record a voice note or type a short text entry, and voice notes are one of the most used inputs. The upload is the trigger, and it happens on the road, during the trip.
2. Read it
The model reads the image, PDF, voice note or text and pulls out the date, time, amount, what the expense was and where it happened. It returns those as named fields, never as a paragraph. A taxi receipt photographed at an angle in a dim car is exactly the input a rule can’t read and a model can.
3. Categorise it
The model assigns each entry a category, such as hotel, flight, visa, taxi or day-to-day spending, and records which trip it belongs to. This is the judgment a person on the accounts team used to make for every line.
4. Check it against the rules
Some checks shouldn’t be left to a model, because a rule is exact and cheap. In a system like this, rules catch the same receipt uploaded twice, a date outside the trip, an amount over the policy limit for its category, or a total that doesn’t match the line items.
5. Score the confidence
Every entry gets a confidence score, a percentage that says how sure the system is that the extraction and category are right. A clear hotel PDF scores high. A crumpled receipt in Arabic and English with a handwritten total scores lower. On this system, scores range from about 80 to 100 per cent.
6. Route it
The score decides where the entry goes. Anything below 75 per cent is flagged for manual review by the accounts team, and anything above goes straight through. A flag should carry its reason, so the reviewer knows where to look.
7. Post it to the accounting system
Entries that pass go into the company’s accounting system through its API, which it already exposed. Nobody retypes anything.
8. Review, correct and learn
The accounts team works through the flagged entries, approving or correcting each one. Every correction is worth saving with the original receipt, because a set of real past mistakes is how you test the next change to the prompt, the model or the threshold before it goes live. Where the threshold sits was the client’s decision, set at the level they’ve become comfortable with so far. For the client, the result is time back for the accounts team and less admin for the travellers after every trip.
This table shows who does what at each step:
| Step | Done by | What it needs | What it guards against |
|---|---|---|---|
| Capture | Traveller | A login and a phone | Receipts lost after the trip |
| Read | Model | The image, PDF, voice note or text | Retyping by hand |
| Categorise | Model | The category list and the trip | Inconsistent coding |
| Check | Rules | The policy and past entries | Duplicates, out-of-policy spend |
| Score | Model | The extraction | Silent errors |
| Route | Rules | A threshold | Low-confidence entries posting unseen |
| Post | Software | The accounting system’s API | Keying errors |
| Review | Accounts team | The receipt beside the fields | Mistakes reaching the books |
Human in the loop: where people stay, and how to decide
Human in the loop means a person reviews or approves some of an AI automation’s work before it counts. Every automation that touches customers, money or records needs a decision about where that person sits and how much they check. Make’s own guide lists human oversight as one of its five steps, between building the automation and testing it. There are four places to put a person.
Before the action
A person approves every output before anything happens. This is the right starting point for a new automation, and for anything sent to a customer or anything that moves money. For example, an automation that drafts replies to customer complaints puts every draft in front of an agent, who edits and sends it.
On exceptions
The system handles the routine cases and sends a person only what fails a rule or falls under a confidence threshold. The expense system above works this way. The accounts team sees the flagged entries, and everything above the threshold goes through on its own.
After the fact
Once you know how often the automation is wrong, a person can audit a sample each week instead of checking everything. For example, a support team might read a random handful of the tickets the system classified and routed that week, and raise the checking level again if the error rate climbs.
On changes
Every change to the prompt, the model or a rule gets tested before it goes live, by running it against saved past cases and comparing the results with the right answers. Models change, and a new version that’s better on average can still be worse on your receipts.
How much checking does an automation need?
Four questions set the level:
- Can the action be undone? A draft can be deleted. A payment or an email to a customer can’t.
- What does an error cost? A wrong expense category costs a correction. A wrong bank detail costs real money.
- How often is it wrong on your test set? Measure it on real past cases before you decide, then keep measuring.
- Does a rule or a contract require a person? Some approvals have to be signed by someone accountable, however good the system is.
Start a new automation with a person before every action, and move to exceptions only once the test set shows it can. The expense system’s threshold is where the client has got comfortable for now.
What the reviewer should see
A review step is only as good as the screen the reviewer works on. They should see the source (the receipt, the email, the document) beside what the system extracted or drafted, the reason it was flagged, and a way to approve or correct it in one step. Each correction should flow back into the test set. If reviewing takes as long as doing the work by hand, people stop reviewing.
AI automation vs RPA and traditional automation
Robotic process automation (RPA) follows fixed steps on structured inputs, clicking and typing through screens the way a person would. AI automation adds judgment, so it can take unstructured inputs such as emails, images and documents. AI agents go a step further, choosing which steps to take towards a goal instead of following a set path. Oracle’s guide describes AI automation as RPA combined with AI models, and the expense example above is the same mix of rules and a model.
| Rules and RPA | AI automation | AI agents | |
|---|---|---|---|
| Input it handles | Structured, predictable | Unstructured (text, images, documents) | Unstructured, plus open-ended goals |
| How it decides | Fixed rules | A model inside a fixed path | A model choosing its own steps |
| When the input is unusual | Stops or errors | Flags low confidence | Tries another route |
| How you test it | Each rule, once | Past cases, each change | Past cases and full runs, each change |
| Best for | Stable, high-volume steps | Reading, classifying, drafting | Multi-step tasks with a bounded goal |
Rules still win when the input never varies. If a spreadsheet arrives in the same format every week, a rule moves it faster and more cheaply than a model. Intelligent document processing (IDP) is the name for AI that reads documents such as invoices and forms, and it’s the reading step in the expense example.
Examples of AI automation by team
One example for each team:
Sales
A new lead fills in the website form. The model reads the message, looks up the company, scores the fit against your ideal customer, and drafts a first reply. A salesperson checks the draft and sends it.
Customer support
A ticket arrives. The model classifies it, pulls the customer’s order history, and drafts a reply with the order attached. An agent approves any refund above a set amount.
Finance
A supplier invoice arrives by email. The model extracts the supplier, amount and due date, rules compare it with the purchase order and the delivery record, and a matched invoice becomes a draft bill. A person approves payment.
Operations
A weekly report used to mean hours of copying numbers between systems. The automation pulls the numbers, the model writes a short summary of what changed and why, and the manager edits the summary before it goes out.
Marketing
A new blog post is published. The model drafts social posts and a newsletter paragraph in the brand’s voice, and the marketer edits them before anything is scheduled.
Four ways to get AI automation
There are four routes to an AI automation, and the right one depends on where your data lives, how unusual the process is, and what you want to own at the end.
AI already inside the software you pay for
Many business tools now include AI automation. Microsoft offers Copilot Studio and Power Automate across Microsoft 365, and Salesforce offers Agentforce inside its CRM. This is the fastest start if the whole process lives in one vendor’s tools. It stops at that vendor’s data and actions.
A workflow platform
Zapier, Make and n8n connect thousands of apps with little or no code, and each lets you add AI steps to a workflow. The platforms section below compares five of them. n8n can also be self-hosted, so the workflow runs on your own servers. A platform suits processes that pass data between common apps. It gets harder when the logic has many branches, when a person needs a proper review screen, or when volume grows and per-task pricing adds up.
A custom build
A custom build is software written for your process. It fits when the process spans systems with no ready connector, needs its own screen for people to review, handles data that can’t leave your environment, or runs at a volume where platform pricing no longer makes sense. You own the code, the prompts, the test cases and the data. The expense system above is a custom build, because the travellers needed a simple upload page behind a login and the accountants needed a review queue tied to their accounting system. If you’re exploring an AI automation or weighing a build, our AI strategy and implementation page covers how we shape and scope one.
Connecting an assistant to your own data
The newest route skips building a workflow at all. An AI assistant such as Claude connects to your systems through the Model Context Protocol (MCP), an open standard Anthropic open-sourced in November 2024. The person then asks questions and gets work done in plain language, inside the assistant they already use. For example, one of our clients found their product analytics in Amplitude hard to interpret and to match against their orders. We connected their Amplitude data and their site’s order data into one MCP connection for their Claude, and they now ask it questions directly and build their own reports.
This table compares the four routes:
| Route | Best for | Time to first result, relative | What you own | Where it stops | How cost grows |
|---|---|---|---|---|---|
| Built into your software | Processes inside one vendor’s tools | Quickest | Configuration | That vendor’s data and actions | Per seat or per use |
| Workflow platform | Passing data between common apps | Quick | Workflows on the platform | Complex logic, review screens | Per task or run |
| Custom build | Cross-system processes, own review, sensitive data | Longest | Code, prompts, test cases, data | Your appetite to maintain it | Model usage and hosting |
| Assistant over MCP | People asking questions of their own data | Quick | The connector | Tasks that must run without a person | Assistant plan and model usage |
Five AI automation platforms to start with
These five run from no-code tools to open-source builders you host yourself. Prices are the entry plans as of September 2026, and cost rises with volume on every one.
Zapier
Zapier connects more than 9,000 apps and lets you add AI steps and agents to its workflows, which it calls Zaps. You build in a no-code editor, and tasks are the unit it charges for.
Ideal for: teams that want the widest set of app connections without writing code.
Price: free plan with 100 tasks a month; Professional from $19.99 a month; Team from $69 a month (pricing)
Make
Make connects more than 3,000 apps in a visual builder, where each workflow (a scenario) is drawn as a map of steps with routers and filters. It also builds AI agents that act across those apps.
Ideal for: teams that want to see every branch of a workflow on one screen.
Price: free plan with 1,000 credits a month; Core from $9 a month for 10,000 credits (pricing)
n8n
n8n offers more than 2,000 integrations and AI agent workflows for technical teams. You can use n8n’s cloud or self-host it, and a self-hosted Community Edition is available on GitHub. It charges for each complete workflow run, however many steps it has.
Ideal for: technical teams that want to run automations on their own servers.
Price: cloud Starter from $20 a month, billed annually, for 2,500 workflow executions; Community Edition self-hosted (pricing)
Langflow
Langflow is an open-source, Python-based framework for building AI applications, with a visual builder for AI agents, retrieval (RAG) apps and MCP servers. It supports the major LLMs and vector databases, and you can deploy it yourself or use its cloud.
Ideal for: developers building agents who want Python underneath the visual builder.
Price: open source; a free cloud account is available (langflow.org)
Microsoft Power Automate
Power Automate runs cloud flows across Microsoft 365 and other apps, plus desktop flows that automate screens the way RPA does. Microsoft’s Copilot Studio adds AI agents on top.
Ideal for: companies already running on Microsoft 365.
Price: 30-day free trial; Premium $15 per user a month, paid yearly (pricing)
How to choose your first AI automation
Your first AI automation should be a process that’s frequent, painful, easy to check, reachable by software, and owned by one person. Those five tests come from how we choose use cases with clients, where we weigh value, feasibility, data readiness, risk and effort:
- Frequent. It happens every day or every week, so the time saved adds up.
- Painful. Someone would happily stop doing it.
- Easy to check. A person can tell quickly whether the output is right.
- Reachable. The inputs and the target system can be read and written by software.
- Owned. One person can say what “right” looks like and will review the exceptions.
For example, compare the expense receipts with the weekly management report from the operations example. Receipts pass all five. They arrive constantly, nobody enjoys keying them, a wrong amount is easy to spot, the accounting system had an API, and the accounts team owned the result. The weekly report passes most of them, but “easy to check” is weaker, because a summary that sounds right can still misread a number. That makes it a good second automation, with a person reading every summary.
Once you’ve picked a process, take it from candidate to pilot in five steps:
- Map the process as it runs today, including the awkward cases people handle without thinking.
- Collect real past cases as your test set, with the right answer written down for each.
- Choose the route from the four above.
- Run it alongside the people doing the work for a set period, comparing its output with theirs.
- Let it act alone on the cases it gets right, and keep a person on the rest.
What goes wrong once it runs
Five problems show up once an automation meets real work, and each has a fix.
Exceptions nobody listed
Real inputs are messier than the examples used to build the system. A receipt in two languages, a hotel bill split across two cards, a supplier who changes its invoice layout. The fix is a confidence score and a review queue from day one, so the unusual case reaches a person.
No test set, so nobody knows if it got worse
Without saved past cases, a prompt edit or a model upgrade can make the system worse without anyone noticing. The fix is to save every correction with its source and run the full set before any change goes live.
Permissions wider than the task
An automation that only needs to create draft bills shouldn’t be able to approve payments. The fix is to give it the narrowest access the task needs, and to log every action it takes.
Running costs that drift
Model usage is charged by volume, and volume grows once people trust the system. The fix is to track cost per run from the start and review it monthly, alongside the time the automation saves.
A team that works around it
If the review screen is slow or the flags are noisy, people go back to the old way and the automation goes unused. The fix is to design the review step with the people who’ll use it, and to tune the threshold until the flags are worth looking at.