# AI and automation: where does each belong in your workflow?
Source: https://humanspark.ai/ai-and-automation/

When someone tells me they want to automate something with AI, my first question is which parts of the work need AI at all.

Do they need help interpreting an unfamiliar document? Do they want information moved between two systems? Or do they want AI to build a tool that follows rules they already understand?

Those can all be parts of the same project, but they're different jobs. Treat them as one thing and it gets harder to decide what to build, which parts AI should handle and how closely the work needs checking.

I've been working through this in my own business and with clients. I haven't got all of it worked out, but separating the two ways AI can be involved has already made those conversations more practical.

## AI can help build the system, work inside it, or both

An automation is work that runs when something triggers it, following a defined process. The trigger might be an incoming form, a scheduled time or a person pressing a button. Some workflows run start to finish on their own, and others pause for someone to approve a step.

AI can be involved in two separate ways.

**AI can help you build the automation.** You describe the job, work through the rules with it and get its help building it. For example, it can help you create a tool that converts a spreadsheet into the format another system expects. Once that tool is built, it can run as ordinary software, with no AI involved in processing the files.

**AI can do a step inside the automation.** It might interpret an email, suggest a category for an unfamiliar transaction or draft a summary from a transcript. The workflow around it decides when that step runs, what information it gets and what happens to its output.

You can use both in the same project, and the two questions are independent of each other:

| Did AI help build it? | Does it use AI when it runs? | Example |
| --- | --- | --- |
| No | No | A scheduled report built using an existing software feature. |
| Yes | No | A file converter written with an AI coding assistant. |
| No | Yes | A manually configured workflow that sends a document to AI for summarising. |
| Yes | Yes | An AI-assisted build of a workflow that uses AI to interpret incoming requests. |

That's why someone can say they automated a task with AI even though the finished automation never calls an AI model. It's also why learning to prompt is only part of [learning to automate](https://humanspark.ai/automation-literacy/).

## Decide where rules, AI and people belong

My working principle is **rules before guesses**. Where the rule is known and stable, write it into the software. Where information needs interpretation, consider AI. Where a decision carries serious consequences, decide who has the authority to approve it.

My bookkeeping shows how that splits out. When a supplier and transaction type reliably map to an agreed category, software can apply that rule. An unfamiliar description is where AI can help with interpretation. A proposed new rule needs checking, because it'll apply to every matching transaction from then on.

For a transaction workflow in general, that might mean software checks the totals, AI suggests how to handle unfamiliar descriptions, and someone with finance knowledge reviews the exceptions and any changes to the rules. No single component has to do every job.

Rules before guesses doesn't mean AI only handles the exceptions. If the main purpose of a workflow is summarising every transcript, AI is doing the main job, and that's fine. The principle asks you to choose deliberately for each step instead of defaulting to AI for all of them.

A fixed rule doesn't guarantee a correct result either. A wrong rule makes the same mistake every time it runs, so the person who understands the work needs to help define it and test it.

## Decide how much runs without supervision

The next decision is how closely someone supervises the work.

When you're working with AI interactively, you can spot a misunderstanding and redirect it. When a workflow runs in the background, nobody is there to do that, so the chances to step in have to be designed into the process.

An automation could prepare a draft reply and wait for approval, or it could send the reply itself. Both automate the work, but the second has more authority and different consequences when it gets something wrong.

My workshop recap workflow shows all of these choices in one place. A transcript becomes a recap script, narration, slides, a video and a web page. Some steps generate or interpret content. Others move files and run software. Reviewing the recap and deciding whether to share it are separate responsibilities again.

Generating the material and sending it to clients are different permissions. I keep the final review, and the preparation time saved is still worth having.

The question to ask is how much the system can complete within limits you can justify. Removing the last human checkpoint isn't a goal in itself.

## Start with one workflow you understand

Choose a piece of work that's repetitive, annoying, time-consuming or prone to errors. Map what happens now, including the exceptions, then ask:

1. Which steps follow rules we already know?
2. Which steps need interpretation, and could AI help?
3. Where should a person make or approve a decision?
4. Could AI help build the workflow, even if it isn't needed when it runs?

Who builds it, who checks it and who looks after it once it's running are separate decisions, and I've covered them in [Who should build your automations, and who should look after them?](https://humanspark.ai/who-builds-your-automations/)

You may end up with a workflow that uses AI throughout, one that uses it for a single step, or one that needs no AI once it's built. Any of those is fine if it improves the work and the people responsible for it can support it.

If you're looking at a workflow and aren't sure where AI would help, [book a 25-minute Focus Call](https://humanspark.ai/call/) and we can talk through the process and a sensible first step.