Can you tell the difference between Automation, an AI workflow, and an AI agent?

If you can't, you're not alone.

The AI world right now is a mess of super confusing terms. Every vendor claims their tool has AI included (even when it’s just bolted on to the tool and delivers no value).

Plus, everything is being called an ‘Agent’ even if it's just automation with a small amount of AI.

So, you end up with these problems:

You buy expensive tools that do the same things as your existing tools in slightly different ways.

You avoid automating simple stuff. Because you think you need AI to do it, when in fact a basic script could work perfectly (e.g VBA in Excel or Apps Script in Google Docs).

Plus, Your team doesn't trust any of it. So adoption stays low and nothing gets used.

But, the good news here is that, once you know what each of these terms means, everything becomes clearer.

You know what processes to keep, what to automate, and where AI should (and should not) be used.

So today, we will make this clear.

​

​
Automation vs AI workflow vs AI Agent

1. Automation: A specific, rule-based process where one action automatically triggers another. It's the most predictable, like a simple IF/THEN statement.

An e-mail rule is an automation - IF I receive an e-mail from this address -> THEN move it to this folder.

2. AI Workflow: This is when instead of a clear rule, you need AI to do the sort of activity that a human would have to think about because the logic is not black and white (fuzzy logic).

IF I receive a spreadsheet of expense transactions THEN use AI to generate the category based on the Column C description.

3. AI Agent: Autonomous systems that can achieve a goal by using various tools and methods, which the AI chooses itself.

In this scenario you would simply say “Find my latest expense transactions and categorize them. Let me know when you’re done.”

Here’re some practical applications.

​

​
Example Practical Applications

1. Automation:

Example: Monthly Data Cleaning in Google Sheets

You have the same messy credit card statement every month. The columns are fixed, the headers are fixed, and your desired output always looks the same.

Here automation is perfect:

• A script can read the raw statement
• It can keep only the columns you care about.
• It can split out cardholder and merchant names into clean fields.
• It can write everything into a clean transactions tab you use for analysis.

How to build this:

Step 1: Show AI your data and describe what you want

Open your messy Google Sheet. Select the first few rows (header + some data). Copy it into ChatGPT and say:

"Every month I get this credit card statement. I need to: keep columns C to E, delete the summary rows, extract the company name into a separate column, extract the cardholder name into a separate column, and put the clean data into a new tab called 'Clean Transactions'."

Step 2: Ask for the script

Say: "Write me a Google Apps Script to automate this."

ChatGPT will give you the code. Copy it.

Step 3: Run it

In Google Sheets: Tools → Script editor → Paste the code → Save → Run.

Google will ask for permission the first time. Approve it. Done.

Now every month, you just open the file and click "Run."

Nothing here requires judgment. Once the script works, you use it forever.

(Note - You could also use VBA here if you’re using Excel / desktop office apps, or Office Scripts if you're working in the cloud)

​

​

2. AI Workflow:

Example 1 (Simple) - Smarter Expense Categorization with a Custom GPT

Now imagine you want to categorize hundreds of transactions where the data is really messy, and the rules needed to transform the data are really fuzzy.

Your team member could have entered ‘lunch with client’, ‘evening meal’, or ‘food’ in a free text field, and no amount of standard automation can account for the number of variables.

This is where AI workflows come in, as you can add a step that acts like the human brain, to spot patterns, categorise, or even generate new data as part of the process.

• Your team uploads the transaction
• A custom GPT proposes categories based on your rules.
• The team reviews anything marked as uncertain and corrects it.

How to build this:

Step 1: Create your GPT

Go to ChatGPT → Explore GPTs → Create. Tell it:

"You categorize expense transactions. Here are my categories: Travel, Food, Supplies, Marketing, Office. Ask me to upload transactions, then propose categories based on the description descriptions. If there’s anything you cannot categorise, output ‘Could Not Categorize’ in place of the category for us to review"

Step 2: Upload your data and let it work

Upload your transaction list. The GPT will categorize everything and flag anything it's unsure about.

Step 3: Review and correct

Check the "uncertain" items. When you correct something (like "Office Depot should be Supplies, not Other"), tell the GPT by adding the human reviewed transactions as examples within the GPT's knowledge so it knows for next time.

Example 2 (Advanced) - Smarter Expense Categorization with n8n

Maybe you want something that runs automatically every month without you pressing anything.

With n8n, you chain multiple tools together and include AI within the process:

• It pulls the latest statement data from your source system.
• It sends chunks of data to an LLM for categorization and location tagging.
• It writes the results back into Google Sheets or your data warehouse.
• It posts a summary into Slack with the top exceptions for review.

How to build this:

Step 1: Set up the workflow

In n8n, create a new workflow. Add a Google Sheets node to read your transaction data.

Step 2: Add AI categorization

Add an OpenAI node. Give it a prompt: "Categorize these transactions into Travel, Food, Supplies, Marketing, Office. Also extract the location if you can identify it from the description."

Step 3: Write results and notify

Add a Google Sheets node to write the categorized data back. Add a Slack node to post a summary of what needs review (largest transactions, anything marked "Other").

Now it runs on its own. You just review the exceptions.

​

​

3. AI Agents:

With AI Agents, you provide a goal, and the Agent figures out the "how" on its own.

What makes a REAL agent:

A real agent isn't just a chatbot answering questions. It's capable of:

• Autonomy: It creates its own to-do list based on your request.
• Tool Use: It can browse the web, write and execute Python code, or connect across multiple systems to get work done.
• Self-Correction: If it tries a method and fails (e.g., a file is missing), it notices the error and tries a different approach without crashing.

Here are three examples of how real Agents are being deployed today.

NOTE: A lot of Agents are still in early stages or 'Beta'. I'd recommend testing on a small use case before trying to get an AI Agent to own any business-critical process.

​

Example 1: The "Accounting Automation" Agent (OpenAI)

Accounting firms are already using AI agents built on OpenAI's platform to automate structured accounting work. The best-documented case is Basis, which builds multi-agent systems for top accounting firms.

What Basis agents actually do:

• Automate reconciliations (bank statements vs GL)
• Generate journal entries based on transaction patterns
• Produce financial summaries for client review

Why this is agentic:

The agent doesn't just answer questions. It breaks down a reconciliation task into steps, calls the right tools (retrieval, code execution, data validation), and works through the entire workflow autonomously. If something doesn't match, it flags it and tries alternate reconciliation logic.

​

Example 2: The "Enterprise Workflow" Agent (Microsoft Copilot Studio)

Microsoft 365 Copilot Agents built via Copilot Studio have been deployed by multiple organizations.

Real example: Capita's Email Triage Agent

Capita (34,000 employees, major UK BPO firm) built a Copilot Studio agent that:

• Processes thousands of incoming emails per day.
• Routes and prioritizes them based on content.
• Escalates urgent items to humans with context.

Why this is agentic:

These agents don't just search a knowledge base. They coordinate across Teams, Outlook, SharePoint, and external systems. When they can't find information, they escalate intelligently. When new data arrives, they adjust their routing logic.

​

Example 3: The "Legacy System" Agent (Claude Computer Use)

This is the newest and most radical form of agency. Anthropic's Claude can now "view" a computer screen and "move" a mouse cursor. This bridges the gap for legacy finance systems (old ERPs, banking portals) that are impossible to automate with standard scripts.

What Claude Computer Use has been proven to do (from independent testing):

• Fill web forms from CSV data (logging in, navigating, entering data row by row).
• Log into dashboards, navigate to reports, and download CSVs.
• Triage "Invoice" emails in webmail (search, label, mark as read).

Why this matters for finance:

Many finance teams are stuck with:

• Banking portals with no "Download All" button.
• Vendor systems that require manual clicks through multiple screens.
• Legacy ERPs where the only way to extract data is through the GUI (User Interface).

The scenario:

You say: "Log into the vendor portal, find all invoices dated 'October 2025', and download them to this folder."

What the Agent does (visually navigating like a human):

1. It looks at the screen, identifies the "Login" box, types credentials, and clicks "Login."
2. It navigates menus like a human would. Sees the invoice list, clicks the first one, clicks "Save as PDF."
3. It repeats until all October invoices are downloaded.
4. It notifies you when done.

If the portal layout changes next month, the Agent adapts. If a page doesn't load, it refreshes and tries again.

​

​

The Bottom Line

Finance teams are wasting time on manual processes that could be 80% automated with simple scripts.

But instead of starting there, they're buying 'AI Agent' platforms because that's what everyone's talking about.

I understand this. Agents sound powerful (just look at the use cases I’ve described above) and you should experiment...

But, most of what you need doesn't require autonomy across systems. It just needs a workflow that runs the same way every time, or AI that helps you work faster while you stay in control.

So, pick one manual process. Map the steps. Decide what tool you will try, and give it a go.

You'll know more in 30 minutes of testing than 3 days of vendor demos.

Best,

Your AI Finance Expert,

Nicolas

P.S. - What did you think of this approach? Hit reply and let me know if you're planning to try this for your team (I read all replies).

P.P.S. - As a bonus, I am sharing with you my video tutorial on my top 3 AI strategies to eliminate manual work.​ ​Watch it now!​​​​​