I’ve said before that AI agents don’t really exist. Most tools that call themselves “agents” were just AI workflows. A step-by-step process with some AI built into one or more of the steps.
But the latest tool updates have made me wrong about this.
So what changed?
The tools got their own computers.
ChatGPT started with Agent Mode but then it really took over when Claude launched Cowork.
You don't see it but when you use Cowork, it opens a virtual machine session in the Cloud where Claude operates and works inside tools like Excel, Powerpoint, Python, etc… This is this technology which made Claude so good at creating professional grade financial models and presentations. Now you find this also in ChatGPT and Copilot with their "Cowork/Work" functionality.
And recently, even Google announced its Gemini agent. You give it objectives, not instructions. It can get its own email address and calendar, and it can keep working for hours or days. OpenAI also launched ‘Dots’. They’re always-on agents that work toward the goals you set.
Before, you wrote a prompt and the AI gave you one output. Now the AI has a computer to work on. If you let it use your files and apps, it starts doing things without you having to explain every step, or prompt it each time.
So tools like Claude Cowork are basically agents now.
But there is a bit risk with this…
They have existed for years, and were used previously for a range of things like powering websites, apps and managing databases.
Now, AI can use them to carry out tasks in a similar way to how you’d use your computer. But they can do it without having to hijack your screen, keyboard and mouse.
Imagine one of these agents with its own mini-computer, with free run of your files and connected applications.
What if it’s missing a key bit of information? What if it guesses at the goal? This is a big risk, no?
Not many people know this. But I used to be a DJ.
And my goal was to get the crowd dancing. But to do this, I’d need to know: Who comes to the party, the best tracks to use, the order to play them in, and when to switch records.
An AI might think that ‘I’m a Barbie Girl’ is a good track to get people dancing (it almost always is).
But if it’s played to a wedding of hard rock fans, people will leave, and this will make the bride and the group quite disappointed for what was supposed to be the best day of their life
Yes this is me as a DJ as a 20 years old!
So like more than 20 years ago when I needed to have all my CDs prepared for a good party today I’m giving you what you need to prepare to make sure any finance agent you build works towards the right goals.
Performance = Intelligence x Context
AI Agents Don't Sleep, But Still Need Context
In one of our AI Connect sessions Eduardo (one of our AI Finance Club experts) explained that when his agent gives him a bad answer – or a draft email that's just off – it's "rarely because the model isn't smart enough. It's because it doesn't have your context."
So he gave this formula: Performance Equals Intelligence x Context.
Intelligence is how smart the AI is. This is what the AI companies release and you can’t change this capability. All you can do is choose from models.
Then use the mid-range models (Sonnet) for 90% of your work, and simple models (Haiku) for simple information retrieval and brainstorming.
Context is what the AI knows about your company, your data, and your rules. This will be our focus for today, which is super important to get right.
As we said earlier, an AI agent is an AI that works on a goal by itself, setting itself steps, and continuing the work until it’s done.
But, say you tell an agent: "Prepare the Q3 sales report." With no context, it has to guess. Which files? What counts as a sale? Which three months are "Q3" for your company?
When it guesses wrong, it is wrong fast (and confidently)
To fix this, all you need is a folder on your computer or in the cloud, with plain text files inside (and maybe some databases / tables – from Excel or a connected tool).
4 things go in it:
Business context: who you are and what you sell.
Structured data: your numbers, in clean rows and columns.
Unstructured data: emails, contracts and PDFs, made light.
A data guide: a note that explains your files to the AI.
These four files are what the agent reloads every morning. Without them it starts from zero.
The next wave of agents will be really intelligent. But without the right context, their performance will always be poor, and you’ll end up wasting a lot of time and money unless you build in the right way.
The 4 AI Agent Friendly Foundations
The way AI ‘sees’ is different to humans. So you need to ensure that whatever context you give, that it’s AI friendly.
A) To make sure that AI doesn’t make things up (or solve for the wrong problem) and B) to make sure you don’t consume all of your AI usage in 5 minutes.
More autonomy = more consumption.
1. Write one page about your business
Open a plain text file. Name it business-context.md, and write four things: where you work, what products you sell, which countries you sell in, and what industry you are in.
For example: "We are a 120-person software company. We sell yearly subscriptions. Most customers are in Germany and France.”
Do it one time and you never have to re-explain anymore. You can reuse the same file in other tools and projects.
Tip – Save your LinkedIn profile as a PDF too. Click the three dots on your profile, then "save to PDF". It takes 30 seconds, and it tells the AI who you are.
Without this file, your agent is like an employee with no idea what their company does.
Or a DJ who doesn’t know where they’re playing, or who the audience is.
2. Define your structured data
Structured data means numbers in rows and columns. It can come from your ERP (the system where your company records its money), but it doesn’t have to. An Excel table, a Google Sheet or a database all count. Export it as Excel, or as a CSV file. A CSV is the plain text version of a spreadsheet. Or connect your AI tool straight to the source with a connector or plugin.
Just be careful with connectors. AI can only hold so much information at one time.
Tip – Tell it where to look. In your project instructions, add one line like: “Use the ERP connection for customer data only. Ignore the rest of the system.” We showed a similar trick for Gmail labels and Drive folders in our September newsletter: This CFO refuses to connect AI to his inbox (3 mistakes to avoid).
Note – I’ve a way for you to handle larger datasets at the end, it’s a bit more advanced, but more than achievable.
Also, make sure you don’t use pretty reports as context. AI hates human-style reports with merged cells. Raw cleanly formatted data is best.
Follow these rules. One transaction per row. One thing per column. The same format for every date, number and currency. No subtotals, no merged cells, and no two values in one cell.
Here is what good looks like:
Date | Customer | Product | Amount | Currency
2026-09-14 | Acme | Plan A | 1,200 | EUR
2026-09-15 | Borel | Plan B | 3,400 | EUR
Tip – Put the date on every source. If two files disagree, you can make sure the agent uses the newest one. Put the date in the file name, like sales_2026-09-30.csv.
Without clean rows, a "Q3 Total" line in the middle of your data gets counted as one more sale.
Without clean rows, a DJ doesn't know what track comes in what order on the record.
3. Define your unstructured data
Unstructured data is text that isn't in rows: emails, contracts, call transcripts, PDFs.
A token is a small piece of text that the AI counts. More tokens means you hit your daily usage limit soon.
PDFs are token heavy, because the AI reads the structure, the images, the fonts and the hidden information too. One PDF page can use 1,500 to 3,000 tokens.
The same words as markdown cost 200 (markdown is just plain.txt with specific heading symbols, like #, ##, ### etc)
Pick one way to convert:
Ask the AI to convert the file once.
Use the free markdown converter from Microsoft.
Paste the text into a Google Doc or Word file, then save it as markdown.
Also super important – Never ever treat a summary as a source. Keep the original contract in the folder, not only your notes about it.
4. Write a data guide
A data guide is a short note that explains your files to the AI. Think of the legend on a map. Or a timestamp for a DJ to know when to cut the record over.
Write four parts: what each file is, what each column means, what the AI must never add up, and your own definitions. Here is a short one:
# Data guide
sales_2026-09-30.csv: one row per sale, from our ERP.
Amount: price before tax, in the Currency column.
Never add the Q3 column to Jul, Aug and Sep. It is their total.
Net revenue means sales minus refunds and discounts.
Why the "never add" line? Some sheets have January, February, March, then a quarter total. The AI may want to add all four columns. It may also compare columns you shouldn't compare.
So test it. Before any analysis, type: "Tell me how you understand this data. Do not analyze yet."
If it's wrong, correct it. Then ask for the work.
Now point the agent at the folder
You can use several tools for this: Claude (Cowork), ChatGPT Work, Copilot Cowork.
Up until now, most AI tools have split their interface into 2 views:
Chat – Where you go back and forth with the AI, and collaborate together
Cowork / Work – Where you give the AI instructions, for it to carry out by itself, checking in with you as it goes
On 16 September, Anthropic brought Claude Chat and Cowork together in one experience. The rollout is starting with Pro and Max plans.
Using Copilot Cowork? It reads .md files, so your markdown files already fit. Just keep the folder in OneDrive or SharePoint, because it can’t open files saved on your computer.
And all you have to do is point them at the folder with your context in it:
Build the folders. Make a folder called workspace. Inside it, make three: context, project and output.
Fill them. context holds your business page and your data guide. project holds the files for this job, like your CSV and converted contracts. output is the only place the agent saves its work.
Back up first. Copy everything to another folder. Cowork can delete files in a folder it can reach.
Choose the folder in your tool. Cowork asks which folder to use. It warns you that everything in it can be deleted and edited. In Claude Projects, you can connect a Google Drive file instead.
Give it four rules.
1. Read the data guide first.
2. Use only the files in the project folder.
3. Write only to the output folder.
4. If two files disagree, the newest one wins.
We’ve given the example of one folder here.
But now imagine you have multiple folders, with context that’s easy for AI to understand. You can then point your AI at multiple folders, to carry out tasks across different datasets and task types, without having to worry about it guessing.
Advanced: when the folder gets too big
Say you have 500 files. The agent can't read them all. So you build a search tool for your files. Tech people call it RAG. It works in four steps:
Cut each file into small pieces. Each piece is a chunk.
Put a label on every chunk.
Store the chunks in a searchable list. That list is an index.
When you ask a question, the AI pulls the best chunks and reads only those.
Labels matter. Take a chunk that says "revenue grew 3%". Which company? Which quarter? Add a label: "Acme Ltd, Q3 report, page 4". Now the AI knows. Tech people call this label layer the "semantic layer".
And the tool flow is roughly Source Data → Database (like Supabase) → AI Workflow with RAG (n8n) → Connected to Claude/ChatGPT.
NIKI, our AI Finance Club knowledge search tool, works this way. It keeps articles, masterclasses, podcasts, YouTube videos and meeting transcripts in separate databases. Every result comes back with a title, a source link and timestamps.
If you’re interested in setting up something like this. Speak to your IT team, or join the AI Finance Club and get support from our experts.
The One Thing to Remember
Autonomy without context is a recipe for poor performance.
And a workspace that isn’t AI friendly increases the risk of wrong results, inaccurate numbers and increased AI costs.
Spend more time on your context, and you won’t have to worry about AI Agent accuracy again. Get AI to know your business, and your team will be dancing all night with or without Barbie Girl 😉
Best,
Your AI Finance Expert
– Nicolas
P.S. The best way to learn from me directly is to join my weekly masterclasses. 60 mins + Q&A where I'll get all your AI in Finance questions answered. Join me here.
