An AI agent quoting inside an ERP gets the tariff right, checks stock, and creates the order without anyone having to review it. The same kind of agent, pointed at a pricing spreadsheet that a person maintains by hand, starts inventing columns that do not exist, mixing up two different versions of the same file, or returning a price that changed three months ago. The second model is not worse. The place where the data lives is different, and that difference decides whether the project works before a single line of code gets written.
The question almost nobody asks before hiring an agent
When a company considers putting an AI agent into its operation, the conversation almost always circles around the model: how "smart" it is, whether it hallucinates a lot or a little, whether it understands the language well. That is the wrong question, or at least one asked too early. The one that actually predicts whether the project works is different: where does the data that agent will touch live, and who guarantees that data always has the same shape?
An ERP or a CRM impose structure by design: every product has the same fields, every customer has a single record, every price follows the same calculation rules. A spreadsheet imposes nothing. Every person who edits it can add a column, merge two cells, write "N/A" where someone else wrote "0" and a third person left the cell blank. An AI agent cannot tell the difference between "the data does not exist" and "the data is written wrong": in both cases, unless someone told it explicitly, it has to guess. And guessing is exactly what an AI agent should not do with real data.
Why it works in the ERP: two real cases
In an AI sales agent project for an industrial distributor, the agent quotes over WhatsApp inside Odoo across a catalog of more than 50,000 products (full case). It works because every one of those 50,000 products has the same field structure, the same rule for calculating the tariff per customer, and the same system checking stock in real time. The agent does not have to interpret anything: it validates the customer, applies the rule, checks stock, and creates the opportunity. None of those steps require the model to "understand" the business; they require the data to already be in order.
In another project, for a group of pharmacies, the AI reasons over more than one million rows from several different pharmacy branches (full case) and ended up reconstructing a sale that the ERP was calculating incorrectly, validating it down to the cent. That is only possible because, before the AI touched anything, the data from the different branches was unified into one place with one structure. The agent did not fix the data through intelligence: the data was already in a state where an agent could reason over it without guessing.
Why it falls apart in Excel: three concrete reasons
- There is no single source of truth. A spreadsheet gets copied, forwarded by email, saved as "final version" and then "final version 2". The agent working on one of those files has no way of knowing which one is current.
- The structure changes without warning. A column called "Price" today becomes "List Price" tomorrow because someone edited it their own way. For a person, that is a minor change; for an agent that interprets columns by name, it is a field that just disappeared.
- Blank cells do not always mean the same thing. An empty cell can mean "not applicable," "not filled in yet," or "it is zero." The agent has to decide which of the three, and when it decides wrong, it does not fail with a visible error: it gives a confident, wrong answer, which is the most expensive kind of failure because nobody catches it in time.
The decision rule
An AI agent performs where the data has a structure that does not depend on a person respecting it every time. You do not need an expensive, full-blown ERP: you need a single source of truth, with a fixed format, that the system, not the good will of whoever edits it, forces to stay consistent. If your business today lives in a folder of spreadsheets that each person edits their own way, the first job is not hiring the agent: it is putting that data somewhere its structure is guaranteed, even if that is a simple database before a full ERP.
Checklist: is your spreadsheet ready for an AI agent?
| Question | If the answer is no... |
|---|---|
| Is there a single version, with no copies forwarded by email or chat? | The agent may be reading an outdated version without knowing it |
| Do the columns always keep the same name, no matter who edits them? | The agent loses entire fields every time someone renames a column |
| Does a blank cell always mean the same thing (never "not applicable" and "zero" at once)? | The agent has to guess the meaning, and guessing is where it starts inventing |
| Is there a person or a system preventing the format from breaking on edit? | Anyone can break the structure by accident, and the agent will not notice |
| Does the data update in one place and show the same way everywhere it is read? | Two people (and the agent) can end up looking at different versions of the same fact |
If you answered no to two or more, the problem is not which agent to hire: it is that whichever agent you pick will inherit the mess in the spreadsheet. We cover this in more detail in the signs your company's data is a mess and in why data comes before AI.
What to do if you live in Excel today
You do not need to migrate to a full ERP overnight to put in an AI agent that actually works. The path we follow at AutoBoost, in practice, is:
- Identify the real source of truth, even if it is currently scattered across several files, and decide which one rules.
- Put that source somewhere with guaranteed structure: it can be as simple as a lightweight database with a form in front of it, not necessarily a full ERP from day one.
- Migrate the editing process, not just the data: if people keep editing the old spreadsheet in parallel, the mess comes back within weeks.
- Add the agent last, once the data no longer depends on someone maintaining it by hand.
The order matters: an AI agent built on top of a messy spreadsheet does not fix the mess, it automates it. And an agent that automates a mistake makes it faster, and with more apparent confidence, than a person would, which is exactly what makes it dangerous.
If your business runs on spreadsheets today and you are considering AI to quote, invoice, or answer for you, at AutoBoost we look at where your data actually lives before we talk about which model to use. See how we work at our services or get in touch and we will look at it together.