Before you paste the ledger into a chatbot
Public AI assistants are useful. The question for a finance team is where the data goes, what the model can see, and who decides what reaches the books.
Somewhere in most finance departments, someone has pasted an aged payables listing into a public chat assistant and asked it to tidy the columns. It worked. It usually does. The assistants are good at summarising, reformatting and explaining, and it would be odd to pretend otherwise.
The difficulty is not whether the tool is clever. It is that the listing has now left the systems your company controls, under terms your company did not negotiate, with no record on your side of what was shared or why.
What changes when the data leaves
Inside your accounting system, data has an owner, a set of people allowed to see it, and a history of who changed what. Paste it into a personal chat account and none of that travels with it.
- The terms belong to someone else. How long a conversation is kept, who at the provider can see it, and whether it may be used to improve the service depend on the product and the account settings. Few finance teams can answer those questions for every tool their staff use.
- There is no company record. If a vendor’s banking details or a customer’s balance went into a personal account, the company has no log of it, and no way to withdraw it later.
- The model only knows the extract. An answer built from a pasted slice of the ledger cannot check Spire, cannot see the invoice behind a line, and has to be typed back by hand. The copying that AI was meant to remove comes back at both ends.
None of this means the tools are unsafe for everything. Drafting a polite collections email from scratch, or explaining a revenue recognition rule, puts nothing of yours at risk. The line is the company’s own data.
A short policy beats a ban
Banning AI outright tends to push its use out of sight. A short, specific policy does more. It can fit on a page:
- Sort data into a few plain classes: public, internal, confidential. Customer, vendor, payroll and banking details are confidential.
- Name the AI tools the company has approved, and for which classes of data.
- Never paste credentials, bank account numbers, personal information or payroll into a personal account, whatever the tool.
- Prefer AI that works inside your systems, with your permissions and your audit trail, over AI that needs the data carried to it.
The useful question is not whether the model is clever. It is who decides what gets written to the books.
What to ask of an AI that works on your books
Once AI sits inside the finance systems rather than beside them, the questions change. These are the ones worth putting to any vendor, including us:
How Ingrid answers them
Ingrid keeps your data private: it uses a fast model by default and a stronger one for the harder steps. Documents are read by a single approved model, and if that model is unavailable the request fails rather than going to another provider.
Ingrid never gets unrestricted access to your data. It reads only through named tools that check your company and your Ingrid role, and in Ask Ingrid each person can link their own Spire login, so Ingrid sees and does only what that Spire user can.
It can prepare a change, such as a quote, a purchase order or a draft invoice, but nothing is written until a person confirms it. Ingrid’s own code then carries it out, once, and reads the record back from Spire. Clicking confirm a second time returns the same result instead of a second entry.
Every step, from the prompt and the model chosen to the proposal, the confirmation and the execution, is recorded in an audit trail for your company, and the audit records of approvals, postings and Ingrid’s steps cannot be edited or deleted; the system refuses it.
Text inside documents is treated as data, not as instructions. Names, memos and document text are marked as untrusted before the model sees them, and Payables warns when a PDF carries hidden text or instructions aimed at Ingrid.
Each company’s data is kept private to that company, and passwords and sign-in details are kept out of logs. The security page covers the rest.
The short version
Use public assistants for public things. For the company’s own numbers, use an assistant that comes to the data under your rules, prepares work instead of doing it, and leaves a record a controller can read.