AI Finance Assistants: How Companies Give the Whole Team Answers Without a Finance Login

By Stony Grunow, CoFounder at Kipper

Here is a scene that repeats in most companies several times a day, and nobody writes it down.

A sales rep is fifteen minutes out from a renewal call. She needs one fact: did this account pay last month's invoice, or is it still open? She has no login to the accounting system, so she messages finance and waits. Some days the answer lands before the call. Some days it lands after.

Now picture the same small stall happening to a support agent chasing a purchase order number, a warehouse lead checking stock on a single SKU, a project manager who just wants to know whether a vendor bill went out. None of them can see the data themselves. All of them ask someone who can.

It adds up faster than anyone realizes. McKinsey put a number on the broad version of this problem years ago: people burn close to a fifth of the week just hunting for information, or hunting for the colleague who has it. In finance the maddening part is that the information already exists. It is sitting in NetSuite, QuickBooks, or Xero, fully formed, behind a login almost nobody outside the finance team holds.

That gap is where a new kind of tool has shown up. The label that has stuck is the AI finance assistant, and the idea behind it is almost boringly simple: let anyone ask a question about the company's financial data in plain English and get the answer back in seconds, without ever opening the accounting system.

Simple to describe. Historically very hard to build. Here is why it works now.

Why the numbers stayed locked up for so long

Blame licensing first, since that is the obvious culprit. A NetSuite seat often runs north of $1,200 a year per person. No finance leader is going to hand that to two hundred people in sales and the warehouse so they can occasionally check an invoice. QuickBooks and Xero seats cost far less, but their permission settings are blunt, and none of these systems show up where the work actually happens, in Slack, in Teams, in a text message.

Cost is only half the story, though, and honestly the less interesting half.

Give a busy executive a NetSuite login tomorrow and there is a real chance it never gets used. Not because they are not sharp. Because these systems were built for accountants, and they assume you already know the chart of accounts, already know which of forty reports answers your question, and have ten free minutes to run it. For everyone else that is a locked door with the instructions written in another language. So the question becomes a favor, asked of the same finance team, over and over.

What actually changed

Talking to your data has been promised for a decade, and for most of that decade it disappointed. The chat box was never the hard part. The wiring was: connecting a language model to a live business system safely, with real permissions, without hand-building a fragile integration for every single tool.

Late in 2024, Anthropic released the Model Context Protocol, an open standard for precisely this. Instead of a custom bridge per app, MCP gives an assistant a consistent way to see what a system can do and to read from it under controlled access. That sounds like plumbing, and it is, but it is the reason this current wave feels different from the chatbots before it. The model is not improvising. It is querying a real system through a defined door and reporting back what it found.

What these assistants are good for, and where they should stop

This is the part worth being blunt about, because the category attracts a lot of overselling.

The genuinely useful version does lookups. Plain, factual lookups. Is invoice 4021 paid? What does this customer still owe us? Which of this vendor's bills are overdue? How many units of a part are on the shelf? You ask in Slack or Teams or over SMS, and the number comes straight from the connected system. A good one is read-only and inherits the permissions you already set, so it can tell someone a figure without ever letting them change it, and without surfacing anything they should not see. Kipper, the product I work on, does exactly this against NetSuite, QuickBooks, and Xero, so a rep or a warehouse manager gets the fact they need without a finance seat. NetSuite is where this tends to bite hardest, since the seats are the priciest and the system is the hardest to learn cold, so if you want the technical side, we put together a NetSuite MCP guide that compares the ways to connect an assistant to it. It is one of a growing set of fintech tools built on the same premise.

Now the harder, more important half: where these tools should stop.

An AI finance assistant is not a stand-in for the finance team's own software, and any vendor worth trusting will say so plainly. It should not close your books. It should not reconcile a bank feed, generate a P&L, or forecast cash. Those are workflows, run by trained people inside the finance system, not questions you fire off from a chat window. The moment a demo promises to "spot anomalies" or hand you board-ready analysis from a one-line prompt, slow down and ask how, exactly. The honest ones keep to their lane: direct questions, direct answers, pulled from real records.

How to size one up before you commit

If you are weighing a tool in this space, a handful of questions sort the serious ones from the demos.

Is it strictly read-only, and does it respect the roles and permissions you have already set up? Where does your data actually get processed, and what gets kept afterward? Which systems does it read, and which records inside them, because "invoices and payments" is a very different promise from "your whole ledger." Can people use it inside Slack, Teams, or SMS, or does it just add one more dashboard nobody opens? And, quietly the most telling one: does the vendor draw the same line I just did between looking things up and doing analysis, or does the pitch promise the moon?

Ask those five, and the honest tools tend to out themselves quickly.

The real shift

Step back from any single product and the pattern is bigger. For decades, who could see company data came down to who held a license, which made sense back when the only way to read the data was to operate the system storing it. Natural-language access, standing on open protocols like MCP, quietly retires that rule. The question stops being "who has a login" and becomes "who has a question."