The phrase “AI CFO” is being attached to two products that share almost nothing. One is a chat window sitting on top of reports a business owner already has. The other is a system holding transaction-level records that can answer a question the owner genuinely cannot answer today, such as which SKU lost money last month after fees, refunds, and freight. Both get sold with the same three words. Only the second one changes what a business knows about itself, and the gap between them has very little to do with which model runs underneath.
The chat layer is a real improvement and still a UI change
A conversational interface bolted onto an accounting product earns its keep. It kills the click path. Rather than recalling that the report you want lives under Reports, then Sales, then Sales by Customer Detail, you type a sentence and a table comes back. For an owner who opens the books twice a month and forgets the menu structure between visits, that is real time saved.
What it does not do is add information. The answer is pulled from the same ledger that was already sitting there. Take an illustrative case: a marketplace deposit of $38,412 posts to the ledger as a single line called Amazon Sales. A chat layer will tell you, correctly and quickly, that the deposit was $38,412. It cannot tell you the deposit contains gross sales, referral fees, fulfillment fees, refunds, and an advertising charge, all netted together, because those components were never written down anywhere the model can reach. The model is not the constraint. The record is.
The question that separates the two
Here is the test I use. Ask the software: which SKU lost money last month?
It is an ordinary question. Any seller with more than about thirty SKUs has asked it, usually while staring at a healthy top line and a bank balance that refuses to grow. Answering it honestly requires five separate facts joined at the unit level. Units sold per SKU. Landed cost per unit for the specific inventory layer that shipped. Marketplace fees allocated per unit rather than netted at the deposit. Returns matched back to the original order, including the ones that come back three weeks after the sale. Advertising attributed to the SKU rather than the campaign.
Miss any one of those and the answer is a guess with a confident tone.
Published plan pages show where the data usually stops
You can see the boundary without talking to a salesperson. On Intuit’s published QuickBooks Online pricing page in August 2026, inventory tracking begins at the Plus plan, listed at $115 per month, while Simple Start at $38 and Essentials at $75 do not include it. On Xero’s US pricing plans page in the same month, Inventory Plus is a paid optional add-on available only on Growing at $55 and Established at $90, and Xero has posted that subscription prices increase from October 1, 2026.
That tiering is not a flaw. A general ledger has to serve a law firm, a plumbing contractor, and a coffee roaster with one data model, and unit-level inventory is a specialized load. The point is narrower and more useful: a ledger configured without SKU-level cost and settlement detail was never given the raw material to reason about SKU profitability. No amount of model quality fixes an absent field.
Category validation and the ceiling show up together
Both Intuit and Xero have been shipping AI features into their accounting products, and their plan pages make the capability tiering explicit. Two incumbents of that size moving in the same direction settles the question of whether the category is real. It is.
It also draws the ceiling in plain view. A general ledger that never received a SKU-level settlement breakdown cannot reason about SKU profitability, regardless of how good the assistant on top of it is. That is a statement about plumbing, not about product quality, and it applies equally to every ledger of every brand. The assistant answers from what the schema holds.
The unglamorous half nobody markets
Nobody runs a campaign about field mapping. The work that makes an AI CFO answer real questions is mostly ingestion and normalization, and it looks like this:
- Pulling settlement reports at the line-item level from each marketplace instead of recording the net deposit
- Mapping every fee type to a category that survives when the marketplace renames it
- Tracking inventory cost by layer so cost of goods sold reflects what actually shipped
- Matching refunds and reserves back to the originating order across a month-end boundary
- Holding a channel dimension on every transaction so multi-channel sellers can split results without a spreadsheet
- Reconciling all of it to the deposit that hit the bank, to the penny
Six unglamorous jobs. Do them, and a competent model answers hard questions from clean data. Skip them, and the same model produces fluent sentences about a number that was never broken apart. This is the reason the data-first vendors and the interface-first vendors will diverge, even though today they use identical language on their home pages.
Crunch, the AI CFO built into ConnectBooks, is in active beta and sits on that second approach, working from marketplace settlement detail and SKU-level records rather than summarized journal entries. Whether any particular implementation delivers is a separate question from whether the architecture makes the answer possible at all.
What to ask before you believe a demo
Four questions, all answerable in a live session:
1. Show me one settlement, expanded
Ask to see a single marketplace payout decomposed into its components inside the product. If the software shows one number and a category, the underlying detail is not stored.
2. Which SKU lost money last month, and why
Watch whether the answer names a cost basis and a fee allocation, or hedges toward gross margin by category.
3. What happens to a refund processed after month end
Returns that straddle a period boundary are where summarized systems quietly drift.
4. Where does the cost per unit come from
Purchase orders, landed cost allocations, and the valuation method all have to exist somewhere upstream. If the answer is “we use the average from the ledger,” you have found the ceiling.
Where this settles
The interface-first tools are not going away, and they should not. Plenty of businesses have straightforward finances where the ledger truly does contain the answer, and a conversational front end over an accurate ledger is a genuine upgrade. The U.S. Small Business Administration’s guide to managing business finances still describes the fundamentals any owner has to get right first, and no assistant substitutes for them.
For product businesses with inventory, multiple channels, and settlement-based revenue, the distinction is decisive. Buyers should stop evaluating these tools on the quality of the conversation and start evaluating them on what the database holds. Ask what data the system captures, at what grain, and how it ties back to the bank. The answer to that question predicts everything the assistant will ever be able to tell you.