The bank feed is where “the firm keeps control” stops being a philosophy.
Coding bank activity is the one part of bookkeeping that could touch hundreds of transactions while nobody is watching. So the controls are the feature, and this page describes them precisely, including what they cost you.
What this means for your firm.
Numbers Game codes bank-feed activity in QuickBooks Online under firm-set controls: every connection starts at propose-only, the confidence bar is the firm’s to set between 0.75 and 1.00, confidence is computed outside the language model rather than reported by it, and a list of structural conditions holds a transaction no matter how confident anything is.
The one place automation could do real damage.
Bank activity is the part of bookkeeping that could code hundreds of transactions while nobody is watching. Every other capability in Numbers Game proposes work a person then approves; the bank feed is the one where the volume is high enough that “a person approves it” has to mean something more specific than a button.
So the useful question is not whether it works. It is what it refuses to do, and who decides. The rest of this page is that answer, including the part that costs you an hour per client.
The firm sets the autonomy, per connection.
Not per firm, and not per client. Per bank connection, because that is the level at which risk actually varies.
| Mode | What posts | Typical use |
|---|---|---|
| Propose only (default) | Nothing at all | Every new connection starts here. Read the proposals for a cycle before you move. |
| Bank rules only | Lines matching the firm’s own rules | A client whose recurring merchants are already mapped. |
| Above the confidence bar | Lines scoring at or above the bar the firm set | A seeded client with a reviewed rule set. |
| All clearing checks | Everything that passes the structural checks | The widest setting, and still bounded by the holds below. |
A client with several bank connections is governed by its strictest one, not its loosest. That is the opposite of how most permission systems compose, and it is deliberate: the safe default when two settings disagree is the cautious one.
The firm sets the confidence bar, and the floor has a reason.
The bar defaults to 0.90. An owner or admin can move it between 0.75 and 1.00.
The floor is not a round number someone liked. 0.75 is the highest a transaction can score when the bank supplies nothing but a raw payment descriptor, so a bar set below it would be posting on the amount alone. Lowering the bar is treated as an escalation and takes the same confirmation as widening the autonomy mode.
The model does not grade its own homework.
This is the question we would put to every AI vendor in this category, and the answer separates them faster than any feature list.
A language model asked to rate its own certainty will hand you a number that reads as calibrated and is not. Confidence here is computed outside the model, from four things that can be checked:
- Agreement with the firm’s own bank rules. Your coding standards, held per client, are the strongest single input.
- How consistently that merchant has been coded before on this client’s books.
- The quality of the bank descriptor. A clean merchant name and a raw payment string are not worth the same.
- The aggregator’s own categorization signal, treated as one input among four rather than as an answer.
None of those is a number the language model reports about itself. If a vendor cannot tell you where their score comes from, the score is a vibe with a decimal point.
Your bank rules are the firm’s coding standards, and they expire.
Bank rules are held per client: this merchant books here, this vendor is never billed. They are taught by confirming proposals rather than filled in on a form, so the rule set grows out of decisions your team already made.
The product then asks the firm to re-review the whole set every five weeks. That cadence exists because we watched our own rules quietly stop being true: a standing instruction outlives the reason somebody wrote it, and a rule that keeps matching a descriptor which now means something different is worse than no rule at all.
Some things are held no matter how confident anything is.
These are structural. They are not scored, they are not overridden by a high confidence number, and widening the autonomy mode does not reach them.
- An unmapped register, or an accounting basis nobody has stated
- Foreign currency
- A rule pointing at a different account than the proposal
- A failed ledger lookup
- An unidentified card payoff
- Unbalanced splits
- An entry already in the books
- Payroll, which is cleared and never coded: the provider posts the journal that splits the run into wages, employer taxes and deductions, and the bank line’s job is to clear that journal, not to become an expense
When a person approves a held line, the books record that it was a person.
Approval is not a silent override. The approver and their stated reason are written onto every resulting entry, so the record answers “who decided this, and why” without anyone reconstructing it from memory a year later.
And approving lifts the confidence bar for that line and nothing else. Every structural check above still holds. There is no gesture in the product that turns them off.
Budget an hour per client, once, or the feed will look broken.
A newly connected bank has no rule history and no merchant history. An unseeded first run therefore holds most of the batch, which is the system behaving correctly and looks exactly like the system not working.
Expect roughly an hour per client, once, reading the recurring merchants out of imported history and saving how that client has already been coded as rules. Firms that skip that hour conclude the feed does not work, and they are not wrong about what they saw. They are wrong about why.
We would rather quote the seeding hour up front than let you discover it in week two. It is a real onboarding cost and it should be in your plan.
157 posted unattended. 332 held.
From a week against real client banks: on a client whose rules were already in place, 157 lines posted unattended. On the same client’s unmapped merchants, 332 were held.
That is the system behaving correctly in both directions, and it is the right way to read the feature. A number that only went one way would mean the controls were decorative. The feed is a confidence mechanism, not an autopilot.
What it structurally cannot do.
These are tool-level exclusions rather than policy promises, which is the difference between a limit and an intention. Numbers Game cannot permanently delete QuickBooks records, only void or mark inactive. It cannot change Intuit account settings. It cannot initiate payments or ACH. It does not touch payroll filings.
Read more on how access is scoped, logged and revoked, or see the wider throughput picture in seven ways to accelerate transaction creation.
The details firms ask first.
Does the bank feed post transactions automatically?
Only if you move it there. Every connection starts at propose-only, which posts nothing at all. A firm widens it deliberately, and a client with several bank connections is governed by its strictest one.
Where does the confidence score come from?
It is computed outside the language model, from agreement with the firm’s own bank rules, how consistently that merchant has been coded before, the quality of the bank descriptor, and the aggregator’s categorization signal. It is never a number the model reports about itself.
Can we set our own confidence threshold?
Yes. It defaults to 0.90 and an owner or admin can move it between 0.75 and 1.00. The 0.75 floor exists because that is the highest a transaction can score when the bank supplies only a raw payment descriptor, so a lower bar would be posting on the amount alone.
What is never coded automatically?
Unmapped registers, an unstated accounting basis, foreign currency, a rule pointing at a different account, a failed ledger lookup, an unidentified card payoff, unbalanced splits, an entry already in the books, and payroll. Payroll is cleared against the provider’s journal, never coded as an expense.
How much setup does a new client need?
Roughly an hour, once. A new connection has no rule or merchant history, so an unseeded first run holds most of the batch. That hour is spent reading recurring merchants out of imported history and saving how the client has already been coded.
Is the bank feed available to my firm today?
Yes. Every new connection starts at propose-only, which posts nothing, so you can run it alongside your current process and read the proposals before you widen anything.
Bring a real client workflow. See what changes.
Book a walkthrough to see Numbers Game on the work your team is doing now.