Common issues
- I imported 200 invoices and Tally still asks me to confirm new ones. That is by design — imports give context, not authority. Only live reviews on that vendor count toward the thresholds above.
- Tally flagged a vendor as new even though I've imported its history. Check the document — most often this means the vendor name is spelled differently in the import vs. the new document. Edit the new document to use the canonical name; Tally will merge them and stop flagging.
- Firm memory isn't showing on a client's document. Firm memory shows in your bookkeeper view of a document from the first approval you record for that vendor in any of that firm's client workspaces. It matches the vendor by name, so a vendor spelled differently in another workspace has no firm memory yet, and what you recorded in one firm never shows in another firm's client workspaces.
Firm memory
If you're a bookkeeper working across several client workspaces, on any plan, Tally keeps your own firm memory in each firm you work with: for each vendor, the category you most recently approved or corrected for it across that firm's client workspaces you work in or have worked in, with the number of approvals behind it. "Adobe Creative Cloud → Software Subscriptions" — once you have approved an Adobe invoice with that category, an Adobe invoice in another of that firm's client workspaces shows you that category. Tally does not fill the category in.
Firm memory stays inside the firm. It is each bookkeeper's own record within a firm: it is not shared with your firm's other bookkeepers, and if you work with more than one firm, you keep a separate firm memory in each firm, and nothing crosses between firms. This is not cross-client data sharing. Firm memory holds vendor categories and approval counts — not raw documents, amounts, or financial data. The actual numbers, attachments, and notes for one client are never visible inside another client.
How Tally earns more autonomy
Autonomy is on Pro and above, and in a bookkeeper's own demo client workspace. Tally has four autonomy stages per vendor, each with explicit conditions. The Autonomy Dashboard a bookkeeper sees on the Memory page labels the same four levels Manual, Suggest, Auto with Review and Fully Auto.
- Watching. You decide every document. The starting stage for every vendor.
- Recommending. Tally proposes values for you to confirm or correct. Requires at least 3 live human reviews on that vendor and a live-human approval rate of at least 80%.
- Acting with guardrails. Eligible to approve a document automatically, with the decision visible and overridable. Requires at least 5 live human reviews (a separate minimum that applies to every automatic stage), at least 10 live human reviews, an approval rate of at least 90%, and a document amount at or below $2,000.
- Autonomous. Eligible at a higher ceiling. Requires at least 25 live human reviews, an approval rate of at least 95%, and a document amount at or below $10,000.
These conditions are necessary, not sufficient. Meeting them makes a vendor eligible. Every document is still checked on its own, and a missing critical field, a suspected duplicate, an anomaly, an amount above the stage's ceiling, or a recent run of reviews that went against Tally will route that document back to you. Vendor Autonomy Journey lists the checks.
The approval rate counts live human reviews only. Tally's own automatic approvals never count toward promoting a vendor — otherwise she would be promoting herself.
There is no setting to pause or switch off autonomy yourself. Your reviews are what hold a vendor back: a rejection counts against its approval rate, and a recent run of reviews that went against Tally stops automatic approval for that vendor until the pattern recovers. Rejecting a document flagged as a possible duplicate resolves the copy: it is not counted in the vendor's approval rate, review count or stage. On a plan below Pro, automatic approval pauses, except in a bookkeeper's own demo client workspace.
Imports give Tally context. Live reviews give Tally authority.
When you import history — a CSV from QuickBooks, a Xero export, a year of bank transactions — Tally absorbs that data fast. She learns who your vendors are, what amounts are typical, what categories you usually pick. That's context. It's why a freshly-imported workspace doesn't show every vendor as "new" — Tally already recognizes them.
But context isn't the same as authority. Tally won't auto-approve a vendor just because that vendor appeared 200 times in your import. Imports tell her who exists; they don't tell her how you actually want each one handled going forward. To earn that, she needs to watch you make real decisions inside the product.
The distinction has a practical consequence: an import never unlocks automatic handling. Live reviews on that specific vendor are the only thing that advances it. The exact conditions are in Vendor Autonomy Journey.
It's a deliberate balance: giving Tally a head start from your history, while protecting you from a system that confidently does the wrong thing because of bad imported data.
Reading the Memory page
- Vendors Known counts every vendor Tally has a document from, once each: names you linked count as one vendor. Show all N vendors lists the same vendors.
- Approval Rate is the share of all human reviews on your documents that were approvals, each review counted once. Rejecting a document flagged as a possible duplicate resolves the copy: it is not counted in the vendor's approval rate, review count or stage, nor in this share.
- Reviews Learned counts the approvals and rejections people have recorded.
- Autonomy Rate (bookkeeper view) is the share of all approvals and rejections on your documents that Tally made on her own. It shows a dash until there is one.
- Override rate (bookkeeper view) is the share of Tally's last 20 automatic decisions that a person overrode; a dash until one is decided.
- First seen is the date of the earliest document Tally has from the vendor, by the document's own date. An undated document counts from the day Tally processed it.
- Confidence is High at an approval rate of 90% or more, Medium from 70%, Low below that, and Learning before the first review. The line beneath compares the newer half of the vendor's ten most recent reviews with the older half; with fewer than four it reads as steady. Rejecting a document flagged as a possible duplicate resolves the copy: it is not counted in the vendor's approval rate, review count or stage, nor among these reviews.
- All Baselines (bookkeeper view) lists each figure Tally keeps for the vendor, with the number of records behind it: approval rate (reviews), average amount (the charges averaged), documents (every document except possible duplicates nobody has confirmed), documents from imports, and average processing time.
- Health (bookkeeper view) is a 0–100 score: the weighted average of the factors Tally has evidence for, each shown with its weight. A factor without evidence is left out, and the others count in proportion to these relative weights: Amount consistency 25, Low dispute rate 20, Extraction quality 15, Activity trend 15 and Anomaly-free history 15. Amount consistency is one minus the spread of the vendor's charges divided by their average, from two charges. Low dispute rate is the share of the vendor's recent decisions that were not rejections. Extraction quality is Tally's average confidence in reading its recent documents. Activity trend compares its documents in the latest 30 days of its history with the 30 days before (80% when the earlier period has none). Anomaly-free history is the share of a vendor's checked documents that Tally has not flagged as unusual. It appears only once Tally can judge that vendor's amounts (its average holds more than five amounts); documents processed before then count as not flagged. A vendor with no factor to score shows no health score. From six decisions on, the arrow shows whether the newer half of its recent decisions approved more (up) or fewer (down) than the older half, by more than 10 points.
- Intelligence Feed. A warning appears when a vendor's approval rate is below 70% after at least three reviews; it gives the rate and the number of reviews. New vendor detected appears while a vendor's first document, by its own date, is within the last 30 days. A milestone appears when a vendor reaches 5, 10, 25, 50 or 100 documents (possible duplicates nobody has confirmed don't count) and stays until it has three more. The feed shows five items, newest first; Show all N insights counts the items under the selected filter, up to 20 of each kind.
- Seasonal patterns cover vendors whose first and last dated documents are at least twelve calendar months apart (counting months, not days). A month's figure is the vendor's spending in that calendar month, averaged over the years in which it had spending in that month. Only months with spending are compared, and the average month is the mean of their figures. Peak months are up to three months whose figure is above one and a half times the average month, largest first. Strength measures how far the months' figures spread around the average month (their standard deviation, taken over those months): Strong when that is more than half the average month, Moderate when it is a fifth of it or more, otherwise Subtle. The arrow compares the vendor's spending in the twelve months ending at its newest document with the twelve months before them, counting in each only the calendar days, a year apart, that both periods cover. It points up or down when the change is more than 10%, and stays level otherwise, or when the earlier period has no spending. Expected next month shows next month's figure, in whole dollars, when next month has documents with spending. Based on partial year is added when fewer than twelve calendar months have documents, or when fewer than two of the vendor's documents fall in next month. Spending in a month that has not yet ended counts toward that month's figure once it ends.
- Open Loops files each item under one type, the first that applies: failed push, possible duplicate, possible anomaly, missing field, new vendor, awaiting review. The number beside each filter counts its items, so the filters add up to the total.
The boundaries
Three boundaries:
- No cross-client document data. Two clients in the same firm don't see each other's documents, vendors, or numbers — ever. The bookkeeper sees both because they're assigned to both; Tally never blurs the boundary.
- Firm memory stays inside the firm. If you work with more than one firm, you keep a separate firm memory in each firm, and no one else sees it. Apart from your own firm memory and review patterns, which stay inside that firm, nothing Tally learns crosses between clients, and nothing crosses between firms. Firm memory never fills a category in.
- No autonomy without your reviews. Imports don't unlock autonomy. Auto-approve needs Pro and above (or a bookkeeper's own demo client workspace), and happens only after live human reviews build the track record.
You stay in control of these boundaries through plan tier and access settings — but the boundaries themselves are baked into how Tally works, not just config you have to remember to set.
What's next
- Documents — review and approve
- Vendor Autonomy Journey — the visible four-stage indicator that shows where each vendor sits today
- Bookkeeper workspace — close readiness, ROI reports, firm memory
- Intelligence Digests — the recurring email summary of what Tally handled
- Resolution Playbooks — how Tally surfaces patterns from your past decisions
- Missing Document Detection — when expected recurring documents haven't arrived
- Plans and billing — which features are at which tier
- Security and privacy — the boundaries explained more fully
What Tally remembers per vendor
For every vendor she's seen, Tally tracks:
- Provenance. How many documents Tally has seen from this vendor — broken into "from imports" (historical context) and live reviews (your decisions).
- Typical amount. What's normal for this vendor — average, range, frequency. The typical amount shown is the average of the vendor's charges in your business's currency, under every name you linked to it. Tally leaves out rejected documents, possible duplicates nobody has confirmed, and unusual amounts until a person approves them. On a document flagged as unusual, she shows the average she compared it against when it arrived.
- Typical category and account. Which category you usually pick, which GL account you usually map to.
- Approval pattern. When you approve vs. correct vs. reject. On a document, Vendor pattern summarises the vendor's most recent decisions (up to ten): "Approved once", "Approved 2×" or "3×", then "Usually approved" from four approvals; rejections read the same way, and a tie shows no line. Most common correction is the category people most often corrected it to recently.
- Anomaly history. When a document fell outside the normal range and what you did with it.
- Open questions. Anything still unresolved (missing W-9, pending follow-up).
You can see this on the Memory page for any vendor. Tally surfaces "47 documents from imports + 12 from live reviews" as separate signals so you know what's grounding her judgment.
What you'll learn
- The difference between imported context and live review authority.
- What Tally remembers about each of your vendors.
- How Tally earns more autonomy over time (automatic approval is on Pro and above).
- What's per-business, and what a bookkeeper's own firm memory holds.
- The boundaries that prevent cross-client leakage.