Anomaly narratives
Live · all tiers. When a document is flagged, Tally states the comparison in plain language: "Stripe usually charges $12–$15/month; this $250 charge is roughly 17× higher. The last similar spike was March." The explanation draws on the vendor's recorded pattern and, on Business, the firm's history.
Ask Tally
Live · all tiers. Questions in plain language against your own records — "what did we spend on AWS this quarter?", "which invoices over $1,000 are unpaid?" — answered from stored data with the records cited. Available as a chat surface and above the document list.
Publicly, on the marketing site, Ask Tally answers questions about the product only. It has no access to any account.
Where it stops. If a surface has nothing in it, Tally says so rather than composing an answer. Each kind of evidence is a separate typed slot, so an anomaly cannot be invented from an ordinary large invoice.
Audit trail
Live · all tiers. Every extraction, autonomy decision and human action is written to an append-only log: the raw output, the reasoning, the model, the cost, the person, the timestamp. A past decision can be replayed exactly as it was made.
What you see. The activity feed, the explanation on each document, and the decision detail behind an automatic approval.
Automation estimate
Live · Pro and above. A nightly job counts the documents Tally handled automatically in the month and multiplies by a configured per-document handling time and labour rate.
This is a model, not a measurement. It is arithmetic over automated volume using assumed rates. It does not measure anyone's hours, and it is not a projection of what any business would experience. It is presented as an estimate wherever it appears, including on the report a firm can send to a client.
Boundaries
- Tally helps prepare 1099-NEC PDFs. It does not file them, transmit them, or take filing responsibility. Filing remains with the user or their accountant.
- Tally works alongside QuickBooks. It is not a general ledger and does not replace one.
- Tally does not replace a bookkeeper, accountant or tax advisor. Final tax, legal and accounting judgment stays with a qualified professional.
- Systems Tally connects to: QuickBooks Online, bank feeds via Plaid, and email ingest via Inbox. Xero, NetSuite, Sage and IIF are not connected systems — their exports can be imported as CSV history, which is not an integration.
- The standalone Inbox product delivers extracted documents to a webhook you configure. It has no QuickBooks connection and creates no bills.
- No push to an accounting system is automatic. No vendor or client email is sent without approval.
- UppaGo holds no security or compliance certification, and does not claim country-specific compliance. Regional availability of connected features varies by setup, plan, bank and local requirements.
- Human review is part of the product. Tally surfaces what needs attention and why; a person decides.
Close readiness
Live · Business. A daily score per client for the current month across four inputs: documents reviewed, bank transactions reconciled, open loops resolved, and expected recurring documents received. A failed push to the accounting system appears as a blocking item inside the score. Contractor tax completeness is tracked separately, not in this score. Clients are bucketed ready, needs attention, or significant gaps, with their specific blockers listed.
Confidence reliability
Live · all tiers. A raw model confidence score is not trusted on its own. For each vendor Tally tracks how often its extractions survive human review without correction, and derives a conservative lower-bound confidence from it. Only human-reviewed documents count; automatically approved documents never vouch for themselves. How often you approve is tracked as a separate signal from whether the extraction was correct, and the two are never combined.
Where it stops. With too little evidence, Tally reports the raw model score as-is and labels the reliability unknown rather than presenting an unearned number.
Conversation memory
Live · all tiers. Multi-turn context is preserved, so a follow-up question is understood against the previous one. Past conversations are listed and can be reopened with their sources.
Documents as facts
Live · all tiers. Every invoice, receipt, bank transaction and imported row becomes one record: vendor, amount, date, category, status, flags. This record is what every list, filter, chart, search and export reads from. Extraction happens once and is stored; the document is not re-read on each query.
Where it stops. A stored fact is Tally's reading of the document. It is editable by you, and your correction replaces it.
Evaluation before release
Internal — no user-facing surface. Changes to Tally's instructions, models or thresholds are measured against documents humans have already reviewed before they ship, and a change that regresses accuracy is blocked.
Firm memory
Live · Business. Mappings and handling a firm has recorded across its own client workspaces become available to that firm's other clients — for example a vendor-to-category mapping the firm has applied consistently.
Where it stops. Firm memory holds patterns from the firm's review behaviour, not client data. No document, amount or note from one client is visible in another, and nothing crosses between firms.
How Tally learns — context and authority
Live · all tiers.
Tally treats two kinds of signal differently, and the difference explains most of its behaviour.
- Imported history creates recognition. A CSV, a QuickBooks or Xero export, a bank-feed sync: Tally records that these vendors exist, what amounts have appeared, which categories were used. Imported documents are not flagged as coming from a new vendor.
- Live human reviews create authority. An approve, a correction or a rejection recorded by a person inside Tally is the only signal that lets a vendor progress toward being handled automatically.
Imports never advance autonomy. A vendor with five thousand imported invoices and no live reviews is treated exactly like a vendor with none: Tally recognises it, and asks you about every document.
Immediate reaction
Live · all tiers. High-severity anomalies, sudden pattern shifts and close-readiness blockers raise notifications when they are detected rather than waiting for a daily batch.
Delivery channels. Email and in-app notification.
Intelligence Digests
Live · all tiers, opt-in. A recurring email — daily, weekly or monthly — summarising what was handled, what changed and what is open. Two audiences: a business-owner digest and a firm digest. Read-only; it never acts on the books. Cadence, send time and time zone are configurable, and every digest carries a one-click unsubscribe that preserves the settings.
Meaning-based search
Live · all tiers. Matching on meaning rather than exact words, so "AWS hosting" and "Amazon Web Services Compute" are recognised as the same concept. Surfaces similar documents and suggests vendor merges where two names appear to be the same vendor.
Missing-document detection
Live · all tiers. Where a vendor has an established interval, Tally surfaces the ones that have not arrived when expected — naming the vendor, the days elapsed, the interval it has recorded and an estimated amount. Four actions are available on each: mark received, expected later, not expected, dismiss. Every action is logged.
Where it stops. It fires only where enough of a pattern has been recorded, so irregular vendors stay quiet. Detail: Missing Document Detection.
Open Loops
Live · all tiers. One prioritised queue of unresolved threads: documents pending review, anomalies awaiting a decision, missing contractor tax information, rejected documents needing follow-up, vendors that have not sent an expected document. Sorted by severity, each linking to its source.
Relationships
Live · all tiers. Tally records which category and account a vendor has been mapped to, which vendor name variants refer to the same vendor, and which client belongs to which bookkeeper. Recorded mappings are offered as pre-fills, with the count they are based on ("based on 12 prior Walmart invoices").
Where it stops. A pre-fill is a proposal. It does not approve anything.
Resolution Playbooks
Live · all tiers. Where the same kind of issue has been resolved the same way repeatedly, Tally records the resolution and offers it on a similar future document as a suggestion with Apply, Edit and Reject. It never resolves anything on its own, and it states what the suggestion is based on.
Reviewer memory
Live · Business. Where a firm has several reviewers, Tally records approval and correction patterns per reviewer, so a partner can see where team members are calibrated differently.
Where it stops. This is a firm-internal view of the firm's own reviewers. It is not shared outside the firm.
Seasonal patterns
Live · all tiers. A weekly job records timing patterns per vendor: monthly cadence, quarterly cycles, and annual spikes. These feed the cash-flow forecast, give anomaly narratives timing context, and underpin missing-document detection.
Table of contents
Trust and autonomy
Live · Pro and above. Autonomy is tracked per vendor and progresses through four levels, each with explicit conditions:
| Level | What Tally does | Conditions |
|---|---|---|
| Manual | You decide every document. | The starting level. |
| Suggest | Proposes values for you to confirm or correct. | ≥3 live human reviews and ≥80% live-human approval rate. |
| Auto with Review | Eligible to approve automatically, visibly and overridably. | ≥5 live reviews (a separate minimum that applies to every automatic level), ≥10 live reviews, ≥90% approval rate, document amount ≤ $2,000. |
| Fully Auto | Eligible at a higher ceiling; exceptions still return to you. | ≥25 live reviews, ≥95% approval rate, document amount ≤ $10,000. |
These conditions are necessary, not sufficient. They make a vendor eligible. Every document is still checked individually, and a missing critical field (amount, total, vendor or a date), a suspected duplicate, an anomaly, low extraction confidence, an amount above the ceiling, or a recent run of reviews going against Tally will route that document to a person regardless of level.
The approval rate counts live human reviews only. Tally's own automatic approvals never count toward promoting a vendor.
What you see. A four-stage indicator on each document — Watching, Recommending, Acting with guardrails, Autonomous — plus the reasoning behind any automatic decision. Pauses, disables and workspace-level shutdowns are shown as side states, not as a loss of stage. Full detail: Vendor Autonomy Journey.
Vendor patterns
Live · all tiers. For each vendor, Tally maintains a picture of what has been seen: typical amount and range, typical interval, typical category and account, approval and correction history, past anomalies. New documents update it.
What you see. Vendor cards on the Memory page ("$2,400 average · 47 documents · 90% approved"), the imported-versus-live provenance line, and the comparison behind any anomaly flag.
Where it stops. The pattern reflects what has been recorded. A vendor with little history has a correspondingly thin pattern, and Tally says so rather than inferring one.
What Tally remembers
Each part below states what it is, what you see, and where it stops. They are not separate features to switch on; a single document usually touches several.
What this looks like
| Situation | What Tally does |
|---|---|
| Fresh import of 200 documents across 11 vendors | Recognises all 11. Flags only documents with real signals: duplicates, anomalies, missing fields. |
| Five live reviews on one vendor | Those five count toward that vendor's autonomy conditions. |
| Import of a vendor never seen before | Recognised, not flagged as new. Recognition does not imply approval. |
| Import of a stale export with outdated category coding | Seeds recognition. Live reviews are what determine how the vendor is handled. |
Both counts are visible: the document detail sheet and vendor cards show imported-document counts separately from live-review counts, so you can see what a suggestion is grounded in.
Why the split exists
An import is a data dump, not a set of decisions. If Tally treated it as approval, a single bad export could set a vendor to the wrong category and then keep applying it automatically. Separating the two means a bad import can misinform recognition — which you will see and can correct — without ever silently gaining the right to act.