Puzzle

Puzzle is accounting software for startups, small businesses, and their accounting teams that combines a general ledger with AI-assisted categorization, reconciliation, accuracy review, close checklists, and configurable tasks. The current public material supports a governed workflow: the system prepares work from connected financial data, while a person reviews and approves the result before it is finalized. [S1][S3][S4][S5]

It is worth shortlisting for a team with a maintained chart of accounts, reliable source feeds, recurring close work, and accountable reviewers. It is a poor choice for a team seeking unattended accounting judgment, or one that cannot show how a transaction enters, changes, is approved, and is corrected. For accounting firms, an added diligence item is material: Accrual says it has agreed to acquire Puzzle's accounting-firm technology and business, while Puzzle's direct startup and small-business operation is expected to continue independently. Confirm the deal status and the actual product, contract, support, and data-migration path before committing. [S6]

Best for: US startups, small businesses, and accounting teams with connected modern finance tools, repeatable close work, named reviewers, and a controlled pilot process.

Not for: A buyer that needs multi-country or unsupported workflow coverage without confirming it; cannot maintain source data or accounting policies; expects agents to replace approval; or needs firm-facing product continuity before the Accrual transaction is clarified.

Prepare and triage transaction categorization

• Trigger: Connected bank, card, billing, payroll, or revenue activity arrives during the period.
• Inputs: Authorized source connections, chart-of-account mapping, transaction context, accounting policy, and exception thresholds.
• Output: Prepared categorizations and an exception set for review. Puzzle describes AI categorization and its close product includes an uncategorized-transactions task. [S1][S3]
• Human checkpoint: The finance owner checks the source, account, entity, period, tax treatment, supporting context, and any change to a learned or saved mapping.
• Success measure: Correct-categorization rate, reviewer override rate, uncategorized-item age, rework, and post-close corrections.

Reconcile connected accounts

• Trigger: A close period reaches the reconciliation step or new source activity arrives.
• Inputs: Connected account records, ledger activity, statement or source evidence, matching logic, tolerance, and an assigned account owner.
• Output: Reconciliation status, matched work, and discrepancies for investigation. Puzzle documents reconciliation and a pre-built bank-reconciliation-review task. [S1][S3][S5]
• Human checkpoint: An accountant confirms completeness and cut-off, investigates breaks, evaluates the proposed treatment, and approves the conclusion.
• Success measure: Correct-match rate, unreconciled balance, exception age, false matches, investigation time, and reviewer rework.

Run an accuracy review before close

• Trigger: The close checklist reaches final review.
• Inputs: Current-period ledger records, prior approved mappings, close checklist, reconciliations, and open exceptions.
• Output: A discrepancy list and actionable follow-up items. Puzzle describes a month-end accuracy task that scans platform records and surfaces discrepancies. [S3]
• Human checkpoint: The controller or assigned reviewer determines whether each item is an error, expected activity, missing support, or an unresolved issue that blocks close.
• Success measure: Material issue detection, false-positive rate, unresolved-item age, post-close adjustments, and time to close.

Configure and review a recurring close task

• Trigger: A team needs the same preparation step each period, such as receipt capture, a reconciliation check, or a client request.
• Inputs: A plain-language instruction, permitted records, expected output, reviewer, deadlines, and an exception path.
• Output: A saved AI or manual checklist task with a reviewable result. Puzzle documents templates, custom tasks, and saved tasks, and describes accountant-defined workflows that become repeatable agent steps. [S3][S4]
• Human checkpoint: The assigned accountant reads the instruction before running it, evaluates the output and evidence, and approves, edits, rejects, or reroutes the work.
• Success measure: Completion rate, evidence completeness, reviewer agreement, errors by task type, rework, and time saved after review.

1. The buyer connects selected finance sources, maps accounts, validates opening balances, and identifies the accounts, policies, owners, and exceptions in the chosen close procedure. Puzzle lists guided onboarding and native finance-tool connections, but the buyer must prove its own connection and data scope. [S1]
2. Puzzle reads the permitted transaction and ledger context, prepares categorization, reconciliation, checklist, or accuracy-review work, and surfaces discrepancies or prepared output. [S1][S3][S5]
3. A reviewer checks the source evidence, accounting interpretation, mapping, period, entity, and exceptions. Puzzle states the work is traceable and reviewable, and its AI Close help article states that no change occurs without explicit approval. [S1][S3]
4. The authorized person approves, edits, or rejects the result. Keep the source snapshot, task instruction or mapping, output, approval record, correction, and close workpaper so the team can reproduce or unwind the decision.
5. If a source feed fails, policy changes, a mapping becomes stale, an exception cannot be explained, or the reviewer cannot reproduce an output, stop the automated path and complete the scoped work manually.

The key control is a readable chain from source transaction through configured logic and reviewer decision to the final ledger state. A polished suggestion is not evidence that its source, accounting treatment, or approval is correct.

Verified facts from current public material: Puzzle describes a general-ledger platform with connected finance tools, AI categorization, reconciliation, accuracy review, AI Close checklists, custom tasks, and a stated human-approval boundary. Its pricing page publishes plans and service options. [S1][S2][S3][S4]

Vendor claims: Puzzle claims automation, time, close-speed, and capacity improvements, including pricing-page guarantee terms. Those claims depend on stated conditions and are not a prediction of a buyer's outcome. [S1][S2][S5]

B2Bagents assessment: Puzzle is best evaluated as a controlled preparation and close-work layer for teams whose source data and accounting process are already defined. The fit weakens sharply if the buyer has fragmented sources, loose mapping ownership, unclear approvals, or an unresolved correction path. For accounting firms, the pending Accrual transaction introduces a current continuity question that should be closed before deployment. [S6]

Unknowns: Public sources reviewed do not establish the exact connector list and authority for a buyer, data-field coverage, write behavior, service-account scope, implementation plan, contractual SLA, report scope, DPA, AI-model and subprocessor terms, residency, retention, deletion, audit-log export, support access, or exit terms. The agreement discussed by Accrual also does not establish the buyer's future contract or migration outcome. [S1][S6]

Good-fit conditions: supported source systems, clean data ownership, a maintained chart of accounts, recurring close procedures, documented review responsibilities, stable accounting policy, and a low-risk workflow that can be checked against a historical close.

Poor-fit conditions: unconnected or unreliable source data; material entries with no timely reviewer; frequent policy or mapping changes; a need for unverified jurisdiction, multi-entity, or multi-currency behavior; or a desire to automate accounting judgment and posting without controlled approval.

Likely failure modes: a missing or duplicated source record produces a plausible but wrong categorization; an outdated mapping selects the wrong account; a reconciliation discrepancy is normalized away; a prepared entry lacks sufficient support; the wrong reviewer approves a result; a connection or task changes without a tested fallback; or firm-facing plans change during the Accrual transaction.

Controls: assign data and accounting owners; version mappings and task instructions; use segregation of duties; run historical parallel tests; sample results; preserve source-to-output evidence; alert on changed permissions and post-approval edits; test rejection, correction, reversal, and connection revocation; and retain a manual workpaper until the pilot meets its threshold.

Start with one account group or repeatable close task. The controller owns accounting policy and close approval; the accounting manager owns the procedure; finance systems owns connections and access; source-system owners validate completeness; and security/procurement own diligence and contracts. Puzzle lists connected systems and onboarding steps, but public material does not establish the buyer's available fields or permissions. [S1][S2]

For each source, document the connection method, accounts, fields, sync schedule, period cut-off, missing-data handling, retries, access roles, external write authority, notifications, export, revocation, and recovery. Keep the existing workflow in parallel across a representative historical period. The exit plan should include exporting relevant ledger records, mappings, task instructions, support, approvals, and audit data; revoking integrations; completing open work manually; and confirming provider responsibilities should the firm-facing transaction change the service path. [S6]

Puzzle states that it uses role-based access, encryption in transit and at rest, audit traceability of activity and approvals, and SOC 2 compliance. Its close help article states that agents provide suggestions and do not change records without explicit approval. These are vendor statements, not a substitute for contract and security review. [S1][S3]

Request the current report and scope, DPA, model-provider and subprocessor list, data-flow and residency documentation, retention and deletion terms, role and approval configuration, support-access model, incident commitments, audit-log export, integration service-account model, write boundaries, and emergency disablement procedure. Verify whether every input, proposal, approval, edit, and final write is attributed and exportable for the intended workflow.

Puzzle's current pricing page lists software plans at $30/month for Starter, $72 for Core, $120 for Complete, and $360 for Scale. The displayed packages differ in included credits, seats, reconciliation and accuracy-review features, integration limits, onboarding, support, and subledgers. It also lists managed service, migration, catch-up bookkeeping, R&D-study, and tax-filing options with separate conditions. [S2]

The published close-time guarantee has eligibility requirements: timely onboarding, supported live API connections, use as the primary ledger through a complete monthly close cycle, and other qualifications. Treat it as a vendor offer, not a pilot threshold. Confirm billing cadence, discount status, eligible plan, credits and overages, add-ons, implementation, support, renewal, cancellation, data export, and transition assistance in writing. [S2]

1. Scope: One account class or recurring task with stable sources, a documented policy, named preparer and reviewer, and a correction path.
2. Baseline: Record volume, current preparation and review time, close day, exception rate and age, correct-match rate, rework, post-close adjustment rate, missing-support rate, and cost.
3. Test: Run Puzzle and the existing process in parallel on historical and current samples. Include missing, duplicate, late, reversal, timing, mapping, source-failure, permission, and out-of-threshold cases.
4. Pass threshold: Set accuracy, evidence completeness, reviewer override, exception-handling, closing-time, permission, audit-export, recovery, and total-cost thresholds before activation.
5. Stop conditions: Any unapproved change; unexplained or unreproducible result; material unresolved difference; missing source support; failed reviewer handoff; lost audit record; untested correction or access revocation; or firm-facing product continuity that no longer matches the signed agreement.

Compare Puzzle with Numeric, FloQast Transform AI Agents, Maxima, Campfire, Rillet, and Basis in the directory. The decision axis is the role Puzzle plays as a general ledger plus AI close layer for a modern finance stack; compare system-of-record dependency, supported sources, workflows, human approval, journal or posting authority, data lineage, implementation burden, service model, price unit, security evidence, and firm-product continuity.

1. What exactly can Puzzle read, prepare, change, or post in our tenant? Demonstrate one workflow from source data through review, approval, correction, log export, and access revocation. [S1][S3]
2. How are incomplete, duplicate, late, or misclassified transactions handled? Run difficult historical samples and show the exception queue, evidence, reviewer decision, and final correction. [S3][S5]
3. What creates a traceable audit record? Show the source transaction, mapping or task instruction, generated work, review, approval, edit, final state, and export. [S1][S3]
4. Which plan and credits are required for our real workload? Confirm plan price, credits, overages, seats, integrations, service add-ons, onboarding, support, and guarantee eligibility in writing. [S2]
5. Which security claims become contractual commitments? Request the scoped SOC 2 report, DPA, subprocessor and model details, retention, residency, support access, roles, logs, and incident terms. [S1]
6. How does the Accrual transaction affect an accounting-firm deployment? Confirm whether and when the deal closes, which entity contracts and supports the product, roadmap, data movement, pricing, migration, and exit choices. [S6]