Puzzle

Almost everything called bookkeeping is mechanical. A transaction arrives, it gets a category, it matches something in a bank feed, and a journal entry records it — the same shapes, thousands of times, and a human doing it is a human doing pattern matching badly at four in the afternoon.

Almost everything called accounting is judgement. Whether a cost is an expense or an asset, when revenue is earned, what a variance means, and whether the numbers tell the story the business thinks they do.

Puzzle's design respects that line. Agents do the first; an accountant reviews and owns the second — which is both the correct answer and the one that keeps the books defensible.

Puzzle In Five Lines Of Accounting

• AI Close runs the month-end close with agents handling categorisation, reconciliation and journal entry preparation.
• Accrual automation for prepaid expenses and fixed assets — the part small-company books most often get wrong.
• Revenue recognition on both a cash and an accrual basis, plus invoicing that reconciles itself.
• Accuracy AI is framed as auditable, which is the only framing that matters when a regulator asks.
• An API and MCP make it agent-native, so workflows can be built on top rather than only used.

Why Does AI Bookkeeping Need An Accountant Anyway?

Puzzle describes AI agents handling categorisation, reconciliation and journal entry preparation while maintaining accountant oversight and control. That last clause is not a legal disclaimer — it is the product's architecture.

The reason is that accounting errors are asymmetric. A miscategorised coffee purchase costs nothing; a misclassified capital expense, a revenue recognition error, or a missed accrual can misstate a company's financial position for a year and be discovered by an auditor, an investor or a tax authority.

Machines are excellent at the first kind of decision and poor at the second, and the difference is not intelligence. It is that the second requires knowing things that are not in the transaction: what the contract says, what the company intends, how the policy has been applied historically.

This is also why the close is the right unit to automate rather than individual entries. A close has a defined start, a defined end and a review step built into it by convention, which makes it a natural place to insert agents whose work gets checked.

The framing of Accuracy AI as auditable is the part worth taking seriously. An automated entry that cannot be traced to a rule and a source document is a problem waiting to surface, and the ability to show why something was booked is worth more than the speed of booking it.

The same split governs the tax side of a business. Instead automates the mechanical parts of filing while a person remains accountable for what is filed, which is the identical architecture applied to a different obligation.

A miscategorised coffee costs nothing. A missed accrual can misstate a company's position for a year and be found by an auditor.

Accrual Automation Is The Hard Part

Puzzle handles accrual automation for prepaid expenses and fixed assets, and this is the least glamorous feature and the one that fixes the most real problems.

The issue is that small companies keep their books on a cash basis by default, because cash is what the bank feed shows. Money left, so it was an expense. That is simple, and it is wrong in a specific and consequential way.

Pay a year of insurance in January and cash accounting records twelve months of cost in one month. The January books look terrible, the following eleven look artificially good, and any decision made from either is made on a distorted picture.

Accruals fix that by spreading the cost across the period it covers, and fixed asset depreciation does the same for equipment. Both are mechanical once the policy is set — which is exactly why they suit automation and exactly why they get skipped when done by hand.

The consequence of skipping them shows up later and expensively. A company raising money or being acquired will have its books converted to accrual accounting by someone else, and discovering at that point that two years of monthly figures were misleading is a bad moment to have it explained.

Integrations Decide Whether The Books Close

Puzzle connects to Stripe, Brex, Ramp, Mercury, Gusto, Deel and Rippling, with QuickBooks supported for data import.

That list is precisely targeted, and reading it tells you the customer. Stripe is revenue, Brex and Ramp are cards and spend, Mercury is banking, Gusto and Deel and Rippling are payroll — which is the complete transaction surface of a modern startup and almost nothing else.

The reason coverage matters more than features is that an unconnected source becomes manual work, and manual work becomes the reason the close takes a week. One system exporting CSVs that somebody imports by hand undoes most of the automation elsewhere.

Payroll is the integration that matters most and is noticed least. It is the largest expense in most software companies, it has the most complex accounting — accruals, taxes, benefits, timing — and getting it in automatically removes both the biggest error source and the most tedious reconciliation.

QuickBooks appearing as import rather than sync is the significant detail. It signals that Puzzle expects to replace a general ledger rather than sit beside one, which is a bigger commitment than adding a tool and should be evaluated as such.

Firms scaling this way usually run a second stack for tax. Platforms such as Black Ore serve that half, and the two rarely overlap.

• Check every transaction source you have against the integration list before committing.
• Treat payroll as the integration that matters most, since it is the biggest expense and the messiest.
• Set accrual policy explicitly rather than letting defaults decide it.
• Insist that every automated entry can be traced to a rule and a source document.
• Run a parallel close before switching, not after.

Trialling Accounting Software Safely

Accounting software is unusually risky to change, because the failure mode is not an outage — it is a set of books that were subtly wrong for three months.

Running in parallel is slower and it is the only responsible way to do this.

1. Pick a closed month whose numbers you already trust, and run it through Puzzle without changing your existing process.
2. Compare the resulting trial balance line by line against what you actually reported, and investigate every difference rather than the large ones.
3. Check the accrual treatment specifically, since that is where cash-basis and accrual-basis books diverge most.
4. Verify each integration pulled everything, by reconciling counts of transactions rather than totals.
5. Have whoever signs off your accounts review the automated journal entries and confirm each is traceable.
6. Only switch after two consecutive months where the parallel close matched, and keep the old system readable for a year.

Who Puzzle Is Actually Sold To

The site names four audiences: accounting firms scaling without adding headcount, startups wanting real-time visibility, SMBs wanting accurate books, and finance teams wanting faster closes.

The firm audience is the most interesting of the four, because it changes the product's economics. A firm adopting Puzzle applies it across many clients, which means one sale covers dozens of companies and the software's leverage compounds rather than being limited to one set of books.

It also means the review capacity the design depends on actually exists. A startup automating its close may have nobody qualified to check the agents' work; a firm has that person by definition, which is who the accountant-oversight model assumes.

For a startup the pitch is different and equally real: real-time visibility. Books closed weeks after the month ended answer questions nobody is still asking, and a company whose runway calculation is based on figures from six weeks ago is flying on old instruments.

The agent-native framing — API and MCP alongside the interface — points further ahead than either. It suggests a future where an agent asks the accounting system a question directly rather than a person opening a dashboard, which is a genuine change in what finance software is for.

For the judgement questions that remain after the books close, a research tool grounded in primary authority — Blue J, for instance — is a different purchase entirely.

Puzzle: Integrations And Audience

Should Your Firm Run On Puzzle?

A fit for your books when

• Your transaction sources are Stripe, Brex, Ramp, Mercury and a modern payroll provider.
• Your close takes weeks and the numbers arrive after the questions.
• You are on a cash basis and know your monthly figures are distorted by prepayments.
• You are a firm looking to serve more clients without proportional headcount.
• Someone qualified will review the agents' work, which the design assumes.

Keep your current stack when

• Key transaction sources are not on the integration list, which means manual imports.
• You have complex, judgement-heavy accounting that agents cannot safely touch.
• Nobody in the business can review a journal entry critically.
• You are unwilling to run a parallel close before switching general ledgers.
• Your industry has accounting requirements a startup-focused product will not cover.