FOR INVESTORS

The ledger is
the test suite.

Accounting is the one back-office function that can verify an agent's work. Mozart is the harness that runs finance workflows unattended on that proof. Deduction recovery for consumer brands is workflow one.

Talk to the founder Prefer email? parth@mozart.financial

Workflow one: deduction recovery
In development with finance teams

01 / THE PROBLEM

The back office is the last big function still run by hand.

Every money workflow in finance has the same three traits. Together they defeated RPA, and they are why generic agents stall here.

The rules are fixed

Match the invoice, find the code, gather the proof, file, post. Same steps every time. This is not a judgment problem.

The systems keep changing

Retailer portals, distributor sites, bank feeds and ERPs change forms, fields and logins without notice. Every hand-built integration breaks.

The proof arrives late

Nothing confirms success at the moment of action. The answer shows up weeks later as cash on a remittance, a denial, or silence.

Why the two existing answers failed

Approach	What it got right	Where it breaks
RPA and scrapers	Deterministic. Cheap per run.	A portal changes a field and an engineer rebuilds the bot. Maintenance eats the budget.
Managed services with humans in the loop	Handles change and judgment. Works for large accounts.	Humans are the verifier, so cost scales with volume and the humans never come out.

02 / THE THESIS

Coding agents matured first because code has a verifier. Accounting has one too.

Double-entry and reconciliation check every action against cash. That is what lets a harness run finance workflows with almost no one watching.

Software

The verifier

Compiler and tests. Every change is checked automatically, cheaply, immediately.

The agents

Coding agents got good, but each team built its own orchestration, memory and retry logic.

The harness

A scaffolding layer made the agents usable: plan, execute, verify, repair, repeat.

Accounting

The verifier

Double-entry and reconciliation. Every money workflow ends in an entry that must tie to a source and clear against cash.

The agents

Finance agents exist, but they are sold as apps with humans checking the output, because nobody uses the verifier.

The harness

Mozart. Workflows are specs; the system executes them, repairs them when a system changes, and the ledger proves the result.

The scaffolding layer for computer-use agents is widely described as an unsolved infrastructure problem. We build it for one function, where the verifier already exists.

03 / THE SPEC

Every money workflow is the same five-part spec.

The harness runs the spec. The customer's ERP, payers and reason codes are configuration. Deductions is the first spec; the next ones fill the same five slots.

1Obligation
What the ledger says is owed or expected
2Evidence
The documents that prove it
3Action
A step taken in an outside system
4Settlement
The real-world outcome, observed later
5Entry
The journal entry, and the reconciliation that proves it

Workflow	Obligation	Evidence	Action	Settlement	Entry
Deductions (now)	Short pay on a retailer remittance	Invoice, bill of lading, promo contract, portal claim detail	File the dispute on the retailer or distributor portal	Repayment, denial or silence, weeks later	Clear the deduction, book the recovery, reconcile to cash
Cash application (next)	Open invoice in AR	Remittance advice, bank line	Apply payment in the bank portal and AR	Bank match	Apply cash, close the invoice, reconcile the bank
Vendor chargebacks and bills (then)	Vendor bill or compliance fee	PO, receiving record, contract terms	Dispute or approve on the vendor portal	Credit memo or payment	Book the credit, reconcile AP

Operating rule inside every spec: code moves the money, the model only repairs a step when a system changes and reads messy documents, and the browser layer is bought from whoever wins it, not built.

04 / THE HARNESS

Accounting already solved memory, trust and failure.

These are the open problems for long-running agents. Finance has had answers for a century. We use them, and we publish three numbers to prove it.

Memory is the audit trail

The agent's memory is the case file and the ledger, not a vector store. Periods close and lock, so nothing compounds across them.

Clean-book rate

Share of recovery entries the accountant accepts unedited at month-end

Trust is a track record per step

Each step, per payer, earns its way out of human review. The gate on "file the dispute" comes off after N clean reconciliations, not on a feeling.

Unattended rate

Share of items from intake to booked recovery with zero human touches

Failure is caught by the books

A wrong action shows up as an entry that doesn't tie. The reconciliation catches it, not the agent's own judgment. When a portal changes, the model repairs the step and re-caches it.

Time to repair

Hours from a portal change to the workflow running again, without an engineer

All three numbers are instrumented from the first live workflow and reported per customer, per week. We would rather show a low number that climbs than a claim.

05 / THE WEDGE

Deductions is workflow one, and the hardest test.

Retailers short-pay brands, lean teams can't fight back, and the proof is pure cash. Every trait of the back office shows up here at full strength.

Up to 10%

of revenue lost to deductions at food and consumer brands. $300K to $1M a year for a $10M brand.

15 to 20

retailer and distributor portals a typical AR team logs into.

2 to 5 people

is often the whole finance team. They set a dollar threshold and eat everything below it.

Why this workflow first

Dozens of portals that change, rules that differ per retailer, and a result that only appears weeks later on a remittance. If the harness runs unattended here, cash application and payables are easier, not harder. And the value is cash, so nobody argues about the baseline.

What a $1,500 shortfall looks like to finance

Invoice	Payment	Shortfall	Reason on the remittance
$50,000	$48,500	$1,500	A shortage code, a promo allowance and a compliance fee. Each needs different evidence.

Deduction-rate estimates are industry-published and vary widely; the shortfall example is illustrative. Initial focus is $10M to $100M brands on QuickBooks.

06 / THE PRODUCT

Workflow one, live: deduction recovery end to end.

From the ledger to the retailer portal and back to the ledger. Human gates come off step by step as each one earns its track record.

Stage	Spec part	What Mozart does	Human gate today	Status
1. Identify	Obligation	Syncs QuickBooks / ERP (GL, trial balance, chart of accounts). Matches invoices, cash and credits. Surfaces unexplained short pays.	Grant accounting access	Live
2. Investigate	Evidence	Maps each payer's reason codes to a canonical set. Pulls delivery and pricing evidence from portals, email and uploads.	Edge cases only	Live
3. Appeal	Action	Builds the dispute packet, files it on the retailer portal with headless browser agents, tracks replies and denials.	Approve the final appeal	In build
4. Collect	Settlement	Tracks repayment and matches cash or credits back to the original deduction.	Confirm settlement terms	Live
5. Account	Entry	Drafts the accounting entries per case and posts them to QuickBooks / ERP on an approval. Reconciles to cash.	Yes / No per entry	In build

Built with thousands of automated tests and an evaluation baseline for every AI step. Automated portal filing and accounting write-back are the next two milestones.

07 / WHERE MOZART SITS

The harness between the ledger and the world.

AI-native ledgers own the books. Accounting-firm agents do the firm's work. Vertical platforms go deep on one industry's data. Mozart runs the workflows that cross into outside systems and reconcile back.

Category	What they own	Relationship to Mozart
AI-native ledgers and ERPs	The general ledger and the close: what happens inside the books	The test suite. Mozart runs workflows against them and writes the result in; QuickBooks first
Agents for accounting firms	The firm's bookkeeping, tax and audit work	Channel. Firms and fractional CFOs each bring several brands
Vertical deduction platforms	One industry's transaction data, sold as an app plus a managed service	Direct overlap on the wedge. They scale people; we scale a harness
Browser and computer-use infrastructure	The generic layer: drive a portal, repair a broken script	Suppliers. We buy the browser layer; the finance harness sits on top
The finance harness (Mozart)	Money workflows that run through outside systems and reconcile in the ledger	Horizontal engine, vertical wedge, unattended by design

08 / THE ROADMAP

Deductions is the wedge. Each spec unlocks the next.

Every workflow fills the same five slots. The harness and the payer connectors are reused; only the spec changes. Each stage is unlocked by the numbers from the one before.

NOW

Deductions

Find, dispute and collect short payments, then book the accounting behind them. Chargebacks, shortages and promo claims across retailers and distributors.

Gate: live now

Receivables

Cash application, reconciliation and collections for the same finance team. Same ledger, same payer connectors, new spec.

Gate: unlocks at 70%+ unattended on deductions

THEN

The back office

Vendor chargebacks, payables and close support. Each ERP connector, payer spec and trust record is reused for the next customer and the next workflow.

Gate: unlocks when workflow two repeats the pattern

What we sell: managed, tailored software at scale

Tailored

Workflows built around each customer's ERP, payers and reason codes. The deeper the integration, the stronger the lock-in.

Managed

We run it for the customer. A lean finance team has nothing new to learn or maintain, and no engineer sits in the loop.

At scale

AI makes the tailoring cheap. Code moves the money; the model only repairs and reads; the ledger verifies.

09 / NEXT STEPS

WANT THE FULL STORY?

Let's talk about
the back office.

The deck, the numbers as they come in, and a live look at the product are a call away.