Encore

Inspiration

Live music is huge, and its margins are tiny.

Live Nation's concerts business generated $19 billion in revenue in 2024, with an adjusted operating income margin of just 2.8%. Coachella grossed $114.6 million in 2017 from about 250,000 attendees. Grand View Research estimates the festival segment of music tourism at $41.5 billion in 2025, growing 16.4% a year. And it's fragile. A record 78 UK festivals were postponed, cancelled, or closed in 2024, more than double the year before.

At margins that thin, small leaks matter. Benchmarks across industries show how common they are. APQC's median benchmark puts 1.5% of supplier disbursements as duplicate or erroneous. World Commerce & Contracting estimates organizations lose an average 8.6% of contract value after signing. The ACFE estimates organizations lose 5% of revenue to occupational fraud, with a median case costing $145,000 and running 12 months before detection. These are general business benchmarks, not leakage rates specific to festivals.

Festivals cram all of that risk into a few chaotic days. A contract says 15% commission while an email says 10%. A supplier bills for six generators when only four showed up. A cheaper quote demands a deposit that drains the cash cushion mid event. By the time anyone notices, the trucks are gone.

At an illustrative 2.8% operating margin, a permanent $7,500 overpayment would consume the operating income generated by about $268,000 in revenue. We built Encore so that unsupported payments don't go out in the first place.

What it does

Encore is an agentic finance team for music festivals. It is built to fill the seats of an accounts payable analyst, a vendor settlement analyst, and a nightly close accountant, while a finance controller keeps the keys.

AI reads the evidence. Code calculates the money. Humans approve consequential decisions.

The moment. A generator supplier invoices $22,500 for six generators, with an $8,000 deposit already paid. A workflow that only reads the invoice could pay the full $14,500 balance. Encore checks delivery first, sees only four generators were accepted, and allows $7,000. It holds $7,500 until the last two are delivered and accepted. Separately, it flags a possible $1,000 service level credit for human review. The held amount becomes payable once delivery is proven. The credit isn't savings unless it's approved and issued. Ask Encore why, and it shows the purchase order, delivery receipts, deposit, and cutoff it used.

It targets the biggest money leaks.

Paying for goods that never arrived. Supplier payments are capped at what delivery evidence supports.

Conflicting commission terms. A 15% vs 10% gap is $5,000 per $100,000 a vendor sells. Encore flags the conflict, and a signed contract outranks an email.

Forgotten credits. Every possible service level credit becomes a review item instead of slipping away. Encore never deducts one automatically.

Duplicate invoices. Exact duplicates are blocked, even with different capitalization or spacing. Bills from the same supplier with the same total within 7 days are flagged for review.

Cash crunches. For every competing quote, Encore shows the lowest projected balance and the first day you'd fall below your safety line. It shows the tradeoffs and leaves the choice to a person.

One wrong rule at scale. You can see every case a specific rule version decided, then pause it, fix it, and preview corrections before any money moves.

Oversized payouts. Vendor payouts above the configured automatic limit of $5,000 a day go to a person for approval.

Across the whole festival. Before the festival, Encore checks that every vendor is ready to transact. During it, Encore matches sales, payments, bank entries, and cash counts and calculates settlements. Every night, it runs the close and pauses for a person on anything unusual. After a decision, it can explain, challenge, and correct it without erasing what it knew at the time.

What a CFO can expect per festival cycle.

Money. $30,000 in payments prevented or recovered.

Time. 260 analyst hours freed, worth $13,000.

Return. $28,000 net on a $15,000 cost, a 187% ROI. Money alone is a 100% ROI.

This is an illustrative model, not measured results. It assumes $5 million in vendor and supplier spend, 2% of it at risk, Encore preventing or recovering 30% of that, and analyst time at $50 an hour. Freed hours aren't automatically payroll savings.

How we built it

Python API with permission checks on every action that moves money.

PostgreSQL for storage, with balances locked while payments are checked.

LangGraph runs the nightly close per vendor and saves checkpoints in Postgres, so a vendor paused on a hold resumes exactly where it stopped.

Claude extracts claims from contracts, emails, and invoices, always citing the source document. Claude never decides money.

Plain, predictable code handles reconciliation, settlements, invoice payability, and double entry accounting.

Separate workers handle the nightly close, scheduling, the demo clock, and payments. The API only records that a payment should happen. One outbox worker is the only component that can send money, through a simulated payment system. Idempotency controls stop retries from paying twice, and our tests cover the retry scenarios.

A flight recorder with an event log, lineage, snapshots, correction previews, and an audit log chained with hashes that can be checked for tampering at any time.

A web app built on a client generated from our OpenAPI spec. It never does money math. Every number on screen comes from the backend.

Challenges we ran into

Finding the real money bleeders. "Finance automation" means nothing on its own. We had to pin down exactly where festivals lose money, and every workflow in Encore exists to close one of those leaks.

Building workflows that actually stop the bleeding. Spotting a problem is easy. Holding back exactly the right amount, releasing it the moment proof arrives, and keeping routine payments moving during a live festival is not.

Understanding the blast radius. Automation saves time until a rule is wrong, and then it's wrong hundreds of times. We had to make sure every automatic decision points to the exact rule version behind it, that the record survives after the rule is switched off, and that every correction can be previewed before money moves.

Accomplishments that we're proud of

The generator invoice scenario works end to end. Real API, real database, real workflow, with exact dollar amounts checked by our verification scripts.

Only one component can send money. Payment intent is recorded by the API and dispatched only by the outbox worker, with idempotency controls against duplicate payment.

AI never touches the math. Claude reads and cites. Code calculates. People approve.

History is preserved and verifiable. Encore keeps both what it knew when it decided and the corrected state, and verifying the hash chain detects changes to recorded history.

No stale approvals. Case resolutions are tied to the case version and the evidence the reviewer saw. If either changed, the approval is rejected.

Proof, not promises. Verification scripts run against a real database and live API and check exact dollar outcomes across reconciliation, settlements, rules, nightly close, payouts, supplier bills, corrections, approvals, and the audit chain.

What we learned

The biggest leaks we found weren't fraud. They were payments made before the evidence was in. Our biggest wins came from asking "is this supported right now?" before every payment.

Missing isn't the same as bad. An unmentioned delivery fee isn't zero, and a new supplier isn't automatically unreliable. Treating those the same leads to wrong decisions.

Autonomy is only safe when you can see its blast radius. The question that shaped Encore was simple. If this rule was wrong, which payments did it touch?

AI is great at reading, and code should do the math. Keeping those jobs separate is what makes the system trustworthy enough to hand real seats to.

Testing the data isn't testing the system. Correct sample numbers don't prove the payment controls work, so we tested the controls directly.

What's next for Encore

Real payment rails in place of our simulated one.

More festival operations data flowing in automatically, instead of through imports.

Pilots with festival finance teams to measure what matters to a CFO. Faster close, less time on exceptions, and real dollars of leakage stopped or recovered.

Our demo uses synthetic data, so we don't yet claim measured customer savings. The pilot is how we'll prove the return.

Built With

• anthropic
• claude
• jwt
• langgraph
• openapi
• postgresql
• python
• uvicorn