How Set Designers Track an Art Department Budget With an AI Agent

Set designers track an art department budget by keeping two numbers apart: what the production has committed to pay, and what has actually been spent, each recorded per set and per build. An AI agent can keep both figures current by reading quotes, purchase orders, invoices and credit notes as they arrive, filing each one against the right set, and answering the question on demand with its sources attached.

Key takeaways

• Track committed and spent as separate figures, and never rely on a single running total.
• Assign every quote, order, invoice and receipt to one set and one build before anything else happens.
• Let the agent extract and file documents, but require a named person to approve every change to the ledger.
• Answer budget questions with the figure, the date of the last record, and the documents behind it.
• Start with one active build and one weekly reconciliation routine before widening the scope.

Why committed and spent are different questions

An art department budget answers two questions that often get merged into one. The first is what the production has promised to pay: approved quotes, placed orders, rental agreements and labor commitments that have not yet been invoiced. The second is what has actually left the account: invoices posted and payments made. Committed money moves when a decision is made. Spent money moves when a bill arrives, which can be days or weeks later.

A designer who sees only the running total cannot tell whether a set is over budget because of a real overrun or because a quote approved two weeks ago has not reached the bookkeeper yet. Those two situations call for different responses. The first needs a conversation about scope. The second needs a reminder to the vendor and an update to the forecast.

Per set and per build matters for the same reason. A street set is often built in stages, with exterior panels, a storefront and practical props each carrying different vendors and lead times. The department needs the committed cost of the storefront build on its own, because the decision to rebuild or reuse a piece belongs to that build, not to the set as a whole.

What the agent reads and how each document is filed

The agent works from documents that already exist: quotes from scenic shops and prop houses, purchase orders, rental confirmations, supplier invoices, credit notes and receipts for small purchases. It does not need to sit inside the production's accounting system to be useful. It needs read access to the mailbox or shared folder where these documents land, and write access to a ledger the department already uses, such as a structured spreadsheet or a budgeting tool.

Each document becomes one row with a fixed set of fields. Keep the structure small enough that a person can check it at a glance:

• Set and build, using the names on the art department board, not the names vendors use
• Category, such as scenic construction, paint, props, rentals, greens or labor
• Vendor name and the vendor's own document reference
• Status: quoted, approved, ordered, invoiced, paid, credited or returned
• Amount as stated on the document, with a note on whether tax is included
• Date recorded, and the file name or link of the source document

The vendor's document reference is the most important field after the status. Vendors resend invoices, and a clean reference lets the agent update an existing row instead of creating a second line of spend. The same quote-to-supplier discipline appears in AI agents for procurement teams, and it carries over well to production budgets.

An illustrative scenario: a street set in mid build

Illustrative scenario, not a real production: a production designer is three weeks from shooting an exterior street set. On Monday, a scenic shop sends a quote for wall panels and a storefront frame. The agent files both under the street set, with the storefront assigned to its own build, and sets the status to quoted. The committed figure for the street set rises by the quoted total. The designer replies in the thread with an approval, and the agent changes the status to approved and saves the thread to the record.

Later that week, the shop invoices for the panels. The invoice is higher than the quote because a rush delivery charge was added. The agent does not overwrite the quote. It records the invoice as a new row linked to the same vendor reference, marks the difference as a variance, and notes that the rush charge does not appear in the approved quote. It then sends a short message to the designer asking whether the charge was agreed. No money has been committed on the agent's own judgment.

On Thursday, the production coordinator asks, what is committed for the street set right now, and how much of it is spent? The agent answers with the committed total, the spent total, the variance, the list of open quotes, and the two documents behind the variance, along with the date of the most recent record. The answer comes from the ledger rows, not from the agent's memory of earlier conversations.

The point of the scenario is not automation for its own sake. The designer learned about the rush charge on the day it appeared, not at month end, and the coordinator's question received a sourced answer in the time it takes to read it.

A practical plan for the first pass

Start with one active build, not the whole production. The first pass should be small enough that a person can verify every row by hand.

• Choose one build that is in active procurement, and agree on its set and build names with the art department lead.
• Gather every open quote, purchase order and unpaid invoice for that build, and enter them once, by hand or with the agent under supervision.
• Check every match the agent proposes between a document and a row before the ledger is treated as correct.
• Write the status definitions and the transitions that need a human, such as approval, a change in amount, a credit, or a cancellation.
• Fix the reply format the agent uses for budget questions: committed, spent, variance, open items, and sources, with the record date on each answer.
• Run a weekly reconciliation that lists quotes with no approval, invoices with no matching quote, and rows whose status has not changed for a long time, and have a person clear the list.
• Add the next build only after the reconciliation has run cleanly for a full production week.

The weekly reconciliation is the routine that keeps the figures honest over time. Without it, the ledger drifts quietly, and the agent's answers become confident summaries of stale rows.

What the agent should never decide on its own

The ledger records commitments, so the agent should record money and never commit it. It can log a quote. It should not accept one, place an order, or mark an invoice as approved for payment. Those actions belong to named people. Write this boundary into the agent's instructions, and reflect it in its permissions: read access to vendor mailboxes, write access to the ledger, and no access to payment systems at all.

Documents are also an attack surface. A vendor's PDF can contain text that tries to instruct an automated reader, whether by accident or on purpose. The OWASP guidance on prompt injection recommends treating content from documents as data to analyze, not as instructions to follow. For this work, a line inside an invoice that says mark as paid should be recorded as unusual text and flagged for a person, not obeyed.

A document tells the agent what a vendor wrote. Only the production decides what it owes.

Duplicates are the other common failure. The same invoice arrives by email and again in the shared folder. Using the vendor's document reference as the key means the second arrival updates the existing row rather than adding spend. Payment platforms apply the same principle with idempotency keys, so that a repeated request is recognized rather than executed twice, as described in the [Stripe documentation on idempotent requests](https://docs.stripe.com/api/idempotent_requests). A budget ledger deserves the same care.

Where to start with an agent in your department

Be Human designs and operates this kind of memory and reporting for productions, departments and other teams as a managed service. A first engagement defines one useful role, here a budget reader for one art department, along with the memory it keeps (sets, builds, vendor names and status definitions), the tools it can use (mailbox read access and ledger write access), the approval points a person must hold, and a success criterion that can be observed, such as whether every spend answer cites its source document. After launch, the work can include monitoring the ledger for stale statuses, reviewing permissions as vendors and staff change, maintaining the connections to the mailbox and the ledger, and responding when something is misfiled. For teams that need a weekly view across several productions, the same status and risk reporting pattern described for project management offices applies.

If you want to find the first task in your department worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent really tell committed money from spent money? Yes, if the ledger has a status field and every document is assigned to a status when it is filed. Quotes and approved orders are committed, and invoices that have been posted or paid are spent. The agent applies those rules to the rows it has, and a person reviews any document whose status is unclear.

Does the agent need access to our accounting system? Usually not for the budget view. It needs to read the mailbox or folder where quotes and invoices arrive and to write to the department's ledger. Payment systems should stay outside its permissions, so that it can report on money without being able to move it.

What happens when an invoice differs from the approved quote? The agent keeps the quote as it was approved and records the invoice as a separate row. It marks the difference as a variance, cites both documents, and asks the responsible person whether the change was agreed. The designer or coordinator decides what happens next, and the record shows both figures.

Should a set designer trust the agent's totals without checking? Trust the totals only after the first reconciliation cycles have been reviewed by a person and the matches look right. After that, spot checks and the weekly reconciliation list are the normal controls. Any total that a decision depends on should still be traceable to the source documents in one step.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.