Skip to content
be-human.ai
← Back to the journal

AI by industry

Auction House AI: Consignor Files, Lot Sheets and Payout Sign-Offs

How an auction house can use AI for consignor records, lot lists and settlement drafts, with the human sign-offs that protect every payout.

Be Human typographic cover with the title: Auction House AI: Consignor Files, Lot Sheets and Payout Sign-Offs
Conceptual illustration created with AI · Be Human

An AI system for an auction house can keep consignor records current, turn intake notes into clean lot lists, and draft settlement statements for staff to check. The agents prepare and flag. People approve every payout, every change to a consignor's terms, and every estimate or reserve that reaches a catalog. Set up this way, the paperwork moves faster while the decisions that move money stay with the people accountable for them.

Key takeaways

  • Start with consignor records, because lot lists and settlements inherit every error in them.
  • Let agents draft lot sheets and settlement statements, but keep estimates, reserves and payout release with named staff.
  • Give each agent read access to the records it needs and write access only to drafts and flags.
  • Require a second person to confirm any settlement that changes a consignor's net amount or payment details.
  • Judge the work by fewer corrections after release, not by how many documents the agent produced.

Where AI fits in the consignment workflow

An auction house runs a chain of handoffs: consignment intake, valuation, cataloging, the sale itself, buyer invoicing, and finally settlement to consignors. Each handoff copies information from one record into another, and most errors appear at the copying step. A consignor's payment details change in one system but not another. A lot described as restored in the intake form is listed as original in the catalog. A returned lot is paid out because nobody linked the return to the settlement.

AI is useful at these seams. A consignor records agent can keep one current profile per consignor with the agreement version, authorized signers, commission schedule and open items. A cataloging drafter can turn intake photos and notes into structured lot descriptions. A settlement preparer can assemble the statement from the sale results and the consignor's terms. What none of these agents should do is accept a consignment agreement, set a reserve, or release funds. Those are judgment and authority decisions, and they belong to named people.

How do you keep consignor records accurate?

A consignor record is the foundation for everything downstream, so it needs a clear scope. It should hold the legal name, the consignor's contact preferences, the signed agreement and its version date, the commission and expense terms, the authorized signers, the payment method (stored masked), and a short history of sales and disputes. Anything the agent states should carry a source and a date, so staff can see where a fact came from.

The main risk is stale or duplicated data. Two records for the same consignor, or an old bank account still attached to a profile, can send money to the wrong place. A practical rule is that any change to payment details requires verification through a channel the consignor did not supply in the same message, such as a call to a number already on file. The agent can detect that a change was requested and open a task, but it should not complete the change itself.

When an agent needs to answer questions from a long consignor history, retrieve the relevant entries rather than loading everything into its context. Retrieval-based designs, described in the Google Cloud and AWS architecture guidance, keep answers tied to the records they came from, which makes them easier to audit.

Preparing lot lists from intake

Lot lists are where errors become public. A wrong dimension, a missing provenance note or a condition that was never checked can lead to a dispute after the sale. The goal of the agent is to produce a complete draft and an honest list of open questions, not a finished catalog entry.

A workable sequence looks like this:

  • Confirm that the consignor record and the signed agreement for this sale are current.
  • Extract the fields for each item from intake photos and notes: title, category, medium, dimensions, condition, and provenance.
  • Mark any missing or contradictory field as open, rather than filling it with a plausible guess.
  • Match each lot to its consignor and to the sale it is entering, and flag lots with a restriction or a pending return.
  • Produce a draft lot sheet and a separate list of open questions for the cataloging specialist.
  • Log what the agent changed, which sources it used, and who approved the final version.

Intake forms and uploaded documents are data, not instructions. A consignor's note that says "waive the commission on this lot" should be treated as a request for staff, not as a command the agent follows. The OWASP guidance on prompt injection prevention describes this separation, and it matters more in a business where outside parties write into the same fields the agent reads.

Consignor notes describe the item. They do not change the agreement.

Drafting settlements without releasing money

Settlement drafts are the most sensitive output in the workflow, because a consignor reads them as a promise. A good draft shows the hammer result for each lot, the buyer's premium that is not passed to the consignor, the commission, the agreed expenses such as photography or insurance, any holdback for an open dispute, and the net amount. Every line should point to a source document.

Illustrative scenario: a consignor's lot sells, but the buyer later returns it for a condition issue. The settlement preparer notices the return in the buyer ledger, places a holdback on that lot's proceeds, and drafts the statement with the holdback shown as a separate line marked pending. It flags the consignor record for a staff decision on whether the return is valid. The statement is not released until the finance lead clears the holdback. The agent did the reconciliation and the staff member made the decision.

Payout instructions need one more safeguard. A retried request must never pay twice. Payment APIs such as Stripe describe idempotency keys for this purpose: each instruction carries a unique reference, so repeating the request returns the original result instead of creating a second payment. The settlement system should generate that reference once per statement and keep it in the audit trail.

The sign-offs that protect payouts

The sign-offs are what make the workflow safe to run. They should be written down, assigned to roles, and checked by the system as well as by people. A practical checklist for a settlement cycle:

  • Consignor identity and payment details are verified through a known channel before the first payout and after any change.
  • Statement totals reconcile to buyer invoices and to the recorded hammer results for every lot.
  • Every deduction has its own line and a source document.
  • Any lot with an open dispute, a returned item or an unpaid buyer is held, not released.
  • A second staff member approves every statement that changes a consignor's net amount beyond routine lines.
  • Each payout instruction is issued once, with a unique reference.
  • Agents can draft and flag. Only named finance staff can release funds.

The last point is an access decision, not a policy statement. The OWASP authorization guidance recommends deny by default and least privilege, which in practice means the settlement agent has read access to sale results and consignor terms, write access to a drafts folder, and no permission to call the payment system. The NIST AI Risk Management Framework is a useful reference for documenting who is accountable for each of these decisions and how the house will review them over time.

Measuring quality and keeping controls current

A workflow that drafts statements needs an evaluation set. Pick a group of past sales where the correct statements are known, run the agent against them, and compare its output line by line. Repeat the test whenever the model, the prompt or the connected systems change. Evaluation guidance from OpenAI and Anthropic describes this pattern in general terms, and the same discipline applies to an auction house.

The useful metrics are operational: corrections found after a statement is released, open questions that staff resolved, time from intake to draft lot sheet, and exceptions raised per sale. Log each draft with its sources and approver so that, when a consignor asks why a figure changed, the answer is in the record. OpenTelemetry's concepts for tracing and observability give a vocabulary for this kind of logging, even if the house uses a simpler tool.

Permissions also need periodic review. Staff roles change, integrations get reconnected, and access granted for a one-off project tends to stay. A quarterly check of who and what can read, write or release is a small task that prevents most of the quiet failures.

How Be Human runs this as a managed service

Be Human designs these agents as part of a managed service. A first engagement defines one useful role, usually the settlement preparer or the lot sheet drafter, along with the memory it needs, such as the consignor profiles and signed agreement versions, the tools it can call, such as read-only access to the sale and ledger records and write access to a drafts folder, the approval boundaries that keep release with finance, and one observable success criterion, such as corrections found after release over a defined period. Ongoing operations cover monitoring, recurring evaluations on past sales, maintenance of the integrations with the auction and accounting systems, permission reviews, and incident response when a source system changes or an output is wrong. Engagements run remotely, in English or Portuguese, and the boundaries are agreed before any agent touches a payout workflow. If you want to find which task in your auction house is the right first one to hand to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent set reserves or estimates for auction lots? It should not. Reserves and estimates depend on market judgment, consignor expectations and house policy, and they carry authority that should stay with named specialists. An agent can gather comparable sale records, summarize condition notes and flag where a valuation is missing, so the specialist decides with better information.

Should an AI agent be allowed to release consignor payouts? Release should stay with finance staff. The agent can reconcile the statement, draft the payout instruction and flag anything unusual, but the final release is a decision with financial and legal consequences. Give the agent no permission to call the payment system, and issue each instruction with a unique reference so a retry cannot pay twice.

How much consignor data should an agent be able to see? Only what its task requires. A settlement preparer needs the agreement terms, the commission schedule and the sale results, but it does not need full identity documents or unrelated correspondence. Apply least privilege, mask payment details, and review access on a regular schedule.

What does a first engagement with Be Human look like? It starts by choosing one role with a clear boundary, usually consignor records, lot list drafting or settlement preparation, and defining its memory, tools, approvals and success measure. The first version runs against past sales so the team can compare its output with known results. Only after that review does the agent move closer to live work, and it stays under monitoring and permission review afterward.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

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

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →