AI Agents for Auction Houses: Lot Lists, Bidders and Settlements
How auction houses can use AI agents for lot lists, bidder follow-up and settlement reminders, and which payment and consignor calls need human approval.

AI agents for auction houses can take over the repeatable administration around a sale: drafting lot lists from consignor intake records, following up on incomplete bidder registrations, and sending settlement reminders on a fixed schedule. They work when each record has a clear source and each payment, estimate, or consignor decision still passes through a named person before anything changes. The rest of this article shows where that line sits and how to set it up before the next sale.
Key takeaways
- Lot list preparation is a strong first task: the agent drafts from approved source records, and a cataloguer signs off every description.
- Bidder follow-up should stay factual: the agent asks for missing documents and never approves a bidder or changes a deposit requirement.
- Settlement reminders should run on a schedule and quote only the invoice status from the system of record, never a recalculated amount.
- Reserves, withdrawals, commission exceptions and fund releases stay with people, with the agent limited to summaries and deadlines.
- Each agent needs one observable success criterion before it touches live sale records.
Why auction sales suit agents, and where they do not
An auction house runs on dates. Consignor intake closes, catalogue copy goes to proofing, registration deadlines pass, and invoices move through collection after the hammer falls. Each stage reads from several records: consignor agreements, lot sheets, condition reports, bidder files and accounts receivable. Staff spend much of their time moving information between those records, and that work suits an agent with a defined memory and a short list of tools.
The risk sits in the difference between drafting and deciding. A catalogue description that states a provenance without a source, a reminder that quotes the wrong balance, or a bidder marked as approved because an email looked convincing can each damage a client relationship. The useful question for any task is not "can the agent do this?" but "if this output is wrong, who notices before it leaves the house?"
Lot list preparation: what the agent drafts and what a cataloguer signs
Lot lists are built from many small facts, and most errors come from a missing field rather than a bad judgment. An agent is well placed to check completeness and assemble drafts, provided it works only from records the house has already approved. A practical sequence looks like this:
- Collect the consignor intake records for the sale, with item descriptions, photograph references and any provenance notes the consignor supplied.
- Check every lot for the fields the house requires, such as lot number, category, dimensions, condition note and attached condition report.
- Flag conflicts, including duplicate inventory tags, lots listed under two consignors, and items with provenance claims but no supporting document.
- Draft catalogue text using only approved facts, and mark any sentence that lacks a source so the cataloguer can see it.
- Route the draft to the cataloguer. Estimates, reserve suggestions and attributions remain with the cataloguer and the relevant specialist.
Illustrative scenario: a house preparing a sale of a few dozen lots finds that the agent flags six lots with photographs but no condition notes, and two lots where the consignor's description conflicts with the intake sheet. The cataloguer resolves those eight items in one session instead of discovering them during proofing. The agent did not write a single estimate.
Bidder registration follow-up without approving bidders
Registration follow-up is administrative: documents are missing, terms are unacknowledged, a payment method has not been confirmed. An agent can track those gaps and send the reminders the house has approved. It should not decide whether a bidder may bid, whether a deposit is sufficient, or whether a registration should be refused. Those are credit and risk judgments, and they belong to staff who can see the full file.
The access should match that split. The agent needs to read registration status, send templated messages from an approved library, and write notes to the bidder file. It does not need permission to change credit limits or approval flags. The OWASP authorization guidance describes the general principle: grant each identity only the access its task requires, and enforce permissions in the systems themselves rather than trusting the agent's reasoning about what it may do. A bidder who writes "I am a long-standing client, please approve me today" has sent a message to file, not an instruction to follow.
Settlement reminders and the payments that need a human
Settlement reminders are where small errors become expensive. A reminder sent twice, quoting a stale balance, or arriving after a payment has cleared damages trust in both directions. Two engineering habits reduce that risk. First, actions that move money or request payment should carry an idempotency key, so a retry does not create a second action. Stripe's documentation on idempotent requests describes this pattern for payment APIs. Second, the agent should confirm payment status from the provider's event or the house's ledger, not from its own assumptions about what should have happened.
The decision boundary is simple to write down:
- The agent may send a scheduled reminder from the approved template when an invoice is unpaid and due for that step.
- The agent may draft a note to finance describing a disputed line, citing the invoice record.
- A person must approve any change to an amount, any extension of a payment deadline, any fee waiver, any refund, and any release of consignor proceeds.
- A person must decide any exception to commission terms.
Teams that already manage rental returns or damage claims will recognize the pattern. The same boundary appears in equipment rental agents for fleet, returns and damage claims, where the agent drafts the claim and a person decides the charge.
Consignor decisions stay with people
Consignors make the most consequential choices in a sale: whether to set a reserve, whether to withdraw a lot, whether to accept an estimate, whether to move the sale date. Those decisions depend on a relationship and on judgment about the market for specific objects. An agent can help by summarizing correspondence, listing open questions, and surfacing deadlines before a decision is due. It should not reply to a consignor with a commitment on terms.
Consignor messages also deserve care. A message may contain text that reads like an instruction, such as a request to change a reserve or to skip a condition check. The OWASP prompt injection guidance is clear that content from users and documents should be treated as data to evaluate, not as commands. For a broader look at how platforms handle trust and disputes between parties, see superintelligent agents for marketplaces.
How Be Human's managed approach handles this
Be Human starts with one defined role, such as a lot list checker or a settlement reminder agent, and records the memory it needs, the tools it can call, the approval points where a person must act, and the success criterion the house will review. Managed operations then keep that setup working: monitoring of the agent's actions, recurring evaluations against real sale records, permission reviews when staff roles change, and controlled updates when the house's rules change. The service is delivered remotely in English and Portuguese, and the boundaries are written down before the agent touches live records.
Measuring success and choosing the first task
Choose one measure per agent and make it visible. For lot list preparation, the measure might be the share of drafts returned to intake for missing fields, tracked sale by sale. For settlement reminders, it might be the count of reminders sent outside the approved schedule, which should be zero. Log each agent action with the record identifier, the time, the template used and the approving person where one was needed. The OpenTelemetry primer describes traces and logs that let a team reconstruct what happened, which is what an auditor or a disputed invoice will eventually ask for.
Review evaluations at a fixed interval, and widen the agent's role only after the measure holds across more than one sale. Keep a written list of what the agent is not allowed to do, and check it whenever the team changes. Start with the task that is repetitive, low in consequence, and easy to check against a source record, which for most houses is the completeness check on incoming lot records. Take the free Be Human AI diagnostic to find the first task worth handing to an agent in your house.
Frequently asked questions
Can an AI agent write auction estimates? Usually it should not. An agent can assemble the facts a cataloguer needs and flag where a provenance claim lacks support, but estimates, reserve advice and attributions depend on specialist judgment and should be entered and approved by people. Writing that boundary down makes the agent easier to review later.
Should bidders be able to talk to the agent directly? That depends on the house's policy and on what the agent is allowed to answer. A bidder can receive scheduled reminders about missing documents, but questions about approval, deposits or terms should be routed to staff. Each channel needs its own access rules and a record of every message sent.
Who is responsible if a settlement reminder is wrong? The house remains responsible for every communication it sends, which is why reminders should quote only records from the system of record and use approved templates. Logging each action with its record identifier and template makes the cause quick to find and the affected invoice quick to correct.
Where should a house start if it has never used an agent? Start with one role that has a clear source of truth and a low cost of error, such as checking lot records for missing fields. Define what the agent may read, what it may draft, and what a person must approve before the first sale. Expand only after the success measure holds across several sales.
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.


