AI Agents for Auction Houses: Consignment Intake and Lot Records
An auction house agent can capture consignor details, photos, condition notes and draft lot records. Specialists sign valuations and provenance claims.

An AI agent for an auction house can take over the repetitive work around consignment intake: collecting consignor details, sorting item photographs, recording condition statements and drafting lot records for review. It should not set estimates, reserves or authenticity opinions, and it should not certify provenance. The practical question is which fields the agent fills, which fields a named specialist must approve, and what evidence sits behind each one.
Key takeaways
- Consignment intake is a sound first workflow for an agent because its inputs are repeated, structured and easy to check against a written record.
- An agent can draft lot records, flag missing documents and keep the consignor thread current, but valuations, reserves and attribution stay with specialists.
- Every field in a lot record should carry its source: the consignor's statement, a photograph, a document or a specialist's note.
- Approval boundaries should be written down before the agent touches live consignments, and they should be reviewed as the workflow changes.
- Success is measurable: complete lot records that reach cataloging without rework, with every signed claim traceable to a named person.
What an agent can capture at consignment intake
Consignment intake usually starts with a conversation, an email or a walk-in visit, followed by a consignment agreement, a list of objects and a round of questions about history, condition and documentation. Much of that information is scattered across messages, attachments and handwritten notes. An agent can read these inputs, pull the relevant details into one structured intake record and point out what is missing before a specialist spends time on it.
The useful scope is narrow. The agent can extract the consignor's name and contact preferences, the item description as the consignor states it, dimensions, materials and visible inscriptions if they appear in the text, the photographs received and whether each object has a complete set of images. It can also check the intake against the house's own style guide and past lot records, so that new descriptions follow the same vocabulary. Retrieval from the house's own documents is the pattern behind most of this work, as described in the AWS overview of retrieval-augmented generation: the model answers from material the house supplies rather than from memory.
What the agent should not do at this stage is interpret. A consignor's claim that a clock was made by a particular workshop is a claim, not a fact, and the record should label it that way.
Which lot record fields come from which evidence
A lot record is only as trustworthy as its weakest field. Before deploying an agent, map each field to the evidence that supports it. A simple structure works well:
- Consignor identity and agreement reference: taken from the signed consignment form, with the date the form was received.
- Object description: drawn from the consignor's list and the specialist's first examination, with the source of each sentence visible.
- Condition notes: recorded by the specialist from direct inspection, with photographs attached. The agent may summarize them but should not add conditions the inspector did not report.
- Stated history and provenance: recorded as the consignor's statement, with each supporting document listed by name.
- Estimate, reserve and attribution: left blank by the agent, completed only by the responsible specialist.
This separation matters because a clerk reading the record later needs to know which sentences describe what the house observed and which sentences describe what the consignor said. The same discipline appears in quality management, where nonconformance records must show who observed what. The approach in the article on AI for Quality Management: Nonconformances and Audit Prep applies here: a record is useful when its evidence can be audited after the fact.
Which valuations and provenance claims a specialist must sign
Valuations and provenance statements carry financial and legal weight. An estimate affects how a lot is presented, how a consignor is advised and how buyers decide. A provenance statement can be relied on for title, authenticity and reputation. Neither should be produced by an agent acting alone, regardless of how capable the underlying research tools become.
The agent can still help a specialist work faster. It can assemble comparable sales from the house's archive, summarize prior lots by the same maker, list the documents a consignor has offered and note where the chain of ownership has gaps. An export or import file for a cross-border lot may also need its own paperwork tracked, a task similar to the one described in AI Agents for Import Export Companies: Documents and Status. The output is a research packet, not a conclusion.
The specialist signs three things: the estimate and reserve recommendation, the attribution or authenticity opinion and the provenance statement that will be published or given to a buyer. Each signature should name the person, the date and the evidence reviewed. The NIST AI Risk Management Framework is useful here because it treats human oversight and accountability as part of managing AI risk, not as a formality added at the end. As research tools mature, the agent may draft more of the packet. The signature line does not move.
A first-workflow plan and an illustrative scenario
A first deployment should be small enough to measure within a few cycles of intake. A practical sequence looks like this:
- Choose one consignment category, such as small decorative objects, and exclude high-value or legally complicated lots from the first phase.
- Write the lot record template with the source label for every field, and agree it with the specialists who will sign.
- Give the agent read access to the style guide and a sample of past lot records, and write access only to draft intake records.
- Define the success criterion before launch: for example, the share of intake records that a cataloger accepts without changing any sourced field.
- Log every agent action so that a specialist can see which document a sentence came from.
- Review errors weekly during the first period, and widen the category only after the criterion holds.
Illustrative scenario: an executor sends a box of silver flatware, a painted tray and a letter that mentions a previous owner. The agent creates an intake record, lists the photographs received, notes that the letter is not yet matched to an object, asks the executor for a clearer image of the maker's mark and flags that the tray has no condition report. A specialist inspects the items, signs the condition notes and decides whether the letter supports the provenance. The agent never writes an estimate, and it never sends the executor a value.
How Be Human runs this as a managed service
Be Human designs and operates AI agents that keep their identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role for the house, the memory and tools that role needs, the approval boundaries and an observable success criterion, such as accepted intake records or the number of specialist corrections per batch. Managed operations can then include monitoring, recurring evaluations of the agent's outputs against sampled lots, maintenance of the connections to the house's systems, periodic permission review, incident response when a record is wrong and controlled changes to the workflow. Service is remote and available in English and Portuguese.
Controls and limits for consignor data
Consignment files contain personal data, financial terms and documents that may be sensitive. Access should follow the principle that each user and each agent receives only the permissions its task requires. The OWASP guidance on authorization makes the same point for any system that acts on behalf of people. An intake agent does not need access to buyer payment records, and a permission review should confirm that it has none.
Consignor emails and attached documents are also a security surface. A message may contain text that tries to instruct the agent, for example to change a reserve or to disclose another consignor's details. The OWASP prompt injection cheat sheet describes why model inputs should be treated as data, not as commands. In practice, the agent should have no path to publish, send or alter a signed record, so a malicious instruction has nothing to act on.
Finally, the house should be able to see what the agent did. OpenTelemetry's observability primer describes the value of traces, logs and metrics for understanding system behavior, and the same idea applies to an intake agent: each draft should record which inputs it used and which tool calls it made. Where lots involve disputes between consignors and buyers, the handling rules described in Superintelligent Agents for Marketplaces: Trust, Disputes, Sellers offer a useful parallel for keeping the record defensible.
Where to start
The best first task is usually the one with clear inputs, a known checklist and a specialist who is already reviewing the output by hand. For most houses, that is the intake record for a single consignment category, not the whole pipeline. Pick the category, write down the sourced fields and decide who signs each claim before the agent is involved. If you want help identifying the first intake task worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Can an AI agent set the estimate or reserve for a lot? No, and a sound design keeps that out of the agent's scope entirely. The agent can assemble comparable sales and summarize the archive, which helps the specialist reach a view faster. The estimate and reserve recommendation remain the signed responsibility of a named specialist.
Who is accountable when a lot record contains an error? The house remains accountable, and the record should show who reviewed each sourced field. A clear log lets the team trace whether the error came from the consignor's statement, a document, the agent's draft or a specialist's correction. That traceability is what allows the workflow to be fixed rather than guessed at.
Does the agent need access to the full client database? It does not. Intake work needs the consignment file, the style guide and a sample of past lot records, and access should be limited to those sources. Buyer payment and account data are outside the task and should stay outside the agent's permissions.
How do we know the agent is ready for live consignments? Readiness is shown by results, not by a demonstration. Set the success criterion in advance, sample the agent's drafts on a schedule, record every correction a specialist makes and widen the scope only when the corrections stay low and the sourcing holds. Recurring evaluations are the way to keep that judgment current as the workflow changes.
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.


