An industry agent is not a prompt with industry vocabulary
By Equipo Quantum Developers
Summarize:
An AI agent does not become industry-specific because it speaks the language of logistics, healthcare, or finance. The falsifiable test is stricter: it is ready for a domain only when it can declare what data it accepts, which decisions it may make, which actions are forbidden, and who owns its exceptions. If two agents share the same operating contract and differ only in tone and vocabulary, they are not two domain capabilities. They are two interfaces over the same risk.
That difference is easy to miss in a demonstration. A model can answer persuasive questions from hand-picked documents, while production receives late data, conflicting states, and requests outside the intended scope. The failure is rarely that the model lacks more industry terminology. It is that the organization has not converted domain knowledge into a testable contract.
Put the contract before the persona
A minimum data contract is more than a list of files the agent may read. It identifies the business object, its key, the authoritative source, the time observed, acceptable freshness, required fields, rejection rules, and usage rights. It must also distinguish a missing value from zero and a corrected record from a new observation.
NIST assigns AI design actors responsibility for documenting objectives, assumptions, context, requirements, metadata, and dataset characteristics. Its actor model explicitly includes domain experts, data providers, product owners, and governance specialists (NIST AI RMF actor tasks). The operational consequence is important: the modeling team cannot write the contract alone. Operations defines meaning and exceptions; data teams define lineage and quality; security and legal define rights; the process owner accepts the consequences of a decision.
For high-risk uses within its scope, the EU AI Act requires data-governance and human-oversight measures and describes operators who have the competence, training, and authority to intervene (Regulation (EU) 2024/1689). Not every business agent falls into that legal category. The design lesson still travels well: a reviewer who cannot stop an action is decoration, not oversight.
Draw a decision boundary with verbs
A useful boundary separates four kinds of authority:
| Level | The agent may | Required condition |
|---|---|---|
| Observe | read, summarize, detect inconsistency | no business state changes |
| Propose | recommend a category or next action | a person or independent rule decides |
| Execute reversible | open a task, request evidence, save a draft | idempotency and tested undo exist |
| Execute irreversible | approve, pay, cancel, commit inventory | separate authorization and stronger evidence |
The contract also needs explicit prohibitions: do not fill an absent value as fact; do not use an expired source; do not change a beneficiary; do not publish without approval. The NIST AI RMF calls for documented scope, differentiated human-AI responsibilities, and defined oversight processes (NIST AI RMF Core). In operating terms, every action needs an owner, a policy, and a record explaining why it was allowed.
A domain-readiness scorecard
The scorecard below is an article-specific heuristic, not an external benchmark. Score each dimension from zero to three: zero means absent; one means implicit or manual; two means defined and tested on a sample; three means instrumented and reviewed periodically.
- Data contract: identity, source, schema, freshness, and rejection behavior.
- Stable semantics: states and terms mean the same thing across teams.
- Decision boundary: permitted, prohibited, and approval-gated actions.
- Reversibility: compensation or a hard block exists for material actions.
- Exception ownership: a queue has an owner, a clock, and an escalation route.
- Evidence: input, policy version, decision, approver, and outcome remain linked.
- Rights and security: purpose and access for each source are authorized.
- Review capacity: qualified people have enough context and time to decide.
A practical rule is that any zero blocks autonomy. A set of ones may support a read-only assistant. Mostly twos can support a pilot with approval. Threes do not automatically grant autonomy, but they show that the organization can observe and govern behavior. This heuristic never replaces legal, privacy, or security assessment.
Illustrative example: industrial spare parts
Consider an agent that classifies spare-part requests. This is an illustrative design, not a customer result. Inputs are an asset number, free-text description, criticality, and location. The contract authorizes the asset master and maintenance history, but forbids the agent from inferring compatibility when the asset model is missing.
A valid asset with one exact match may produce a reservation draft. Two plausible matches create an exception with reason AMBIGUOUS_PART. An unknown asset produces ASSET_NOT_FOUND, never an invented component. Maintenance makes the technical choice, procurement checks availability, and the agent preserves the evidence it used. The design is genuinely domain-specific because it embeds identity, authority, state, and responsibility—not because the prompt contains machine names.
What should be visible in operations
A control plane such as Quantum Automation Center can expose an agent catalog, execution status, timelines, artifacts, logs, and operational analytics. Those surfaces do not replace the domain contract; they make it inspectable. For each agent, an operator should be able to answer:
- Which contract and policy version ran?
- Which business object was affected?
- Which source was considered current?
- What action was proposed or performed?
- Who approved it, and where is the evidence?
- Which exception remains open, and who owns it?
The documentation for AI agents and business objects can help align those fields with the surrounding architecture. The goal is not to pull every business rule into one product. It is to create a common view of limits, runs, and outcomes.
The counterargument: discovery should not become bureaucracy
An industry-informed prompt can be entirely appropriate for document exploration, question preparation, or summaries that a person will review. Requiring eight fully mature dimensions before testing whether a problem deserves investment can create months of design without evidence of usefulness. For informational, reversible work, begin with a light contract: approved sources, a visible uncertainty policy, and no authority to act.
Controls should grow with consequence. The mistake is not using a prompt to learn. It is presenting that prototype as an operating capability before deciding who answers when data is late or a recommendation is wrong.
When not to use a domain agent
Do not use one when a stable rule can be implemented more predictably with deterministic automation. Stop as well when there is no authoritative source, business objects lack consistent identity, nobody accepts the exception queue, or human review exists only on paper. If an irreversible decision must occur faster than qualified review is available, reduce scope: observe or propose, but do not execute.
The final test is simple. Remove the industry name from the prompt. If you still cannot distinguish the agent by its data contract, decision boundary, and responsibility model, you do not yet have an industry agent. You have a themed demonstration.
Sources
Article topics
