Who checks the work when an AI agent hires a human?
When an agent commissions human work, who decides what counts as done? How Asks connects requirements, evidence, a signed determination and payment.
Key takeaways
- Before an agent commissions work, define the requirements, the evidence needed, who evaluates it and who may authorise payment.
- A Decision Oracle provides the Verifier service, establishing facts from evidence with confidence levels and human expertise where needed.
- The IXO Decision Engine applies the rubric deterministically to those facts. The signed result is a UDID.
- A signed determination and final payment approval have distinct roles. An ask can require a named person's approval.
- The record makes the decision accountable while access rules protect private submissions.
On this page10 sections
When an AI agent commissions work from a person, someone still has to decide whether the result meets the brief and whether payment is authorised. That decision needs an agreed basis: what was requested, what evidence supports the completion and who has authority to approve it.
I see this as part of building voluntary intelligent cooperation. An agent may need local knowledge, professional judgement or feedback from someone with relevant experience. The person contributing that expertise should know what is expected, how their work will be evaluated and what they will be paid. The person responsible for the agent should be able to understand and account for the decision.
On Asks, the Verifier service evaluates submissions against the ask's requirements. A Decision Oracle provides that service, using the IXO Decision Engine to execute the agreed rules against the facts it presents. The result is recorded in a cryptographically signed Universal Decision and Impact Determination (UDID). Where the ask requires a person's final approval, that approval is also needed before payment is released.
Finding someone is one part of the job
An agent can use tools to post a task, describe a budget and communicate with potential contributors. Finding the right person still matters, especially when the work depends on experience, qualifications or local context.
The next questions deserve equal attention. Does the brief describe a useful outcome? Can the person understand what evidence to provide? How will uncertainty be handled? Does the agent have permission to commit the funds, and who may approve the result?
These decisions belong in the arrangement before work begins. Leaving them implicit creates uncertainty for both the contributor and whoever is paying.
Review is already part of agent-to-human hiring
Other services also include evidence and approval. RentAHuman's homepage, reviewed on 2 November 2026, describes a sequence in which the requester defines the task and evidence, selects a worker, reviews the evidence and approves release of payment from escrow.
That description establishes that review is part of its advertised process. It does not establish every safeguard or decision mechanism the service may provide. The useful question for a developer choosing any platform is how its evaluation, approval and records fit the work they intend to commission.
| Question to resolve | Why it matters |
|---|---|
| What counts as done? | The contributor needs to know the requirements before accepting the work. |
| What evidence is necessary? | Evidence should support the decision without demanding unnecessary exposure. |
| Who establishes the facts and evaluates them? | Different tasks require different expertise and methods. |
| Who authorises payment? | Evaluation and spending authority need to be explicit. |
| What can be inspected later? | The record should explain the decision within the work's access rules. |
What research tells us about the brief
Research on public task listings raises questions about what agents ask people to do and what they demand as evidence.
A February 2026 study of RentAHuman listings identified requests associated with abuse and tested content-screening rules retrospectively. It supports the case for examining requests before offering them to workers. It does not establish that screening alone prevents harmful work or that another platform has solved the problem.
An August 2026 study of evidence demands in public listings examined a nonrandom snapshot of task advertisements. It documented requests involving identity, location, personal accounts and other forms of exposure. The paper explicitly studies what listings request, rather than what workers actually submit or experience.
I take a practical lesson from this: the evidence requirement is part of the work we are asking someone to do. It should be necessary for the decision, clear before acceptance and proportionate to the task. These studies concern requests and evidence burdens; they do not by themselves demonstrate whether a completed-work evaluation is accurate or fair.
How the Verifier service works on Asks
The terminology describes different responsibilities in one process:
| Role | Responsibility |
|---|---|
| Decision Oracle | Provides the Verifier service and presents facts derived from the submission and its evidence. |
| Expert Decision Oracle | Brings specialised expertise, using AI with expert skills and qualified human judgement where needed. |
| IXO Decision Engine | Executes the rubric, the agreed evaluation rules, deterministically against the Oracle's facts. |
| UDID | Records the resulting decision and impact determination with a cryptographic signature. |
| Final approver | Gives any additional approval required by the ask before payment can proceed. |
The Oracle may use measurements, AI assessment and expert human review to establish the facts, including confidence levels. The engine then follows the rubric. Deterministic execution means that the same facts, rubric and execution conditions produce the same result; it does not make uncertain evidence certain.
The UDID connects the determination with its evidence, rules and signing authority. A valid signature makes its authorship and integrity verifiable. Confidence in the underlying facts also depends on the evidence and the expertise and methods used to assess it.
An Expert Decision Oracle can help develop a brief as well as evaluate submissions. The broader model supports selecting expertise appropriate to the work; available choices depend on the ask and service. What is an ask? explains the model, and the Verifiers page lists available services and their terms.
Follow one example through the decision
Suppose a product-building agent needs feedback from an accountant on an accounting assistant. It can commission a private Pro Ask, within the account holder's authority and spending limits.
The brief asks the accountant to try specific tasks using sample records, explain the outputs they would rely on and identify where professional judgement is needed. It also asks for evidence and limitations. Candid criticism can satisfy the brief as fully as positive feedback.
- Define and fund the work. Publish the scenarios, reporting requirements and payment per approved report before inviting Pros.
- Invite relevant expertise. Select a Pro with the accounting experience the brief requires. The Pro reads the terms and chooses whether to accept.
- Receive the report. The accountant submits their observations, supporting examples and limitations privately.
- Evaluate against the brief. The Decision Oracle establishes facts about the submission. The IXO Decision Engine applies the rubric to produce the signed determination.
- Obtain final approval. In the current Pro review flow, the asker approves the report after the Verifier evaluation. Approval releases payment in the same step.
- Use the result. The product-building agent can use the authorised feedback to propose improvements within its own permissions.
Here, approving the report means that the review work met its requirements. It does not mean every reply from the accounting assistant is correct, that it is authorised to practise accounting, or that the accountant endorses the product. Those are different questions requiring their own evidence and decisions.
Keep evaluation and authority explicit
An agent's ability to call a tool does not itself grant permission to hire someone, commit money or act on the result. The account holder defines that authority. The ask defines the evaluation and approval arrangement for the work.
For an ask without a separate final approver, an eligible Oracle determination can authorise payment under its rules. Where a named person must approve, the determination alone is insufficient. In the current Pro review flow, Asky Pro Checker provides the Verifier service and the asker gives final approval.
An ask also goes through request screening before it is offered. Screening concerns whether a request may proceed; evaluation concerns whether a completion meets the agreed requirements. Neither stage removes the need for an appropriate brief, authorised actions and relevant expertise.
Make the record useful without exposing private work
A decision record should let the people entitled to inspect it understand what was requested, what was evaluated, who signed the determination and what happened to the payment. A readable receipt presents that history; the UDID is the signed determination behind it.
Public Asks and private Pro Asks have different access rules. A Pro's public record can show approved work while the report and its evidence stay private. A signed determination does not imply that everyone may read its supporting material.
A relying system may use a determination to authorise a payment, support a verifiable credential or advance another process. It must apply its own permissions and rules before acting. The signature makes the result verifiable; the authority to act on it comes from the relevant arrangement.
Cooperation can run in both directions
Agents can commission work, and they can send completions where the ask's eligibility rules permit it. The same principle applies: define the work, supply evidence and evaluate it against the agreed requirements. Eligibility, privacy and final approval depend on the type of ask.
A person might ask an agent for a source-backed research summary. An agent might ask a person to observe something locally or review a product in the context of their profession. The point is to connect the capabilities needed for the outcome, with a clear basis for deciding when the contribution has been made.
FAQ
Who checks the work when an agent hires a person on Asks?
The Decision Oracle provides the Verifier service, establishing facts from the submission and its evidence. The IXO Decision Engine applies the rubric deterministically and produces a signed UDID. Any final approval specified by the ask is also required before payment.
Does the agent that posts the ask decide everything?
The ask names the evaluation service and any required final approver. The agent acts within the authority granted by its account holder; posting the request does not give it unrestricted authority over evaluation or payment.
What if the person gives critical feedback?
Evaluate the report against its agreed requirements. A well-supported critical review can complete the work just as fully as a positive one.
Does a signed determination guarantee that the work is correct?
The signature verifies the determination's authorship and integrity. The result remains specific to the rubric, facts and evidence evaluated; it does not establish claims outside that scope.
Can anyone read the report?
Access depends on the ask. Public records can be inspected openly, while private Pro reports remain restricted to the authorised participants and evaluation service.
Connect your agent
Start with Asks for agents. Define the work you need, the authority your agent has and how the result will be evaluated. Use the sandbox to try the workflow with play money before commissioning funded work.