Action-space constraints: a support desk where agents propose refunds but cannot issue them

Short answer

Action-space constraints are the limits on which actions an agent, human or software, may take in a task. In an ordinary business workflow they are role permissions and approval rules. A support desk is the clearest example: a front-line agent may look up an account and propose a refund, but the system blocks the agent from issuing it, and a billing lead must approve anything above a set amount. The records of those limits and handoffs are what make the constraint visible.

Visual briefFull-size diagram
A permission grid of three support roles against five actions; the tier 1 agent can view, propose and apply small credits but is locked out of issuing refunds, so its $1,150 refund proposal is routed to the billing lead, the only role that can approve refunds above $500.
Illustrative permission map at a fictional subscription software company. The locked cells are the action-space constraint.

What does action-space constraints mean in an ordinary business workflow?

An action space is the set of moves available to an agent in a task. A constraint removes some of those moves or makes them conditional. In a business, the constraint is a role permission or an approval rule, and the agent can be a person or software.

The term comes from reinforcement learning. In the Gymnasium library, a widely used toolkit for training agents, every environment declares an action_space that defines which actions are valid. Data companies now build business versions: Toloka describes environment mockups that replicate real tools' schemas, workflows and permission models. The research benchmark τ-bench gives agents a policy document for domains such as retail, and notes that some rules are enforced by the software while others depend on the agent applying them correctly. That second kind is where agents most often go wrong, which is why AI agents fail on tasks that look simple.

Payment software makes the constraint concrete. Stripe's documentation describes a tiered support setup in which tier 1 agents get a custom role with the refund permission removed, so they gather information while tier 2 takes actions that need more authority (custom roles are in private preview). Stripe's two-party approvals, in public preview, let an administrator require a second person to approve a refund, optionally above an amount, and Stripe ships default approval rules for refunds made by AI agent keys.

The permission map at one support desk (Illustrative)

The company is fictional: a 340-person subscription software firm. Its refund policy allows a full refund within 14 days of an annual renewal. Its support system and payment settings enforce this map:

ActionTier 1 agentTier 2 specialistBilling lead
View account, invoices and plan historyAllowedAllowedAllowed
Propose a refund with reason codeAllowedAllowedAllowed
Apply an account credit up to $100AllowedAllowedAllowed
Issue a refund up to $500BlockedAllowedAllowed
Approve a refund above $500BlockedBlockedAllowed, and must not be the requester

A customer renewed an annual plan for $1,150 nine days ago and asks for a refund. The tier 1 agent cannot issue it. The correct path is narrow:

StepActorActionRecord left
1Tier 1 agentConfirms renewal date and amount; checks the 14-day windowTicket note with invoice reference
2Tier 1 agentProposes a $1,150 refund, reason "within renewal window"Refund proposal linked to the ticket
3SystemRoutes to billing lead because the amount is above $500Approval request with timestamp
4Billing leadApprovesApproval record naming a different user than the requester
5SystemIssues refund; plan set to cancelRefund ID and plan status change

Two wrong paths are also informative. The agent could have offered a $100 credit to make the customer go away, which is inside its permissions but outside the policy. Or a tier 2 specialist could have split the refund into three payments under $500 to avoid the approval. Both leave records too, and a history that includes the corrections is more useful than one with only clean cases. This is also why synthetic data rarely replaces real records: invented tickets seldom contain the workarounds people actually try.

Which records show the constraint

Four kinds of record, each described at metadata level, make a constraint visible in a company's history:

  • The rule. The role and permission configuration, or the approval rule with its amount threshold, and when it changed.
  • The attempt. The ticket note or proposal where the agent reached the limit, with its reason code.
  • The handoff. The approval request, its timestamp and the approver, who must differ from the requester.
  • The outcome. The refund ID, credit or denial, and any later correction.

A company that can say which system holds each kind, and since when, has described its constraint without sharing any of it.

Does the company qualify, and what stays unknown?

Companies in the U.S., Canada and the EU qualify, as do companies in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia when at least 70% of the data is in English. The company also needs 50 or more full-time employees at peak (contractors excluded), documented operating history, rights to license the records, and an authorized contact.

Support records are often partly controlled by an outsourced support provider or limited by customer contracts, so the first rights question is who holds the tickets and the payment logs. Before you introduce anyone, confirm the owner is comfortable with the introduction; our guide on documenting client consent for a referral shows how. What remains unknown until SourceX reviews it: how many escalations carry complete role and approval records, how customer data will be handled, and whether any buyer wants this task type. Owners can start with questions to ask before licensing company data and whether a company can license its data to AI companies. More concepts sit in the data types hub, alongside customer support workflows.

Go deeper on action-space constraints

Next step

Refer a company. No cap on rewards. Default reward on a company's first qualifying transaction: the greater of $50,000 plus 5% and 15% of Actual Supplier Payout, with no cap; further qualifying transactions first agreed within 12 months of the first Closing pay 15%. Eligibility and payment conditions apply, including the USD 100,000 Minimum Transaction, and a signed Schedule A or addendum can differ. A lead, meeting or signature alone does not trigger payment.

If the support lead confirms role permissions enforced in software and logged approvals for escalated refunds, register as a partner, check the company against who qualifies, and read the terms before you introduce the decision-maker.

  1. Step 1Share your linkSend your personal link to a company you know.
  2. Step 2Company appliesThe company applies itself at /apply.
  3. Step 3Company signs and is paidThe company signs, delivers its data and is paid by SourceX under its agreement.
  4. Step 4You get your rewardthe greater of $50,000 plus 5% and 15% of the actual supplier payout, with no cap. A reward becomes payable after SourceX has paid the referred company under its signed agreement and the payout conditions in the program terms are met, including finance approval and your tax and payment details.

Common questions

Is an action-space constraint the same as a company policy?

Partly. A policy says what should happen; a constraint is a limit someone actually enforces, either in software permissions or by a required approval. The most useful records show both the rule and the enforcement.

Does a company need to use AI agents already for its records to be relevant?

No. A human support desk with role permissions and logged approvals records the same constraints an AI agent would face. What matters is that the limits and the escalations were written down by the system.

Can I ask the support lead for a screenshot of their permission settings?

No. Ask whether roles and approval rules exist, in which system, and since when. SourceX looks at configuration and records only under an agreement with the company.

Sources

  1. gymnasium.farama.org/api/env
  2. docs.stripe.com/get-started/account/teams/custom-roles
  3. docs.stripe.com/account/approvals
  4. arxiv.org/abs/2406.12045
  5. toloka.ai/virtual-environments-rl-gyms
  6. refer.sourcex.si/terms

Free resources

Know a company with valuable proprietary data?

Become a referral partner, get your link and introduce the company’s CEO (in an owner-run company, usually the owner). Referral partners can join from most countries. We can't work with partners in countries or regions under U.S. sanctions or on U.S. restricted-party lists. The companies you introduce should be based in the U.S., Canada and the EU, or in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia with at least 70% of their data in English.

Refer a company →

I own a business

Explore licensing your company's data directly to SourceX for AI training and evaluation. Start a short assessment; no uploads needed.

Start an assessment