What records demonstrate that an agent should ask for clarification before acting?
A complete clarification record has five linked parts. Missing any one of them weakens it.
- The instruction as received. The request, work order or ticket in its original wording.
- The trigger. What did not fit: a conflict with a contract, a requester without authority on file, a missing document.
- The question. Who asked whom, when, through which channel.
- The answer. Who replied, and what they decided.
- The consequence. What was done after the answer, and ideally a later record showing whether pausing was right.
Most companies keep parts 1 and 5 in their systems of record. Parts 2 to 4 are where clarification lives, and they are often scattered across email, ticket comments, approval tools and hold reason codes. Whether a company has them linked is the first thing worth finding out.
Why the moment of asking is hard to find in public data
Public text is full of completed tasks and short on the pause before them. Research on agents points at the same gap. A study built on an underspecified variant of a software engineering benchmark found that models tend to default to acting rather than asking, and struggle to tell underspecified tasks from well-specified ones; when they did interact, performance on underspecified tasks improved by up to 74%. The τ-bench benchmark makes the same point in customer service: each domain has a written policy, some rules are enforced by the system and some are not, so the agent has to apply the policy itself.
Security guidance names the risk on the other side. OWASP lists excessive agency among its top risks for large language model applications, describing it as damaging actions that follow from unexpected or ambiguous outputs, with excessive autonomy, permissions and functionality as its root causes. A business record of a person declining to act until authority was confirmed is a real-world example of the control OWASP describes.
None of this means every company's clarification records are licensable or wanted. It explains why the category is worth asking about. For more on why agents fail in practice, see why AI agents fail.
Three cases from one company (Illustrative)
All three come from a fictional 380-person commercial property management company. Each shows a different trigger.
| Trigger | What arrived | Question asked | Who answered | What happened | Records that show it |
|---|---|---|---|---|---|
| Conflicting scope | Work order to repaint a whole suite | Coordinator asks the owner's representative, because the approved estimate covers two rooms | Owner's representative | Scope confirmed at two rooms; change order raised for the rest | Work order, estimate, email thread, change order |
| Missing authority | Tenant staff member asks to add a new hire to the building access list | Front desk asks the tenant's named authorized contact | Tenant's office manager | Access granted next day after confirmation | Access request, authorized contact list, confirmation note |
| Incomplete evidence | Emergency plumbing invoice with no completion sign-off | Accounts payable places a hold and asks the property manager to confirm the work | Property manager | Paid after photos and sign-off; one duplicate line removed | Invoice, hold reason code, sign-off, adjusted invoice |
Look at the last column. Each case is only provable because several records point at each other. The access case, for example, needs the authorized contact list as it stood on that date; if the list is overwritten rather than versioned, the reason for the pause cannot be shown later.
The invoice case also shows why a refusal can carry value: the hold caught a duplicate line, and the adjusted invoice proves pausing was right. That later record is what turns "someone was careful" into evidence.
Which companies are likely to have these records?
Look for businesses where acting on a wrong instruction is costly and the rules are written down:
- Accounts payable and treasury teams with hold codes, two-person approvals and vendor-change callbacks.
- Property, facilities and field service companies where work orders, estimates and change orders are separate documents.
- Insurance, lending and benefits administrators with referral rules that send cases to a supervisor.
- IT service desks that verify requester authority before access changes.
- Logistics and procurement teams with tolerance rules on quantities, prices or delivery terms.
Firms in these areas often appear in back-office automation discussions, which is a natural moment to raise the question. For background on the category, see what agentic AI is.
Questions to settle before an introduction
Keep every question at metadata level. You never need to see a case.
- Do holds, referrals or escalations carry a reason code, or only free text?
- Are questions asked inside the ticket or case, or in separate email?
- Are authority lists and approval limits versioned, so the rule that applied on a given date can be shown?
- How many years of holds and escalations survive, and in which systems?
- Do the records belong to the company, or to its clients? A service provider working inside a client's system usually does not own that history.
- Who can authorize a licensing discussion? That is the CEO, usually the owner in an owner-run company.
The company also needs 50 or more full-time employees at peak (contractors excluded), a documented operating history, rights to license, and an authorized contact. 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. Details are on who qualifies.
Limits and unanswered questions
- Free text is hard to judge from outside. Whether questions and answers can be linked to the instruction reliably is only clear in SourceX's review with the owner.
- Personal and counterparty data. Clarification threads name people. De-identification and rights are settled before any delivery, and some threads may be excluded entirely.
- Not every pause was correct. Some holds are excess caution. A useful collection includes outcomes, so a buyer can tell good pauses from unnecessary ones.
- Synthetic substitutes. Invented cases can illustrate a policy but miss the real mix of ambiguity; our comparison of synthetic and real data for AI agents explains the trade-off.
- Two separate payments. Any company compensation is agreed between the company and SourceX. Your reward is funded separately by SourceX and never comes out of the company's payout.
More guides are in the guides hub.
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 a company tells you its holds and escalations carry reason codes and link back to the original request, that is the moment to register, check the basics on who qualifies, and read the terms.