We run the procurement platform
Sticker has powered purchasing workflows for childcare centers, restaurants, direct primary care clinics, cleaning companies, summer schools, and other operating businesses.
Proprietary data sample · RL environmentsSUPPLY BENCH V3 · MULTI-VENDOR PROCUREMENT
Supply Bench
Can an AI agent turn an approved purchase request into correct supplier orders?
Sticker evaluates computer-use agents on the operational work between approval and a reconciled purchase: sourcing products, placing orders across vendors, preserving buyer policy, and proving what happened.
16.7%strict end-to-end pass
0.63mean weighted reward
38.9%reward at or above 0.95
18 fixed-cohort attempts · 3 deterministic tasks · exact model IDs · full trajectories
THE ACTUAL ENVIRONMENT
One request. Multiple systems. A verifiable final ledger.
These are frames from a successful agent trajectory, not design mockups.
Read the approved request
The agent starts in Sticker with PO lines, budget, deadline, and buyer policy.
Place supplier orders
It navigates realistic supplier catalogs, reasons over offers, and checks out.
Reconcile what happened
It maps order lines, handles discrepancies, and commits a final disposition.
PROCUREMENT, IN PLAIN ENGLISH
Approval is the beginning of the buying work.
Procurement is how an organization decides what to buy, from whom, under which rules, and then verifies what was actually ordered, delivered, and invoiced.
In the workflows represented here, an approved purchase request authorizes the spend but does not complete the purchase. A human purchasing operator still has to translate each requested line into supplier orders while respecting exact specifications, pack quantities, budgets, delivery dates, substitutions, and account policy.
- 1Understand the requestWhat is needed, by when, and under which constraints?
- 2Source and orderChoose defensible offers and transact in the correct vendor portals.
- 3Map the ledgerConnect every supplier-order line back to the approved PO line.
- 4Recover and reconcileHandle backorders, price variance, replacements, and final disposition.
WHY STICKER
Built from operating experience, not imagined workflows.
Our data comes from the work
The sample is derived from proprietary, de-identified request and order patterns plus the decisions our team learned by fulfilling purchases manually. Customer records are not exposed.
We encountered the capability gap
We explored computer-use agents for this work and could not rely on them to place and reconcile orders correctly. Plausible mistakes in quantity, mapping, policy, and recovery kept the workflow human-operated.
WHY TRAIN ON PROCUREMENT
This is economically useful computer use with objective ground truth.
Long-horizon executionAgents must preserve intent across a workbench and multiple supplier portals.
Judgment under constraintsThe cheapest-looking product may violate clinical, functional, delivery, or authorization policy.
Consequential precisionA correct narrative is not enough; quantities, mappings, escalations, and terminal state must be right.
Dense learning signalEvery attempt receives structured verifier evidence and weighted reward, including near-complete failures.