Guide // MandateGuard
How do you add spending controls to a purchasing agent?
By the Ingenious AI team · Published 2026-11-14 · Reviewed 2026-11-14
▸ Run the free live demo See pricing
Free covers 1 agent and 1 policy with the full authorize flow, signed receipts and the approval inbox — no card required to start.
Short answer
Add spending controls to a purchasing agent by giving it a MandateGuard policy of its own before it places its first order: a per-task cap sized to its largest legitimate order, a daily cap for its busiest honest day, an allowlist of the suppliers that actually invoice you, and an approval threshold where order value starts to matter. The agent proposes every purchase through authorize and acts on ALLOW, DENY or REQUIRE_APPROVAL; approvals release a one-time token bound to supplier and amount. Meaningful task references — purchase order numbers — join the signed receipts to your own records.
What is a purchasing agent, in control terms?
A purchasing agent is an AI agent whose job includes initiating payments for goods or services: ordering, subscribing, topping up or renewing. Spending controls for such an agent are a policy attached to its identity in MandateGuard: the agent authenticates with its agent key, proposes each purchase through authorize, and receives ALLOW, DENY or REQUIRE_APPROVAL. The policy — caps, allowlist, threshold — is owned by the workspace owner and can be changed without touching the agent's code.
Give the agent an identity before a budget
Controls attach to an identified agent, so start there: create the workspace, register the purchasing agent, and keep its agent key in the agent's secret store like any other credential. The identity matters for accountability — receipts will name this agent's decisions, and if the business later runs a research agent too, the two histories stay separate. Avoid the shared-service-account pattern where several automations spend under one identity; when something odd appears in the chain, the first question is always which automation did this, and a shared identity makes that question unanswerable. One agent, one job, one policy is the unit of control.
Translate the purchasing job into a policy
Write the policy from the order book, not from imagination. The per-task cap is the largest legitimate single order. The daily cap is the busiest honest ordering day. The allowlist is the suppliers of record — the entities that actually invoice — and nothing else. Approve-over sits where order value starts to correlate with consequence in your business: the quantity step, the new product line, the rush order. MandateGuard's free plan covers this shape for a single agent completely, including receipts and the approval inbox; multiple purchasing agents or duplicate detection across orders are the natural triggers for Pro.
Make approvals part of the workflow, not an interruption
A purchasing agent that hits REQUIRE_APPROVAL should park the order, surface it to the buyer or owner, and continue with other orders. Design the surrounding process so approval latency is normal: a daily inbox review clears routine escalations, and the one-time token means an approved order proceeds for exactly the approved supplier and amount. Buyers should decide from the card — vendor, amount, task reference — and open the receipt history when something smells off. Denials from the inbox are process, not failure: they are how the business says not that one, with the decision recorded for the next audit.
Join receipts to your purchasing records
The task reference is the cheapest integration you will ever do. Require the agent to pass the purchase order number, ticket or requisition ID with every authorize call. Receipts then reconcile mechanically against orders and deliveries: every allowed spend has an order, every order has a chain of decisions, and variances show up as receipts without orders or orders without receipts. At month end, finance reads a verified ledger instead of a chat transcript. On paid plans, evidence export packages that ledger for auditors, and vendor trust aggregates each supplier's behaviour across all your purchasing agents.
Handle the catalogue edge cases deliberately
Real catalogues produce edge cases that break naive policies. Partial shipments split one order into several charges; each charge is its own authorize call, so size caps for the shipment shape, not just the order shape. Substitutions change amounts between proposal and charge; the approval token binds the approved amount, so genuine changes re-enter the flow honestly. Minimum-order fees push small orders over thresholds; if that pattern is routine, the threshold — not an exception habit — should move. Record each choice in the policy and, where the mandate needs task-level nuance, use a spend envelope for the exceptional job instead of loosening the standing policy.
Grow the control with the fleet
One purchasing agent under control is a solved problem. Growth brings the interesting risks: a second agent whose orders can duplicate the first's, seasonal volume that stresses daily caps, new suppliers seeking allowlist entry. Duplicate-commitment detection watches across the workspace ledger for the doubled order. New suppliers should enter through a probationary pattern — allowlist entry plus a temporarily low approval threshold — until their own receipts establish ordinary behaviour. The pattern stays constant: identity, policy from real prices, approvals as workflow, receipts joined to orders, and reviews on a rhythm.
Five actions to take now
- Give the purchasing agent its own workspace policy, not a shared team's defaults
- Allowlist the catalogue vendors it buys from, by charging identity
- Set approve-over for first orders and quantity steps
- Require meaningful task references naming the purchase order or ticket
- Review receipts against deliveries weekly to catch drift early
Worked example: a packaging reorder agent under control
A small distributor's purchasing agent reorders packaging stock from two allowlisted suppliers. Its policy: per-task cap 400 dollars, daily cap 900 dollars, approve-over at 250 dollars. Routine reorders at 120 to 220 dollars flow automatically with receipts naming the purchase order in the task reference. A 340-dollar order pauses for approval because it crosses the threshold; the owner confirms the quantity step matches a seasonal order and approves the bound token.
In the third month the agent attempts a 380-dollar order from Supplier B with a task reference naming a product line the distributor discontinued. The amount is in-policy, but the approval card makes the owner look — and catch a stale catalogue mapping before stock arrived for a dead product. Controls did not just bound the money; the approval step put a human eye on the one order that deserved it.
Frequently asked questions
Does the agent need my supplier account credentials?
Keep purchasing authority in the authorize flow. The agent proposes; the mandate disposes. Supplier accounts and payment execution should sit outside the agent wherever possible.
How do task references help accounting?
Receipts carry the reference — a purchase order or ticket — so finance can join receipt history to orders without reading agent chat logs.
What if prices vary by order?
Size the per-task cap to the largest legitimate order and put approve-over below it, so price spikes land in the inbox instead of failing or flowing silently.
Can one agent buy for several departments?
It can, but separate agents with separate policies make receipts attributable. Pro and Team support multiple agents for exactly this.
What happens on supplier price rises?
Over-cap denials and approval volume rise visibly. Renegotiate or adjust the policy deliberately; the receipts quantify the rise.
Keep reading
- How do you set up your first AI agent spending policy?
- How do you set guardrails for autonomous procurement agents?
- How does a vendor allowlist work for AI agent spending?
Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides