Guide // MandateGuard
Glossary: AI agent mandate and spending terms explained
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
This glossary defines the mandate vocabulary operationally, as MandateGuard's mechanisms use it: mandate, policy, envelope and intent as the scope ladder of spending authority; per-task cap, per-day cap, approval threshold, cumulative and per-period budgets as the five distinct limits; ALLOW, DENY and REQUIRE_APPROVAL as the three decisions, with the bound one-time token and the freeze as their companions; receipts, the hash chain, verification, evidence packs, vendor trust and duplicate detection as the evidence system; and workspace, agent and role keys as the access model. Shared, mechanism-true vocabulary is governance infrastructure.
Why operational definitions?
This glossary defines the working vocabulary of agent spending mandates as MandateGuard uses the terms: mandate, policy, caps, allowlist, approval threshold, the three decisions, receipts and chains, tokens, freeze, envelopes, intents and role keys. Definitions are operational — each states what the thing does in the system — because in governance, a term you cannot connect to a mechanism is a term that will be misunderstood at the worst moment.
Authority terms: mandate, policy, envelope, intent
A mandate is scoped spending authority: who may spend, on what, with whom, how much, under what supervision. A policy is a standing mandate attached to one agent — caps, allowlist and threshold, governing all its ordinary work until changed. A spend envelope is a mandate scoped to one task: its own maximum, vendors, period and conditions. An intent is a mandate built to be shared across organisations: registered, referenceable by an intent_id, with cumulative and per-period budgets and an authority window. The four terms form a scope ladder — agent, task, relationship — and most confusion in this field comes from using one rung's word for another's mechanism.
Limit terms: caps, thresholds, budgets
A per-task cap (per-transaction cap) is the most an agent may spend in one authorisation. A per-day cap is the most it may spend across a day. An approval threshold (approve-over) is the amount above which an in-policy spend needs a human decision. On intents, a cumulative budget caps the mandate's total lifetime spend and a per-period budget caps each period's pace. These five numbers do different work — single-event, daily aggregate, supervision, lifetime total, pace — and a mandate missing the relevant one has a hole of exactly that shape. When reading any system's limits, name which of the five each number is; marketing that blurs them is hiding the holes.
Decision terms: ALLOW, DENY, REQUIRE_APPROVAL, token, freeze
ALLOW: the proposal fits the mandate; spend may proceed. DENY: it breaks a hard bound; the reason is recorded and the identical proposal will be refused identically. REQUIRE_APPROVAL: it fits the hard bounds but crosses the supervision line, so it waits for a human. A spend token is what approval mints: bound to vendor and amount, single-use, the only way the approved purchase executes. Freeze (the kill-switch) is an agent state in which every proposal is denied until an owner unfreezes. Suspend and resume are the same idea for intents; revoke ends an intent. Notice the vocabulary's honesty: every term names a state a machine enforces, not an intention a document expresses.
Evidence terms: receipt, chain, verification, trust
A receipt is the signed record of one decision — vendor, amount, reference, decision, reason. The chain is the receipts' hash-linked order, each entry committing to its predecessor. Verification (verify-chain) is recomputing the chain to prove no record was altered. An evidence pack is the exported period record with its verification, for auditors. Vendor trust is the per-vendor summary computed from your own receipts — allowed and denied counts and totals — and duplicate-commitment detection is the ledger pattern-match for the same vendor and amount committed twice inside a window. Evidence terms matter because they are what survive the meeting: when memories differ, the chain is the arbiter all parties agreed to in advance.
People and access terms: workspace, keys, roles
A workspace is the owning account: its key manages agents, policies and approvals. An agent key identifies one agent's proposals; it can ask, never approve. Role keys serve shared mandates: issuer keys write intents and manage their status; merchant and agent keys read intents and verify receipts only. An Agent Passport ID, attachable to an agent record, names the agent's registered identity in the fleet's identity system — separate from spending authority by design. The access vocabulary encodes the trust model: asking, approving, issuing and verifying are different powers, held by different credentials, and blurring them in your own runbooks recreates the risks the separation exists to prevent.
Using the glossary as an operating tool
Pin the glossary where mandates are worked: beside the policy register, inside approver onboarding, at the front of the incident runbook. In incidents, force the vocabulary — say the decision, the reason, the receipt number — because precise terms compress diagnosis (a DENY for daily cap with a retry pattern is a complete situation report). In procurement, use the glossary as a translation test for vendor claims: ask each vendor to map their feature names onto these mechanisms and watch where the mapping fails. And when your organisation coins local terms (approver bands, cost-centre references), annex them here rather than letting them drift through chat. Governance fails in language before it fails in numbers; this page is the cheap defence.
Five actions to take now
- Keep this glossary beside your policy register; link terms in your own runbooks
- When a vendor uses a term differently, ask which definition their system enforces
- Teach approvers the three decisions and the token before their first queue shift
- Revisit the glossary when new mandate shapes (intents, envelopes) enter your fleet
- Add your organisation's own terms (cost centres, approver bands) as a local annex
Worked example: a new approver's first week
A new approver joins a team's rotation and is handed this glossary plus the policy register. Her first week produces three small saves that trace directly to vocabulary: she denies an approval whose vendor is allowlisted but unfamiliar, because she knows allowlist membership and approval judgement are separate questions; she recognises a repeated vendor-and-amount request as a possible duplicate commitment and asks for the original receipt; and she refuses a request to just raise the cap temporarily, because she has learned caps change in reviews, not in queues.
None of those judgements required seniority — they required the words to mean the mechanisms. That is the glossary's job in an onboarding pack, and why it sits beside the register rather than in a wiki nobody opens.
Frequently asked questions
Is this vocabulary standard across vendors?
The concepts converge (the EMVCo draft and AP2 use mandate language similarly), but each system's exact semantics are its own. Always map a vendor's terms to its mechanisms.
What is the difference between a mandate and a policy here?
Mandate is the general concept — scoped spending authority. A policy is MandateGuard's standing implementation of it for one agent; envelopes and intents are other mandate shapes.
Where do I see these terms in the product?
Policies and receipts in the workspace flows, envelopes at /envelopes, intents in the registry at /intents, and every term's behaviour is demoed on the homepage's live demo.
Should our runbooks reuse these definitions?
Yes — copy them, then annex your local terms. Shared vocabulary between the tool and the runbook is what makes incidents communicable.
What term causes the most confusion?
Authorise. In payments it means the instrument checked out; in MandateGuard the authorize decision is the governance act. Keep the two apart in every document.
Keep reading
- What is an AI agent spending mandate?
- ALLOW, DENY or REQUIRE_APPROVAL: how agent spend decisions work
- What is a spend envelope for an AI agent?
Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides