← Go back

Guide // MandateGuard

How do you set up your first AI agent spending policy?

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

Setting up your first MandateGuard policy takes five steps: create a workspace and keep its key safe, register one agent for one named job, price that job honestly from real vendor prices, then set the four policy parts — vendor allowlist, per-task cap, per-day cap and approve-over threshold — in that order. Prove it with four deliberate test purchases covering ALLOW, both DENY shapes and an approval round-trip, and put the approval inbox into a named person's daily routine before go-live. The free plan covers this completely for one agent; the receipts from week one will tell you the one adjustment the policy needs.

What does a first policy need?

A first agent spending policy is the initial MandateGuard policy attached to your first agent: per-task cap, per-day cap, vendor allowlist and approve-over threshold, set from the prices of the one job the agent will do. The free plan supports the complete shape — one agent, one policy, full authorize flow, receipts, approval inbox and approval webhook — so a first policy is a real control, not a trial of one.

Step one and two: workspace and agent

Create the workspace with your organisation's name and email; the workspace key it returns is shown once, so store it like a password — it is the credential that manages agents, policies and approvals. Then register the agent for one named job: not automation or helper, but the role it will spend in, such as media-buyer or enrichment-agent. The agent key is likewise shown once and belongs in the agent's secret store. This identity groundwork feels administrative, but it is what makes every later receipt attributable: decisions in the chain belong to a named agent acting for a named workspace, which is the unit your future audits and incident reviews will think in.

Step three: price the job before touching settings

Collect the job's real numbers first: the vendors it pays, their typical prices, the largest legitimate single purchase, and the busiest plausible day. Founders routinely discover at this step that the job's price shape is not what they assumed — the dataset is 62 dollars, not 40; the busy day is nine purchases, not four. Setting policy from those numbers produces caps that fit like a glove: tight enough to matter, loose enough to work. Setting it from round numbers produces a month of exceptions. If you cannot list the job's purchases, the agent is not ready for spending authority at all — give it a research-only role until its purchasing pattern is concrete.

Step four: the four settings, in order

Set the vendor allowlist first, because it defines the world the other numbers operate in: only the sellers this job pays. Set the per-task cap a little above the largest legitimate single purchase. Set the daily cap around the busiest honest day, remembering retries happen. Set approve-over at half to two-thirds of the per-task cap, so the top band is supervised. That order — who, how much at once, how much per day, what needs a human — mirrors how the engine evaluates, and it prevents the common incoherence of a threshold above the cap (which can never trigger) or a daily cap below the per-task cap (which makes the largest legal purchase consume the whole day).

Step five: prove the mandate with deliberate tests

Before real work, fire four test purchases through the agent's real integration: a small in-policy buy (expect ALLOW), an over-cap amount (expect DENY), an unlisted vendor (expect DENY), and an amount above the threshold (expect REQUIRE_APPROVAL — then approve it and redeem the token). Each outcome proves one branch of your agent's behaviour and one face of the policy. Read the four receipts in the chain and confirm the task references read the way you want your history to read. Teams that skip this step discover their agent's DENY behaviour during a real incident, which is the most expensive classroom there is.

Go-live: the human routine is part of the policy

A policy is half settings, half routine. Name the approver and check the inbox daily from day one, even when the queue is empty — the habit must exist before the day it matters. Schedule the weekly receipt skim and the monthly policy review now, while the workspace is small. Set the approval webhook if your team lives in a chat tool, so pending decisions arrive where people are. Then let the agent work. Most first policies need exactly one adjustment in their first month, and the receipts will identify it unambiguously: the bound that blocks real work, or the band where approvals have become reflexes.

Growing from one policy

The first policy teaches the method; scale is applying it per job. A second agent gets its own policy fitted to its own prices rather than inheriting the first agent's numbers, which is why Pro and Team exist: multiple agents and policies, plus duplicate detection across the workspace, vendor trust from your growing ledger, the kill-switch and evidence export. Resist the urge to create one generous shared policy for convenience — shared policies dissolve attribution, and attribution is what makes the whole system governable. One agent, one job, one policy, tuned from its own receipts: that is the shape that scales.

Five actions to take now

  1. Create the workspace and save the workspace key somewhere safe — it is shown once
  2. Register the agent for one named job and save its agent key
  3. Set the policy from the job's real prices: caps, allowlist, threshold
  4. Run the demo flow or a small live test through all three decisions
  5. Put the approval inbox into a named person's daily routine today

Worked example: a first policy that survived its first month

A solo founder sets up her first policy for an agent that buys stock media and model API credits. Real prices: media packs 15 to 45 dollars, API top-ups 20 dollars, a busy day around 90 dollars. Policy: per-task cap 50 dollars, daily cap 120 dollars, allowlist of the media library and the model provider, approve-over at 35 dollars — so the largest media packs pause for her, and top-ups flow.

Week one produces 26 receipts: 21 allows, 3 approvals she granted, 2 denials — one unknown vendor the agent found via search, one 70-dollar bundle over the cap. She changes nothing: the denials were the mandate working, and the approval volume (about one every other day) fits her routine. The policy she set in twenty minutes carried the month without adjustment.

Frequently asked questions

How long does setup take?

Workspace, agent and policy creation take minutes through the API or demo flow. The thinking — pricing the job honestly — is the real work.

What if I set the caps wrong?

You will learn from receipts within a week, and policies are editable at any time. Start conservative where the job allows and tune from evidence.

Do I need a credit card for the free plan?

Free covers one agent and one policy with the full flow. Paid plans are bought through the PayPal door or crypto checkout on the pricing page.

Can I copy the demo policy?

The homepage demo (cap 50, day 100, approve over 25, allowlist openai and aws) is a reasonable template for a small API-buying agent — adjust the vendors and prices to your job.

When should I upgrade to Pro?

When you run a second agent, need duplicate detection or vendor trust, or want evidence export. The history you built on free carries over.

Keep reading

Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides