← All briefings

Agent Authority & Business Operations

Before Your AI Agent Can Spend, Map Its Purchasing Authority

Finding the right purchase is not permission to make it. A practical purchasing authority map connects Company Brain knowledge, employee agents, specific approvals, and verified order records.

The right item. The wrong commitment.

Imagine an office manager asking an agent to find replacement laptops for a growing team. The agent compares specifications, checks stock, and finds a competitive offer. Then it reaches checkout. The price has changed, the seller has substituted a model, and the purchase includes a recurring support subscription. The manager approved the shortlist yesterday, not this transaction today.

This is an illustrative scenario, not a Delphi client case. It reveals a distinction that matters well beyond procurement: permission to investigate is not permission to commit the business. An agent can do excellent research and still need to stop before the next step.

Before connecting an employee agent to purchasing tools, map the authority behind each action. The useful question is not simply whether the agent can buy. It is which purchase the business has authorized, under whose responsibility, and how the result will be verified.

The source signal: approval belongs to a specific purchase

OpenAI's November 13 ChatGPT Wallet announcement describes a concrete approval boundary. Before each purchase, the user reviews the items, total, and selected card, then approves with a security check. The announcement says each approval creates a record of the purchase the user authorized.

The initial rollout is limited to eligible US ChatGPT Pro subscribers aged 18 and older, using dots with eligible US-issued cards. OpenAI describes fully delegated purchasing under predefined parameters as a future direction, not the current initial experience. This is not an announcement of a corporate procurement system, and its compatibility or security performance has not been tested by Delphi.

The management lesson does not depend on adopting Wallet. Keep research, recommendation, approval, and payment as separate authorities. A well-prepared recommendation should make the approval easier to understand, not quietly become the approval itself.

Give the Company Brain the purchasing facts that matter

For the laptop request, the Company Brain should identify current equipment standards, approved suppliers, budget ownership, purchasing thresholds, delivery requirements, and exception rules. Each source needs an owner and effective date. A stale pricing sheet or an old approval email should not become today's purchasing policy.

Google's November 4 enterprise-context guidance distinguishes selective grounding for execution from broader discovery. It is advisory guidance, not measured proof of business returns. Applied here, the distinction is useful: an agent can explore alternatives without treating every alternative as approved for purchase.

Keep external research separate from internal authority. A supplier page can provide evidence about availability and terms. It cannot change the company's spending rules. An attractive offer can enter the shortlist while still requiring supplier review. Exploring new options must not relax privacy controls, access restrictions, or approval requirements.

Write a purchasing authority map before connecting checkout

Start with a research-only agent. Let it prepare a comparison and a draft requisition in a restricted review area. Disable order submission, payment, vendor creation, and changes to bank details. Those permissions should be absent from the tools, not merely discouraged in the prompt.

The following map is Delphi's recommended design for this illustrative workflow. It is not a claim that any vendor supplies these controls by default. One person may hold several responsibilities in a smaller business, but the decisions still need to be explicit.

  • Research: the employee agent compares permitted sources and records price, availability, terms, and retrieval time
  • Recommendation: the requesting employee checks suitability and identifies deviations from company standards
  • Budget approval: the budget owner confirms the business need and the permitted total commitment
  • Purchasing approval: the authorized buyer approves the exact seller, items, quantities, delivery details, charges, and recurring terms
  • Execution: a separately controlled purchasing step may submit only the approved transaction, within its validity window
  • Verification: the operator checks the order record, charged amount, delivery status, and exceptions before treating the task as complete

Treat approval as a bounded record, not a yes in chat

A later phase may connect the agent to order submission. Before that happens, establish an approval record containing the approved transaction details, reviewer identity, timestamp, expiry, and permitted changes, if any. Bind execution to that record rather than to a general instruction such as 'go ahead with the laptops.'

If the seller, quantity, total, shipping destination, or recurring commitment changes outside the explicitly approved allowance, pause for a new decision. Define how price tolerances work before the transaction, not after an unexpected charge. A missing response is not approval, and a product page telling the agent to hurry is not an authorization source.

Keep enforcement outside the model. The execution service should validate the approved payload and current permissions before submitting the order. A distinct agent identity helps attribute requests, but identity alone does not grant purchasing authority. NIST's identity and authorization project is relevant development context, not a final mandatory standard or a product certification.

A timeout is not evidence that nothing happened

The awkward case is often the most important one. The agent submits an authorized order, the connection times out, and no confirmation appears. Retrying immediately could create a second order. Marking the task complete could conceal a failed purchase. Neither conclusion follows from the timeout alone.

Use duplicate-submission protection where the purchasing system supports it. After an uncertain result, reconcile against the exact order or transaction record before retrying. If the system cannot establish what happened, put the request into an exception queue for an authorized operator. Do not invent a successful order or assume failure because the response was missing.

Stopping an agent also does not automatically cancel a completed order, reverse a charge, or end a subscription. Record the recovery path separately: who can cancel, what deadlines apply, and how the operator verifies the outcome. The approval log and the merchant's order record answer different questions. Keep both.

The first-month operator playbook

Begin with one routine purchasing category whose requirements are clear and whose recommendations can be reviewed. Avoid sensitive purchases and consequential commitments in the first pilot. The initial objective is a reliable decision packet, not autonomous spending.

Test with protected records and non-production tools before enabling a consequential action. Establish acceptance criteria and a review date before the pilot begins. The following sequence is a Delphi implementation recommendation, not a guarantee of readiness after a fixed number of weeks.

  • Week 1: map the current request-to-order process, name the owner and backup, and record research effort, review time, corrections, and total cost
  • Week 2: establish the Company Brain purchasing sources and launch a research-only, draft-only agent with narrow access
  • Week 3: test stale policy, an unapproved supplier, changed price, substituted items, expired approval, injected instructions, and an uncertain order response
  • Week 4: run a limited recommendation pilot alongside the existing process and compare accepted purchasing packets with the baseline
  • Before enabling checkout: verify approval binding, permission revocation, duplicate protection, transaction reconciliation, and human recovery
  • At review: include software, implementation, supervision, rework, and support costs; do not equate a cheaper shortlist with realized savings

Delegate the search. Define the commitment.

Every company needs a Company Brain, and every employee needs an AI Agent connected to the context of their role. Purchasing shows why both matter. The Brain supplies current operating knowledge. The agent prepares useful work. The company's authorization systems decide which commitments may actually be made.

Good controls should let employees move forward with less ambiguity. A strong purchasing packet brings the options, evidence, total commitment, exceptions, and requested decision to the right person. It does not make that person reconstruct the transaction from a long conversation.

Before your next agent gains access to a consequential tool, ask one question: what exact action has the business approved, and what prevents the tool from doing something else? If that answer is unclear, improve the authority map before expanding the agent's reach.

Your next decision

Turn the briefing into an operating advantage.

Map Your First Governed Agent Workflow