A payment system normally assumes that a person sees an offer, approves a cart and initiates the transaction. Agentic commerce changes that sequence. A person may give an AI agent an instruction in the morning, while the agent compares options and completes a purchase hours later. The interface becomes easier; the trust problem becomes harder.
My thesis is that agentic commerce is an authorization business. The durable value is not simply helping an agent buy faster. It is proving that the agent is genuine, that its action reflects the user's intent, that it stayed inside defined limits, and that every participant can reconstruct what happened if the outcome is disputed.
The protocols are converging on evidence
Google's Agent Payments Protocol describes authorization, authenticity and accountability as the central questions created when an agent transacts for a user. Its proposed answer is a chain of cryptographically signed mandates: an intent mandate records the instruction, and a cart mandate records the specific items and price. In April 2026, Google said it was donating AP2 to the FIDO Alliance and released version 0.2 with support for human-not-present payments.[1][2]
OpenAI and Stripe's Agentic Commerce Protocol takes a complementary approach. The merchant remains merchant of record, validates the order, calculates tax, handles fulfillment and decides whether to accept or decline. Its delegated-payment flow can bind a one-time request to a maximum amount and expiry before a payment token reaches the merchant. These are important boundaries; they reduce credential exposure but do not remove merchant, payment-provider or regulatory responsibility.[3]
Visa's Trusted Agent Protocol focuses on recognizing an approved agent and verifying signed request data, including the agent's purpose and commerce intent. Reuters reported in September 2026 that Visa, Mastercard and Ant International had also launched an initiative for common agent-identification and verification standards. That initiative signals demand for interoperability, but the public announcement did not establish a completed universal standard or deployment timetable.[4][5]
A purchase becomes a control system
Once an agent can act without the user present, a purchase is no longer one approval screen. It is a control system. The system must bind identity, instruction, policy, product, price, timing, payment credential and final outcome. It must also handle substitutions, price changes, unavailable inventory, delivery constraints and attempts to manipulate the agent.
Consider an illustrative instruction: buy a flight if the total refundable fare falls below SAR 1,800 before 18:00. A responsible system should not interpret that as permission to buy any itinerary under the limit. The departure city, dates, baggage, connection rules, traveler identity, refund conditions, currency, tax treatment and notification threshold all matter. The mandate needs enough structure to prevent a plausible but unwanted purchase.
The entrepreneurial mechanism: trust infrastructure
The entrepreneurial opportunity sits between the agent, merchant, payment provider and regulated institution. Products may include agent identity and attestation, mandate creation, policy engines, transaction limits, secure token exchange, consent records, risk scoring, exception routing, receipt verification and dispute evidence. The best product will not merely produce logs. It will make the authorization state understandable to a user, merchant, risk team and regulator.
A possible business model could combine integration fees, transaction-based verification, enterprise policy management and compliance evidence. That is a hypothesis, not a claim that any venture has earned such revenue. A pilot should test whether the trust layer reduces fraud, manual review or dispute time without adding so much friction that customers abandon the channel.
For FinTech founders, this creates a sharper product boundary. Building an autonomous shopping assistant is one product. Holding funds, initiating payments, issuing instruments, operating payment infrastructure or making regulated decisions may trigger different obligations. A persuasive demonstration should never be treated as evidence that the underlying regulated activity is licensed or operational.
The counterargument: payment networks already solve trust
A reasonable counterargument is that cards, wallets, tokenization, authentication and fraud systems already allocate responsibility. Agentic commerce could become another channel on top of mature payment rails, with no need for a separate category of companies.
That may be true for the transfer of value. It is less complete for the transfer of intent. Existing rails can show that a credential was used, yet the dispute may concern whether the agent was allowed to choose this merchant, accept this substitution, disclose this data or act at this time. Protocols can standardize evidence, but businesses still need policy, integration, monitoring and a human resolution path.
Saudi Arabia: permission must survive regulation
In March 2026, the Saudi Central Bank issued an updated oversight framework for payment systems and their operators, aligning supervision with the Law of Payments and Payment Services and its implementing regulations. The framework emphasizes oversight tools, self-assessment, assessment by the authority and obligations aligned with applicable regulation and international standards.[6]
SAMA's implementing regulations require risk-based transaction limits for payment service users, protection of personalized security credentials and secure handling of client data. They also condition payment-account access for initiation and information services on user consent and secure communication. An AI agent does not remove these requirements; it makes the evidence of consent, limits and secure delegation more important.[7]
Saudi personal-data rules are also relevant. The PDPL defines processing broadly, including automated operations, while its implementing regulation addresses transparency when decisions are made solely through automated processing. Agentic-commerce design should therefore minimize data, explain the decision path where required, preserve withdrawal and correction mechanisms, and avoid treating broad access to personal information as implicit purchasing authority.[8]
For national entities and critical infrastructure, the National Cybersecurity Authority's Essential Cybersecurity Controls reinforce documented governance, approved responsibilities and accountable cybersecurity management. The practical lesson for founders is that agentic-payment security cannot be left inside a model prompt. It needs an owned control environment.[9]
GCC expansion must be re-authorized
A Saudi design can provide a strong base for Gulf expansion, but it cannot be copied as if the GCC were one regulatory market. Payment licensing, consumer protection, tax, data localization, identity infrastructure, dispute rules and accepted authentication methods differ. The mandate and evidence model may travel; the legal and operational acceptance must be revalidated country by country.
Across the wider MENA region, the variation is greater. Some markets have mature instant-payment and digital-identity infrastructure; others rely on cash, fragmented merchant systems or uneven connectivity. Agentic commerce may create value first in bounded enterprise procurement, travel or recurring replenishment rather than unrestricted consumer autonomy. That is a market hypothesis to test, not a regional forecast.
Five controls before an agent can spend
First, authenticate the person and the agent separately. Second, express authority as a narrow mandate with amount, time, merchant, product and substitution limits. Third, separate planning from execution so a compromised recommendation cannot silently become a payment. Fourth, use a scoped credential that cannot be replayed outside the approved transaction. Fifth, produce a receipt that links the instruction, policy checks, merchant decision, payment result and any human intervention.
The acceptance test should include hostile conditions: prompt injection from merchant content, changed prices, conflicting instructions, expired mandates, unavailable products, partial refunds, duplicate callbacks and an offline approver. The useful metric is not how many purchases an agent can complete. It is the percentage completed inside mandate, the exception rate, unauthorized-attempt prevention, dispute-resolution time and losses after controls.
The founder's decision
Founders should decide whether they are building an agent, a trust layer, a merchant integration or a regulated payment service. Buyers should require a clear responsibility map before a pilot: who authenticates the user, who signs the mandate, who verifies the agent, who holds credentials, who can stop execution and who resolves a disputed transaction.
Agentic commerce will not be won by the system that spends most autonomously. It will be won by the system that can act usefully inside explicit authority—and prove it.
Key takeaways
- Treat agentic commerce as an authorization chain, not merely a conversational checkout.
- Bind identity, intent, amount, timing, product limits and payment credentials into verifiable evidence.
- Separate planning from execution and test hostile, expired and disputed transaction paths.
- Revalidate licensing, consent, data and dispute controls separately in each GCC and MENA market.