---
title: "How to make your business agent-ready"
description: "Eight steps that let AI agents find, buy from and get help from your business as customers, with a scorecard, disqualifiers and a checklist for this quarter."
type: Playbook
url: https://sunilkumar.ai/playbooks/how-to-make-your-business-agent-ready/
author: Sunil Kumar (https://sunilkumar.ai/about/)
datePublished: 2026-10-19
lastReviewed: 2026-10-19
version: "1.0"
pillar: Architecture
tags: [agents-as-customers, agentic-commerce, architecture, mcp, governance]
license: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
---

# How to make your business agent-ready

Eight steps that let AI agents find, buy from and get help from your business as customers, with a scorecard, disqualifiers and a checklist for this quarter.

## In short

- Agent-ready means an authorised agent can find your offer, use a door you control, prove who it acts for, pay, be held to terms you wrote, and reach a person when it is stuck. Most businesses pass none of the six today.
- Start by sending an agent to buy from yourself. The list of places it stops is the only readiness audit worth having, and it takes an afternoon.
- The standards exist. Shopify's Catalog and the Universal Commerce Protocol for the offer, MCP for the door, Web Bot Auth for identity, scoped tokens and single-use cards for payment. The work is adopting them, not inventing them.
- Score your business from 0 to 3 on eight steps. Anything under 2 on identity or terms means an agent cannot be trusted at your door, however good the catalogue is.
- Write terms for agents before the market writes them for you. A cap per transaction, a verification requirement and an unwind for a wrong purchase is a contract almost nobody in your category has published yet.

## Steps

1. **Send an agent to buy from yourself** Give a browsing assistant four tasks, find the price, buy something, change the order and get help, and write down every place it stops. That list is the audit.
2. **Publish the offer as data** Products, services, prices, availability, lead times and terms as structured data on your own domain, in the schema your category's agents already read, with the catalogue complete.
3. **Open one door, and watch it** An API or an MCP server for the three things an agent most wants to do, with logging from the first day, rate limits of its own, and a path through your bot defences.
4. **Verify the agent and bound its authority** Accept signed requests from registered agents, give each agent its own identity and scope, cap what it may do, and make revocation immediate.
5. **Take an agent's payment** Accept scoped tokens, single-use cards and advance mandates, and remove the steps in your checkout that only a human can complete.
6. **Write terms for agents, and the unwind** A page that says what an agent may buy, on what terms, what needs a person, and exactly how a wrong purchase is reversed and who pays.
7. **Give the agent a route to a person** A handover that carries the agent's context to a human on your side, with a response time, instead of a CAPTCHA followed by a chatbot.
8. **Measure agent traffic and review it monthly** Agent share of visits, orders and tickets, where agents stop, and what they tried that you do not offer, reviewed by one named owner.

Nine in ten merchants told Checkout.com this year that they are preparing for AI agents as customers. In my experience almost none of them can say what ready means. This playbook is the answer I give: eight steps that take a business from a website built for humans to one an authorised agent can buy from, with a scorecard to run against your own company and a checklist to hand to the team.

I wrote about [why agents are already customers and why businesses keep them out](/writing/agents-as-customers/). This is the practical part. It comes from the other side of the counter: for two years we have built agents that buy, book, reconcile and chase on behalf of clients, and I have watched them fail at other businesses' doors often enough to know where the doors are.

> **Agent-ready business**: A business that an authorised software agent can find, understand, transact with, be held to terms by and get help from, without a person on either side having to intervene in the normal case.

The principle behind all eight steps is one sentence: treat the agent as a customer with a different body. It cannot see a PDF, type a one-time code or wait on hold, and it carries its authority, its limits and its money in forms a human never did. Each step below gives it one thing it needs, and most of them make the business safer, not less safe, because an identified agent on a scoped door is easier to police than anonymous browser traffic.

## Step 1. Send an agent to buy from yourself

Before any plan, run the test. Open any capable assistant with browsing and give it four tasks against your own business: find the price of a specific thing you sell, buy it or request it, change the order, and get help with a problem. Do not help it. Write down every place it stops, what it saw and what it did next.

The output is a short list, and it is more useful than any readiness framework because it is yours. In the audits I have seen the same stops appear in the same order. A bot check on the first page. A price that is in a PDF, an image or behind a form. A checkout that needs a code sent to a phone. A chat widget that cannot hand over to anyone. Terms that the agent reads and cannot apply. Each one maps to a step below, so the list is also the plan.

Run it on a competitor while you are there. If an agent can buy from them and not from you, you have found the sentence for the board.


- [ ] Four tasks run by an unassisted agent: find, buy, change, get help.
- [ ] Every stop recorded with the page, the reason and what the agent did next.
- [ ] The same test run against one competitor.
- [ ] A named owner for the list and a date to run it again.


## Step 2. Publish the offer as data

An agent reads a schema. If there is no schema it guesses, and a guessing agent either buys the wrong thing or gives up. The first thing to publish is the offer: what you sell, at what price, with what availability, lead time and conditions, as structured data on your own domain.

For products, the schema your category's agents already read is the one to use. Shopify merchants are now in Catalog and enabled on the Universal Commerce Protocol by default, which is why Shopify can say that AI searches run on its structured catalogue convert at twice the rate of searches run on scraped pages. On other platforms the same job is done with schema.org product and offer markup, a product feed, or a catalogue endpoint. For services, it is a rate card and an availability calendar in a form a machine can query, which almost no services business has.

Completeness matters more than polish. An agent that finds 80 percent of your range assumes the other 20 percent does not exist. Lead times and stock matter more than photographs. And the terms belong with the offer, not three clicks away, because an agent evaluating an offer evaluates the terms at the same time.


- [ ] Every product or service with price, availability, lead time and conditions as structured data.
- [ ] The catalogue complete, including the things you would rather discuss on a call.
- [ ] Terms attached to the offer, not in a separate document.
- [ ] A way to tell when an agent read it: a feed or endpoint you can log.


## Step 3. Open one door, and watch it

The browser is the long way round for an agent, and the chat widget is a human interface pretending to be an API. What an agent wants is an endpoint it can call. In 2026 that means an API, or an MCP server, which has become the default way agents talk to tools; Shopify's developer edition now lets an agent register a profile and call a public MCP endpoint without approval.

Start with one door and three jobs: the three things an agent most wants to do with you, which the step one audit already told you. Check a price and availability. Place or request an order. Check the status of one. Describe each precisely, because the description is what the agent reads. I set out the economics of this from the agent builder's side in [MCP vs API](/writing/mcp-vs-api/): a well-described tool is cheaper than every agent learning the website, and the same arithmetic holds for the business receiving the agents.

Then watch it. Log every call from the first day: which agent, acting for whom, what it tried, what it got. Give the door its own rate limits so that a busy agent is throttled rather than banned, and a route through your bot defences so that step four's verified agents are not blocked by a rule written for scrapers.


- [ ] One endpoint, API or MCP, for the three jobs agents most want.
- [ ] Each operation described in plain language with its limits.
- [ ] Logging of agent, principal, request and outcome from day one.
- [ ] Rate limits of its own, separate from the bot-management rules.


## Step 4. Verify the agent and bound its authority

This is the step that makes the rest safe, and it is the one most businesses skip because it sounds like someone else's problem. It is not. An agent must be able to prove who it is, who it acts for and what it is allowed to do, and your systems must be able to check.

The standards arrived in the last year. Cloudflare's Web Bot Auth lets an agent sign every request with a key, using HTTP message signatures, so the receiving business can check the signature against a directory of registered agents and see whether the agent has come to browse or to pay. Visa's Trusted Agent Protocol and Mastercard's Agent Pay both use it, each with its own directory of registered agents. Mastercard added, on 30 September, a score that estimates whether a transaction was started by an agent at all, so that issuers can approve legitimate agent purchases rather than decline them as fraud.

On your side the work is three things. Accept signed requests and verify them, with your bot defences letting verified agents through and blocking everything else. Give each agent its own identity in your systems, tied to the human or company it acts for, with a scope: what it may read, what it may buy, up to what amount. And make revocation immediate: 75 percent of merchants in the Checkout.com survey called real-time revocation critical, and they are right, because an agent whose person has withdrawn consent must stop at the next request, not at the end of the month.

The refusal list we write for every agent we put in production has a mirror here. Your door needs a list of what an agent may not do whatever it asks, enforced in code rather than in a policy document; the [note on the agent that could not say no](/notes/the-agent-that-could-not-say-no/) is the same lesson from the other side.


- [ ] Signed agent requests verified against a registry; unsigned automation stays blocked.
- [ ] One identity per agent, linked to its principal, with a written scope and a cap.
- [ ] Revocation that takes effect on the next request.
- [ ] A refusal list for agents, enforced at the door, with a reason returned.


## Step 5. Take an agent's payment

An agent has no card in a pocket and no phone for the code. It pays with a token scoped to a merchant and an amount, a single-use virtual card issued for one purchase, a credential the person released for a bounded purpose, or an advance mandate. Stripe's Sessions announcements in April covered most of these: Shared Payment Tokens, single-use cards issued to agents, Link's agent wallet with spending approvals, and a way to distinguish legitimate agents from fraudulent ones. On UPI, the protocol NPCI is drafting would let an agent be registered, verified and authorised to pay within limits set by the person.

The work on your side is mostly subtraction. Find every step in your checkout that only a human can complete, the one-time code, the image puzzle, the "call us to confirm", and either remove it for verified agents or replace it with a check the agent can pass. Then make sure your payment provider's agent features are switched on; in several providers they now are by default, and in others they are a conversation to start this week.

Keep the human in one place: the approval. A person sets the cap and the scope once, in advance. The agent transacts within them without asking. Anything outside them goes back to the person, which is the same pattern we use when an agent acts on a client's own systems.


- [ ] Scoped tokens, single-use cards or mandates accepted, with your provider's agent features enabled.
- [ ] Every human-only step in checkout removed or replaced for verified agents.
- [ ] Caps and scopes set by the person in advance; anything outside them escalated, not declined silently.
- [ ] A receipt the agent can read and store, in data as well as on a page.


## Step 6. Write terms for agents, and the unwind

Nobody has settled who pays when an agent buys wrong. In the PYMNTS Intelligence survey, 93 percent of US retailers said the agent's provider should carry the loss, 80 percent expect the provider to verify the agent's authority first, and only 28 percent would offer agents their full range on the same terms. The sample was 60 merchants, which is small, but the direction is unmistakable, and the Federal Reserve's Christopher Waller named trust as the biggest barrier to scaling agentic commerce in his Sibos speech. Trust between businesses is written down, so write it.

A terms-for-agents page needs five things. What an agent may buy or request, and whether the range or the prices differ from the human channel; one third of the PYMNTS merchants intend different terms, which is fine if published and fatal if silent. The cap per transaction and per period above which a person must approve. What the agent must present: a verified identity and a principal. How a wrong purchase is unwound: the window, the method and who bears which cost. And how disputes are raised and answered, by the agent, by the person or by the agent's provider.

The business that publishes this first in its category is writing the contract the market has not written. It will get some of it wrong and revise it. It will also get the agents, because an agent chooses the business whose terms it can apply.


- [ ] A public terms-for-agents page, linked from the offer and returned by the door.
- [ ] Range and prices for agents stated, whether the same as the human channel or not.
- [ ] Caps, verification requirements and the approval path written as rules an agent can apply.
- [ ] The unwind for a wrong purchase: window, method, who pays.


## Step 7. Give the agent a route to a person

Every agent we run has a rule for what to do when it is not sure, and the rule is to stop and ask a human. Your business needs the receiving end of that rule. When an agent is stuck, at a question your door does not answer, a change your terms do not cover, a delivery that went wrong, it needs to hand the matter to a person on your side with the context attached, and get an answer it can act on.

Today the route is a chat widget that asks the agent whether it is a robot, or a phone number. Replace it with a handover: an endpoint or an email address that accepts the agent's transcript and the identifiers of the order, a person or a team that owns it, and a response time you publish. The agent's principal gets the answer either way, but the difference between a handover and a dead end is whether the transaction completes or the agent buys elsewhere. In the Checkout.com survey, a quarter of consumers said they would stop using an agent if a purchase were hard to dispute, which is a reminder that the person behind the agent is judging you too.


- [ ] A handover channel an agent can use, carrying the transcript and the order identifiers.
- [ ] A named owner and a published response time.
- [ ] Handover rates tracked per reason, so that recurring reasons become a door or a term.
- [ ] The person's answer returned in a form the agent can act on.


## Step 8. Measure agent traffic and review it monthly

What is not measured will be blocked again by the next security review. From the first day, know the agent share of your visits, orders and tickets, where agents stop, which agents came and for whom, and what they asked for that you do not offer. Shopify now shows merchants orders and sales from AI surfaces in one place; if your platform does not, your door's logs from step three will.

Review it monthly with one named owner, the same way the business reviews any channel. The three numbers to watch are the completion rate of the step one test, run again each month, the share of agent attempts that end in a handover, and the share that end in a purchase. The first should rise to five out of five tasks. The second should fall as recurring reasons become doors and terms. The third is the business case, and in my experience it is the number that turns the security review from a threat into a budget.


- [ ] Agent share of visits, orders and tickets reported alongside the human channels.
- [ ] The step one test re-run monthly and the completion score recorded.
- [ ] Handover reasons ranked and the top one fixed each month.
- [ ] One named owner for the agent channel, with the review in the calendar.


## The scorecard

Score each step from 0 to 3: 0 for nothing, 1 for a plan or a policy document, 2 for something built and in use, 3 for something built, measured and reviewed. Add the eight scores.

| Step | What a 3 looks like | Disqualifier |
| --- | --- | --- |
| 1. The test | Run monthly, five tasks out of five completed by an unassisted agent | Never run |
| 2. The offer | Complete catalogue as data, terms attached, reads logged | Prices in a PDF or behind a form |
| 3. The door | One endpoint for the three jobs, described, logged, rate-limited | "Agents can use the website" |
| 4. Identity | Signed requests verified, per-agent scope and cap, instant revocation | One API key shared by every caller |
| 5. Payment | Tokens, single-use cards or mandates accepted; no human-only steps for verified agents | A one-time code sent to a phone |
| 6. Terms | Public terms-for-agents page with caps and an unwind | Silence |
| 7. The person | Handover with context, named owner, published response time | A chat widget or a phone number |
| 8. Measurement | Agent channel reported monthly with a named owner | No way to tell an agent from a visitor |

Anything under 2 on identity or terms means an agent cannot be trusted at your door, however good the catalogue, and a total under 12 means agents are buying from your competitors. The [downloadable checklist](/downloads/agent-ready-checklist.md) has the grid and the questions for each step.

## If your customers send procurement agents, not shopping agents

Most of the published numbers are retail, but the enterprise version is larger. Gartner predicts agents will intermediate more than 15 trillion dollars of B2B spend by 2028, and Forrester expects 20 percent of B2B sellers to face agent-led quote negotiations in 2026. For a business that sells to other businesses, the eight steps are the same and the three jobs in step three are different: a quote, an order confirmation and an invoice, each as data an agent can take back to its principal.

The honest test for an enterprise is whether a customer's procurement agent can get those three things without a person in your sales team being the API. Where it cannot, the gap is usually step two, because the price list lives in a sales call, and step six, because the terms live in a master agreement the agent cannot read. Both are fixable in a quarter, and both are worth fixing before the agent on the other side starts comparing you with a supplier who has.

## How this looks from our side of the counter

I should say where the eight steps come from. At Ailoitte we build agents that act on clients' own systems, under the tests I call [Agentic in Production](/ideas/agentic-in-production/): real work, bounded failures, an audit trail, a named owner and a budget. Many of those agents are buyers. They reorder stock, reconcile invoices, chase deliveries and compare quotes, and every one of them carries a refusal list, a cap and a rule to stop and ask a person. When one of them meets a business that passes these eight steps, the transaction completes in seconds and the trace shows exactly what happened. When it meets one that does not, it raises a ticket and a person makes a phone call two days later. The playbook is the difference between those two outcomes, written for the business on the other side.

## If you are doing this in your company

Run step one this week, on yourself and on one competitor, and put the list in front of the leadership team; it will make the case better than this playbook can. Give the agent channel a named owner. Publish the offer as data and open one door with logging, which is a month of one team's work. Write the terms-for-agents page before anyone asks for it. Start the identity and payment conversations with your platform and your payment provider now, because those are the steps that depend on other people's timelines. Then score yourself, and score again in a quarter.

If you would like an outside pair of eyes on where your business stops an agent, that is a two-week review, and the [work with me](/work-with-me/) page has the shape of it.

## Sources

- Checkout.com, [Consumer demand for AI shopping is forming fast but trust for agentic commerce is still catching up](https://www.checkout.com/newsroom/consumer-demand-for-ai-shopping-is-forming-fast-but-trust-for-agentic-commerce-is-still-catching-up), press release, 9 June 2026, and the report [Agentic Commerce 2026](https://www.checkout.com/guides-and-reports/agentic-commerce-2026): 89 percent of merchants preparing, 75 percent on real-time revocation, 25 percent of consumers who would stop using an agent if a purchase were hard to dispute.
- PYMNTS Intelligence, _Will the 2026 Shopping Season Go Agentic?_, reported in [AI makes the holiday shopping list, but who gets the sale?](https://www.pymnts.com/news/artificial-intelligence/2026/ai-makes-the-holiday-shopping-list-but-who-gets-the-sale/), 15 September 2026: 2,061 US consumers and 60 US retail merchants; the 93 percent, 80 percent, 28 percent and one-third figures.
- Shopify, [Selling everything, everywhere, all at once: the Spring '26 Edition](https://www.shopify.com/news/spring-26-edition-merchant) and [Agentic commerce for every developer: the Spring '26 Edition](https://www.shopify.com/news/spring-26-edition-dev), 17 June 2026: Catalog and the Universal Commerce Protocol enabled by default, the conversion comparison, agent registration and the public MCP endpoint, orders from AI surfaces in the admin.
- Stripe, [Everything we announced at Sessions 2026](https://stripe.com/blog/everything-we-announced-at-sessions-2026), 29 April 2026: Shared Payment Tokens, Issuing for agents, Link's agent wallet, Radar bot abuse prevention and agent guardrails.
- Lohe, R. and Allen, W., [Securing agentic commerce: helping AI agents transact with Visa and Mastercard](https://blog.cloudflare.com/secure-agentic-commerce/), Cloudflare, 24 October 2025: Web Bot Auth, HTTP message signatures (RFC 9421), Visa's Trusted Agent Protocol and Mastercard's Agent Pay.
- Mastercard, [Mastercard expands Agent Pay with new trust and intelligence services](https://www.mastercard.com/vn/vi/news-and-trends/press/2026/september/new-trust-and-intelligence-services.html), press release, 30 September 2026.
- Waller, C. J., [Payments in the Age of AI Agents](https://www.federalreserve.gov/newsevents/speech/files/waller20260929a.pdf), speech at Sibos 2026, Miami, 29 September 2026.
- Business Standard, as reported by Outlook Business, [India plans AI-powered UPI payments framework through Unified Agent Protocol](https://www.outlookbusiness.com/news/india-plans-ai-powered-upi-payments-framework-through-unified-agent-protocol), 9 July 2026.
- Gartner, _Gartner Unveils Top Predictions for IT Organizations and Users in 2026 and Beyond_, press release, 21 October 2025, as reported by [Digital Commerce 360](https://www.digitalcommerce360.com/2025/11/28/gartner-ai-agents-15-trillion-in-b2b-purchases-by-2028/), 28 November 2025.
- Forrester, [Predictions 2026: the "agentic commerce" race, and some potential regrets, in digital commerce](https://www.forrester.com/blogs/predictions-2026-the-agentic-commerce-race-and-some-potential-regrets-in-digital-commerce).
- Model Context Protocol, [specification](https://modelcontextprotocol.io/specification/latest), for the door in step three.


## Questions people ask

**We are a services business with no online checkout. Does this apply to us?**

Yes, with step five moved later. A services business is bought by agents in two ways, a quote request and a booking, and both are usually a form built for a human. Publish the service catalogue and the rate card as data, open a door for quotes and bookings, write the terms, and add payment when you have agents asking to pay.

**Will opening a door to agents expose us to more fraud and scraping?**

It does the opposite if you do step four. Today your defences cannot tell an authorised agent from a scraper, so they block both or admit both. A signed, registered agent with a scope and a cap is easier to police than anonymous browser traffic, and everything that does not sign can stay blocked.

**Should we give agents the same prices and terms as human customers?**

Decide it deliberately and write it down. One third of US retailers told PYMNTS they would offer agents the full range on different terms, and that is a legitimate choice if the terms are published. What fails is silence, because an agent that cannot find the terms assumes the public ones apply, and the dispute lands on you.

**Which standard should we build to, with so many protocols announced?**

Build to the layer below the protocols. Structured data for the offer, an API or MCP server for the door, signed requests for identity and tokenised payment. The commerce protocols, Google and Shopify's Universal Commerce Protocol among them, sit on top of those and are far easier to add once the layer below exists.

**How long does this take?**

The audit in step one takes an afternoon. A first door, the terms page and the handover path are a month of one team's work for a mid-sized business, less on a platform that already has the catalogue and payment pieces. Identity and payment depend on your providers and are the steps to start conversations about this week.


---
Sunil Kumar, Co-founder and CEO, Ailoitte Technologies. https://sunilkumar.ai/about/