The first two parts of this series assumed something so obvious it barely needed saying: the agent comes through your website. Part one looked at agents as a second kind of visitor to the page. Part two looked at how they depend on your forms, errors and pricing to finish a task. In both, the page was still the way in.
That won't always be true. An agent doesn't need a visual website. It needs a task completed, and for a growing number of tasks the page is a detour: a layer built for human eyes that the agent has to reverse-engineer. Where a structured tool or API exists, it'll use that instead.
The website isn't about to disappear. But for part of what it does today, it may stop being the primary interface. For a CEO, that's a strategic question more than a design one: which of the things your site does should a machine be able to do directly, which should stay in front of a person, and who controls the connection between them. What follows covers that split, what to prepare now, and what's still too early to build.
What an agent actually wants
An agent asked to compare three providers, book a call or request a quote has one goal: get the task done and report back. It has no use for your hero image, your animation or your menu structure. It only touches them because, today, they're the way in.
Going through a page is the slow route. The agent reads the rendered layout, guesses which button starts the process, fills in fields it has to infer, and hopes the page behaves the same way next time. Every redesign can break it. A structured tool removes the guessing: the site, or a service behind it, states which actions exist, what each one needs and what it returns.
Structured access comes in several forms: an API a company already runs for its own systems, a defined set of tools an agent can call through a protocol such as MCP, or, at proposal stage, browser-level standards such as WebMCP, which would let a page declare functions and annotated forms as tools for agents working in the browser. WebMCP is in early testing in Chrome, and support elsewhere is not settled, so treat it as a direction, not a decision. The details will change. The principle won't: the site tells the agent what it can do, instead of the agent working it out.
Two layers, two jobs
Right now your website does two things at once, and it helps to pull them apart.
The human-facing layer explains, persuades and proves. It carries your positioning, your people, your evidence, your pricing logic and the reasons someone should pick you.
The machine-facing action layer does things. It searches a catalogue, checks availability, takes a booking, accepts a quote request, returns an order status.
On most B2B sites the two are fused: the action is a button at the end of a persuasive page. For a person, that works. For an agent, the action is buried in a layout designed for a different reader.
Splitting them doesn't mean building two websites. It means recognising that a task and the page that sells it are different jobs, and that the first can be offered to machines without the second changing at all.
Which tasks are worth exposing
Not every action on your site is a candidate. A task is worth making directly available to machines when it's well defined, repeatable and low in risk. Five common examples show the range:
- Search. Looking up services, products, documentation or availability. Read-only and structured, so it's the safest place to start.
- Product or service comparison. Returning your offer's attributes in a consistent form so an agent can set it next to the alternatives. This only works if the underlying information is consistent: same fields, same units, conditions stated.
- Booking. Reserving a call, a demo or a slot. Bounded and reversible, as long as the confirmation is explicit.
- Quote request. Taking a structured enquiry. This works when the inputs can be defined. If a quote needs a scoping conversation, the tool can capture the brief and leave the pricing to a person.
- Account actions. Checking an order, downloading an invoice, changing a plan. Valuable and sensitive: these need authentication, permissions and confirmation before anything irreversible happens.
Two things should stay out of the action layer. Anything where the decision itself is the sale, like positioning, bespoke pricing or scoping a consultative project, needs a person on the buying side and a person on yours. And anything irreversible shouldn't run on an agent's judgement alone: a confirmation step costs seconds, while a wrongly cancelled contract costs a relationship.
What this does to the role of the website
If tasks can be completed without a visit, the website's job narrows and sharpens. It becomes the place where a person decides, not the place where a task gets done.
Picture the sequence. An agent shortlists three providers against a brief. A person looks at the shortlist, then visits the sites to answer what the agent can't: who are these people, what have they actually done, what would working with them be like. That visit is a trust and decision layer, and it rewards what the first two parts kept pointing to: named expertise, specific evidence, stated scope, pricing logic and honest limits.
Two things follow. A site that leans on decoration to look credible is exposed, because the human visit now comes after an agent has already filtered on facts. And measurement changes: a task completed through a tool call doesn't look like a session in your analytics. A quote request that arrives with no preceding pageviews may not be a tracking error. If you judge the site only by traffic and conversion paths, a working action layer can look like decline.
Ownership and platform dependency
Who benefits from an action layer depends on who owns it.
If your services, availability and prices live only inside someone else's platform, such as a marketplace, an aggregator or an AI vendor's agent directory, that party decides how you're described, ranked and compared. It can change the terms, the format or your visibility, and you find out afterwards. Being one row in someone else's comparison is a position, but not one you control.
Four things are worth owning, wherever they're exposed:
- The source of truth. One place where your services, scope, prices or ranges, conditions and availability are kept in structured form. Your website, any API and any third-party listing should all draw from it.
- The rules. What an agent may do alone, what needs confirmation, what's off limits.
- The identity layer. How a person or an agent proves who they are before touching an account.
- The customer relationship. Whatever route a request arrives by, the contact and the history should end up with you, not with the intermediary.
A practical test: if a platform disappeared or changed its terms tomorrow, could you rebuild your presence from your own data within a week? If not, you're renting your interface.
When it's too early to build one
For most B2B service firms, building a dedicated agent layer now is premature. Four things have to be true before it pays back.
First, you have transactional tasks with clear inputs and real volume. A firm whose sales process starts with a conversation has little to expose beyond a booking and an enquiry form, and a well-built form already covers both.
Second, buyers are already arriving through agents, and you've checked rather than assumed. Look at referral sources, at what prospects say about how they found you, at whether enquiries appear without the usual path.
Third, your data is already structured. An agent layer on top of inconsistent information just spreads the inconsistency faster.
Fourth, you can carry the risk. Authentication, permissions and confirmation steps are real work, and an action layer that lets something misfire on a customer's behalf is worse than none.
The standards are also unsettled. Protocols and browser proposals are still moving, and what you build for one may need rework for the next. Spending heavily on an interface whose shape isn't fixed is a poor use of a limited budget.
What to prepare now
You don't need an agent project to get ready for this. The useful preparation is cheap, and most of it is the work the first two parts already argued for.
- Put your offer into one structured source. Services, scope, prices or ranges, conditions, regions, availability. Keep it in your CMS or a single data source and render the website from it. When an agent layer becomes worthwhile, the hard part is already done.
- List the five tasks a buyer, or an agent working for them, would want to do. Compare providers, request a quote, book a call, check a status, find a document. For each, write down what it needs as input and what it returns, as if a stranger had to fill it in.
- Make the forms and bookings you have work as plain, structured HTML. Labelled fields, stated constraints, explicit confirmations. This is the cheapest bridge: a well-built form can later be exposed as a tool with little extra work, and a badly built one can't.
- Decide who owns what. Where the source of truth lives, what an agent may do unsupervised, and who approves changes.
- Strengthen the human layer. Named people, specific evidence, stated scope and honest limits. This is what a person checks after an agent has shortlisted you.
- Start looking. Add a line to your enquiry form asking how people found you, and check AI-assistant referrals in your analytics every month. Evidence of agent-driven traffic is the signal to move from preparing to building.
Where this leaves the website
For people deciding whom to trust, the website stays the place where the evidence lives. What changes is that it may no longer be the only way in, or even the main one, for the tasks around that decision.
Companies that treat the site as one interface among several, with their data, rules and customer relationships kept in their own hands, will adapt with little drama. Companies whose entire presence is a set of pages built for one kind of visitor will find out how much of it was ever about the page.
The website isn't going away. It just may not always be the primary interface.
Not sure which tasks on your site are worth exposing to machines, and which need to stay in front of a person? Our CRO, customer journey and digital experience audits map where your site loses attention, trust or conversion, and what to fix first.