Catering to the Agent.
When the buyer stops being a person, the website stops being a brochure.
Two patterns are running in parallel right now. LLM referral traffic into B2B sites is climbing fast. The CEOs of every major data platform are saying out loud that the next stack has to be rebuilt for software that serves agents. The two patterns meet on the same surface: the company website. This is a short note on what changes, and on the work the gap creates.
Chris Salazar, Innovative Group
The platforms have shipped. The arm that turns them into pipeline has not.
The world is moving from software that serves people to software that serves agents. The databases, the platforms, the entire stack has to be rebuilt for that.
Every quote like this travels through three audiences before it reaches the people who have to do something about it. The CEO reads it and asks the CTO. The CTO names a platform. Marketing finds out the budget got reorganised. By the time the implication lands on the GTM team, the headline has moved on. The implication itself is huge. The website your buyers visit has been built for human eyes scanning a page. The next decade of buyers is software synthesising a thousand pages in a single query, then deciding for the human whether to click at all.
The numbers behind the shift.
Three signals from Databricks' 2026 State of AI Agents tell the story without much commentary.
The first number is demand. The second is the gap between intent and delivery. The third is the structural bet the platforms are placing on what comes next. Read together, the story is that buyers are queueing for this work and the work is hard to ship. That is where the partner economy ought to grow up around the platforms, and where most of it has not.
Agents read differently.
An agent does not browse. It queries. It does not scan a page for the headline that catches its attention. It synthesises the page against a goal, extracts the structured data, quotes the source, and moves on. The pages that win attention from a human visitor are pages that catch the eye in the first three seconds. The pages that win attention from an agent are pages that answer the question cleanly enough to be trusted as a citation.
Two things flow from that. First, the brochure-style website starts losing the visitors who matter most: the agent acting on behalf of a senior buyer. Second, a website that wants to win the agent visitor needs three things the typical B2B site does not have. Real-time signal capture so the page can adapt to the visitor's intent. A clean information taxonomy so the agent finds what it came for. A response surface that decides what to recommend in the same session as the visit, not the next one.
Why this lands on the website first.
Inside an enterprise, AI is a board topic. It gets debated for nine months and ships in fragments. The website is where the debate ends and reality starts, because the website is where every constituency meets the same way: as a visitor with intent. The agent layer that wins the next decade of pipeline is the one that turns the website into a real-time response surface, governed at the data layer, fed from the same signals the marketing automation already collects.
Most enterprises will discover this the slow way. They will run a chatbot pilot, find the chatbot embarrassing, blame the model, and pull back. The actual problem was never the model. The actual problem was that the chatbot was bolted onto a static site with no live signal underneath it. When the agent has nothing to read but the brochure, the agent makes things up. When the agent has the visitor's intent, the company's content classifier, and a real-time graph of what the visitor has touched in the last five minutes, the agent answers the question the visitor came to ask.
Next-Best Action is the wedge.
Inside a marketing organisation, the agent era arrives one workflow at a time. Of all the workflows where it can land, Next-Best Action is the one that pays for itself fastest and proves the case loudest. It is measurable. It is revenue-attached. It uses the data the marketing org already collects. The architecture underneath it (signals to intelligence to agents to activation) is the same architecture every other agent use case will need over the next two years. Whoever builds the NBA layer cleanly is also the team that ends up building the chatbot, the brief-to-assets agent, the recommendation surface, and the conversational analytics on top.
That is the wedge. Not the only place to start, but the place where the buyer is already looking and the platform layer is mature enough to support production.
What we refuse to ship.
A short list of patterns we have stopped agreeing to. Most of these stem from a model where the agency does the slide deck and the platform vendor does the install. That model is breaking under the weight of the agent era.
- A pilot that bills by the query without a forecast.The pilot fails its CFO review at month two. Consumption pricing without FinOps guardrails is not a pilot. It is an experiment that will not survive its first invoice.
- A chatbot bolted onto a static site.Every one of these embarrasses the brand inside three months. Real-time signal underneath, or no chatbot at all.
- A six-month custom build to replace something already in the stack.If your CDP, your CRM, and your event stream already capture the signal, the work is to extend them. Replacing them is a 2022 motion.
- An NBA program that requires a data team rebuild.The data team has work. The work is to ship the agent layer on top of the model that is already running, not to renegotiate the data charter every quarter.
- A reference architecture without a delivery pod attached.The category is full of decks. The shortage is teams that can ship.
Co-Founder of Innovative Group. Operator background across go-to-market, growth, and AI delivery, with the team that has shipped end-to-end AI marketing platforms inside enterprise stacks. Writes about the shift from experiment to operation, and the work it creates.
The platforms have shipped.
The arm that turns them into pipeline has not.