Every month a team somewhere buys an agent. It answers questions, drafts emails, qualifies leads, writes the first version of a report. The demo is real and the first week is exciting. Three months later the function runs exactly as it did before, plus one more tool nobody is sure who owns.
The agent did not fail. It was asked to do a job that no single component can do: change how a function runs. That job belongs to infrastructure, and an agent is not infrastructure. It is the part of the infrastructure you can see.
The pattern is measured, not anecdotal. MIT’s NANDA initiative studied 300 enterprise deployments in 2025 and found that 95% of generative AI pilots produced no measurable effect on profit and loss. The cause the researchers named was not model quality. It was a learning gap: tools that never adapted to the workflow they were dropped into.
The Visible Part
An agent is the thing that talks. It sits in a chat window or an inbox, it takes an instruction, it produces work. Because it is the part people interact with, it becomes the part people evaluate. Is it accurate? Is it fast? Does it sound like us?
Those are fair questions about a component. They are the wrong questions about a function. A sales function is not a set of emails. It is a sequence of decisions about who to talk to, what to say, when to stop, and who picks up when the conversation turns into a deal. An agent that writes a better email inside an unchanged sequence produces better emails inside an unchanged sequence. The sequence is where the time goes.
This is why so many agent deployments produce a number that sounds good and a function that feels the same. Forty percent faster at a task nobody redesigned is still the old function, with one task compressed.
What Infrastructure Means Here
Infrastructure is the part that does not talk. Four things, in our experience, and a function only changes when all four are in place.
The workflow, redesigned. Not the old process with an agent inserted at step three. The work redrawn from the outcome backwards: what the function must produce, which decisions are genuinely human, which are rules that were only ever human because nobody wrote them down, and where the handoffs sit. The agent gets a place in that design. It does not get to define it.
The data it runs on. An agent is only as good as what it can see. If the customer history lives in three tools and a spreadsheet, the agent guesses. If nobody has decided which record is the truth, the agent picks one, confidently. Most of the disappointing agent stories we hear are data stories with an agent standing in front of them.
The rules it follows. Every function runs on judgment that its senior people carry in their heads: what we never promise, how we qualify, which claims we make and which we do not, what “on brand” means when there is no template. Left in heads, this judgment cannot be given to an agent, and it leaves when the person does. Written down as rules the agent reads and a check enforces, it becomes the most durable asset the function owns.
The handoffs to people. The moment the agent stops and a person starts is a design decision, not an accident. Who is told, what they see, what they are expected to do, and how the result gets back into the system. Functions that skip this get the worst of both: agents that go quiet at the edge of their brief, and people who do not know the work has arrived.
Put those four together and the agent becomes what it should have been from the start: the fastest worker on a well-run team. Leave them out and it is a very fast worker on a team nobody has organised.
Expertise Is the Thing Being Installed
Here is the reframe that changed how we describe our own work.
What a good consultant brings to a function is not hours. It is expertise: knowing what a working sales motion looks like, what a brand system has to contain, where operations leak. For decades that expertise was delivered as advice, in decks and workshops, and it decayed the moment the consultant left, because it lived in people’s memories and a binder.
Agentic AI changes what expertise can be delivered as. The knowledge of how a function should run can now be written down as workflow, data structure, rules and handoffs, and installed so that the function runs that way whether or not the expert is in the room. The agents are the moving parts. The expertise is the infrastructure they move within.
That is a different product from advice, and it is a different product from software. Software gives you a tool and leaves you to decide how to use it. Advice tells you how to use it and leaves you to build it. Expertise as infrastructure does the deciding and the building, then hands you the keys.
It is also why we stopped describing what we do as building agents. We build the infrastructure and install it one function at a time: go-to-market, sales, brand and content, operations. The Catalyst products are that infrastructure packaged for a domain, with agents inside. Our own website is run this way: the brand is a rule store, the rules are checked before every release, and the agent that produces the pages could be replaced tomorrow without the brand moving an inch. The agent is the least important part of the system, and the part people notice first.
Why the Order Matters
The instinct is to start with the agent, because the agent is the exciting part and the infrastructure sounds like a project. So the sequence becomes: buy the agent, see what it can do, fix what breaks.
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, for escalating costs, unclear business value or inadequate risk controls. Model capability is not on the list. McKinsey’s 2025 global survey found that 88% of organisations now use AI somewhere, and 39% see any effect on earnings; the small group that does see it was 2.8 times more likely to have fundamentally redesigned its workflows first.
What breaks is everything the agent stands on. The data turns out to be scattered, so the agent gets a data project bolted on. The rules turn out to be unwritten, so the agent gets a compliance review bolted on. The handoffs turn out to be undefined, so the agent gets a process document bolted on. Six months in, the team has built the infrastructure anyway, backwards, under pressure, around a tool chosen before anyone knew what it needed.
Start with the workflow and the rest follows in order. The redesign tells you what data the function actually needs, which is always less than it has. It tells you which rules have to be written, which is always more than anyone expected. It tells you where the handoffs are, which is where the agent’s brief ends. Only then does the question “which agent?” have an answer, and by then it is a small question.
What This Means If You Lead a Function
The good news in all this is that you do not need the whole company to move. Infrastructure is built one function at a time, and a function is a manageable decision: one accountable leader, one redesignable workflow, and no committee between the decision and the change.
That is true whether the function is the whole commercial side of a growing company or a marketing team inside a large one. The scale of the parent does not matter. What matters is whether one person can say yes to redrawing the work, and then hold the team to the new design while the infrastructure settles in.
If you are that person, the question to ask before any agent conversation is not “what can it do?”. It is: if this agent disappeared next quarter, what would still be different about how my function runs? If the honest answer is nothing, you are not buying infrastructure. You are renting a fast worker for a team you have not yet organised.
The Question Worth Asking
Agents are not the wrong purchase. They are the wrong first purchase, and the wrong thing to measure a transformation by. The functions that actually change are the ones where somebody did the unglamorous part first: redrew the work, decided what the truth was in the data, wrote the rules down, and designed the moment where the machine stops and a person begins.
Everything else is a demo that went well.
So, for the function you lead: what expertise does it run on today that exists only in people’s heads, and what would it take to make that the infrastructure instead?
If you want to test the thinking on your own function before talking to anyone, the Work Redesign Audit is a self-serve exercise: run it on one workflow and see what the redesign step actually asks.
FAQ
Frequently asked questions
Is this an argument against AI agents?
No. It is an argument about order. Agents are the fastest workers a function has ever had. They change how a function runs only when the workflow, the data, the rules and the handoffs around them have been designed for it. Without that, they speed up one task inside an unchanged function.
What is the difference between this and AI transformation?
AI transformation is the outcome a company budgets for. Installing expertise as infrastructure, one function at a time, is how the outcome is reached. A transformation that buys agents and skips the infrastructure changes the tooling and leaves the function as it was.
Does a function inside a large company need the whole company to move first?
No. A function with one accountable leader has the same properties as a growing company: one person can say yes, the workflow can be redrawn, and there is no committee between the decision and the change. The infrastructure is built inside the function’s boundary.
How long does the infrastructure take to build?
For one function, weeks rather than quarters: the workflow redesign is days of work with the team, the rules are written as they are found, and the first agent goes in once the handoffs are drawn. The first assets through are slower than the old process. The tenth are not.
Who owns it afterwards?
The team. The point of infrastructure is that it keeps running when the people who built it leave. The workflow, the data model, the rule store and the handoff design are documented in the team’s own tools, and the agents inside them can be replaced without rebuilding anything around them.