Talk to enough operators right now and you hear the same request, almost word for word. The CEO wants an "AI person." So does sales. So does marketing. So does finance. Every function in the company has figured out that AI can change how it works — and every one of them has arrived at the same conclusion about what's missing. Not a better model. Not another license. A person: someone who knows how to deploy agents, who will sit with the team, learn how the work actually gets done, and tailor AI to it.
Engineering teams are moving fastest. They live closest to the tools, and agentic coding has already rewired how many of them work. But the rest of the company isn't waiting politely behind them. Sales, finance, marketing, operations, and the C-suite itself now need agent deployment expertise embedded in their own teams — and it's not enough to put in a ticket and wait until an engineer frees up. By the time one does, the team has either lost momentum or built something brittle on its own.
This is happening at hundreds of companies at once, whether or not they've named it.
I want to make an argument about what that request really means: that frontier diffusion — the capabilities of frontier AI actually spreading through an organization and changing how work gets done — is fundamentally a people and talent problem, and always has been. Then I want to talk about the model we've built at AgentGraph to solve it, because I think the default answers on the market get the economics wrong.
The models are ready. The organization isn't.
There's a pattern economists love to point to about electrification. Factories got access to electric motors in the late nineteenth century, and for decades productivity barely moved. The motors weren't the problem. The factories were still laid out around a single steam-driven line shaft, and most owners simply swapped the steam engine for an electric one and changed nothing else. The gains only arrived once a generation of engineers redesigned the floor — a motor on every machine, workflows rebuilt around the new capability.
We're in the line-shaft phase of AI. The capability is sitting in every company already: frontier models, agentic harnesses, skills, MCP servers, connectors into the systems of record. Licenses have been rolled out. Pilots have been run. And yet in most organizations, the way the finance close works, the way pipeline gets built, the way a campaign gets shipped, looks roughly like it did two years ago.
The capability isn't the constraint. The constraint is the people who know how to rewire the floor.
Diffusion happens at the function level, one use case at a time
AI doesn't diffuse through an enterprise as a single rollout. It diffuses the way the electric motor did — machine by machine, workflow by workflow. A collections workflow in finance. Account research in sales. Content production in marketing. Ticket triage in support. Each one is its own small re-architecture, and each one needs someone who understands both the agentic primitives and the way that particular team works.
That combination is rare. The people who can decompose a real workflow, encode it as skills, wire the tools into the right systems, put evals around it, and get the result into the hands of the team that uses it — those people are among the scarcest talent in the economy. And the demand for them isn't coming from one place in the org chart. It's coming from every place in the org chart, simultaneously.
That's the shape of the problem. It's not "we need an AI strategy." It's "we need an AI person embedded in every function, available when the team is ready to move — and we don't have nearly enough of them."
The people deserve more of the credit than the AI
There's a reason that person is so hard to find: real expertise in building and running agents takes a long time to earn. It doesn't come from a certification or a weekend of tutorials. It comes from reps — from shipping agents into messy environments, watching them fail in ways nobody predicted, and building the instincts for what to fix. Many of the builders on AgentGraph have each run more than a billion tokens through agents over the past couple of years. That kind of mileage shows up in the work. They know which problems an agent should own and which it shouldn't, where a workflow needs a human checkpoint, how to scope a use case so it actually ships, and when a demo is lying to them.
Here's what I've seen across a lot of deployments: the AI projects that do really well, do well because of the people behind them — not because of the model that was used to help build them. Two teams can have access to the exact same frontier model, the same harness, the same tools. One ships an agent that changes how a function works. The other ships a pilot that quietly gets abandoned. The difference is almost never the AI. It's the judgment of the people who architected the solution and steered it toward a business outcome.
That line is getting blurred, and I think it's worth saying plainly. As the tools get more capable, we're giving AI a lot of the credit for outcomes that were designed, scoped, and steered by AI engineers. The model is the motor. Someone still has to redesign the floor — and that someone is the reason the project worked.
Every company needs an internal forward-deployed engineering org
I've written before about why building a forward-deployed engineering org is the next top priority for nearly every company. The short version: the forward-deployed engineer — the embedded, solution-oriented technologist Palantir invented to deliver value through data — is turning inward. Companies are standing up their own FDE capability to deploy AI into their own operations, the same way the last generation stood up analytics centers of excellence.
The "AI person" every team is asking for is a forward-deployed engineer. The request is the demand signal for that org.
But here's the practical problem. Most companies trying to build that capability have one or two people in a pod, and those people are drowning. They're being pulled into the CRO's priorities on Monday, the CFO's on Tuesday, and an executive offsite on Wednesday. They can't be embedded everywhere at once, and they can't be expert in everything. A great agentic engineer who has never lived inside a revenue operations team will spend weeks learning what a sales ops veteran already knows. Hiring your way out takes a year you don't have, at salaries that make the math hard to justify for every function.
So companies reach for the obvious alternative.
The agency model gets the incentives backwards
The default answer is to hire a large deployment firm or agency. And to be fair, those firms can bring real talent. But the economic model works against you.
An agency engagement is pushed into your company. It starts with a large scope, a statement of work sized to the firm's staffing model, and a team whose success is measured by the size and length of the relationship. The incentive is to expand — more workstreams, more seats, a bigger transformation program — whether or not your organization is ready to absorb it. You end up paying for capacity ahead of demand, and worse, you end up dependent: the knowledge of how your AI systems work walks out the door with the consultants, and the only way to keep going is to renew.
That's the opposite of diffusion. Diffusion is your teams learning to work differently. Dependency is someone else doing the work for you, indefinitely, on their timeline.
AgentGraph: an AI person for every team, pulled in when you need one
This is the problem we built AgentGraph to solve. If AI diffusion is a talent problem, the answer is to make the right talent available to every team — without forcing you to hire it all or rent it all from one firm.
It comes down to three pieces.
A network of elite builders across the whole range of the work. AgentGraph is a network of vetted, AI-native experts, and it's deliberately broad. Some are strategic — the people who help leadership decide where AI should go first and how to stand up an internal FDE capability. Some focus on training and education, because diffusion only sticks when the people doing the work learn to work with agents themselves. And many are function-level builders: builders with go-to-market expertise who know what a pipeline motion actually looks like, builders with finance expertise who understand the close, reconciliation, and FP&A, and builders across operations, support, product, and engineering. Your sales team gets someone who speaks sales. Your finance team gets someone who speaks finance. That domain fluency — paired with hard-won mileage building and running agents — is the difference between an agent that demos well and one that a team actually adopts.
A context engine that maps the organization. Embedding talent only works if the talent has context. AgentGraph's context engine builds a working map of your company — teams, systems, workflows, and where the friction lives — so you can see the landscape of use cases and block and tackle them over time instead of chasing whichever one is loudest this week. It's also what lets a builder show up on day one already oriented, rather than spending the first month rediscovering what the last one learned. The map compounds: every engagement makes the next one faster.
A credit system that lets you pull AI in, rather than having it pushed on you. Instead of signing a sprawling statement of work, you make one commitment to one entity and build up credits. Then you spend them where the demand actually is: a week of a finance builder for the close, a training session for the marketing team, a go-to-market engineer to stand up an account research agent, a strategic session to prioritize next quarter's use cases. Teams draw on expertise as they're ready for it. When a use case is done, you stop spending on it.
That last point is the economic heart of it. A credit model is a pull model. Demand originates with the teams doing the work, and capacity follows it — not the other way around. You get the procurement simplicity of one relationship, the breadth of a whole network, and spending that tracks real adoption instead of a vendor's growth plan.
Pull beats push
Go back to the factory floor. The electric motor got the headlines, but it was the engineers who redesigned the floor that delivered the productivity. And electrification didn't succeed because one contractor came in and rewired everything at once. It succeeded as a generation of those engineers spread out across the floor and rebuilt it machine by machine, close to the work, as each part of the operation was ready.
That's what AI diffusion looks like too. Not a single transformation program pushed from the top, but expertise pulled in by every function, one use case at a time, by teams who then own what they built.
AI will keep getting better, and it will keep getting the credit. But the companies that actually diffuse it will be the ones that invest in the people who know how to deploy it. Every team in your company wants an AI person. The companies that win the next few years will be the ones that find a way to say yes to all of them — without hiring an army or signing their future over to an agency.
AgentGraph gives every team in your company access to an AI person: a network of elite builders across strategy, training, and every business function, a context engine that maps where AI should go next, and a credit model that lets you pull expertise in as you're ready. If you're figuring out how to diffuse AI across your organization, let's talk.