You paid for the seats and ran the training, but the organization still works the way it did before. Adding AI tools was never going to change how it works, and one more set of tools will not either.
A company with all the tools
Take a made-up software company where the seats and the training did everything they were bought to do. Every person uses AI all day, and every one of them is very good at it.
Now follow one customer request through it. The account manager writes up the request with AI. The product manager turns it into a spec with AI. The developers implement it with a coding agent. The tester writes the test cases with AI. The account manager drafts the update to the client with AI.
Five steps, in the same order as before. The same people hand the work to the same people, and it waits in the same places for someone to pick it up. Each of them reads and writes their part faster. That is all that changed.
This company has no tool left to buy and no training left to run. And it does not work agentically. It works the old way, just read and written by AI assistants.
How work moves between people
What an organization gets done is set by how the work moves between people: where it is handed over, and where it waits. A tool in one person's hands does not reach that. It changes that one person's own task.
Three things follow.
First, AI does not make everything better. It amplifies what is already there: whoever is good gets more done, and whoever struggles gets deeper into the same trouble. In Google's DORA research, AI "magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones."
Second, speeding up one person moves the bottleneck to the next. Think of it this way: put a faster machine at one step of an assembly line, and the pile in front of the next step grows. Faros AI saw this in its own telemetry from more than 10,000 developers: high-AI teams merged 98% more pull requests while review time rose 91%.
Software development shows it. The process from account manager to product manager to developers to tester and back to the client was paced by coding. When AI made coding fast, the bottleneck moved out to every other step. It did not go away.
Third, even when no work piles up anywhere, making one step faster barely moves the total time. A real process has tens of steps. Say it has twenty and each takes an hour: make one of them twice as fast with AI, and the work takes nineteen and a half hours instead of twenty. 100% gain only bought you 5% in reality.
The handoffs nobody owns
Each person can fit AI only into their own tasks, because their own task is all they control. The developer cannot redesign how requirements reach them. The tester cannot decide that testing should happen somewhere else in the process. So everyone improves their own step, and the handoffs between the steps stay untouched.
The numbers that get reported on AI tools miss the handoffs too. Seats, usage, people trained, agents built: every one of them counts something that happens inside a task. None of them says anything about how the work moves.
My own process, redesigned
In my own custom product development, each piece of work starts at issue refinement. When an issue is created, an agent refines it, and when it needs more information it has to ask someone.
That someone is usually me, including when the actual answer sat in the customer organization. In the worst case the loop went like this: the agent asks me, I email someone at the customer company, they work out the right decision, they send it back, and I feed it to the agent. I was the messenger between an agent and a customer. The agent took few minutes to realize it needs more information and it took me a day to get the necessary details back to it.
So I redesigned that process. Users at the customer organization now file requests and answer the agent's questions directly inside their application.
The overhead dropped drastically. The reason is the process: the question now goes to directly to the person who has the answer, and nobody carries it there by hand.
Now picture if we would try to supercharge that process with an AI assistant at every step. I write my email to the customer with AI. They write their reply with AI. The same messages go round the same loop, written faster. Adding ChatGPT at every step of the old process does not produce the new process. Only rethinking the whole process does. And it gets even worse if we look the request coming all the way from the customer in the first place. Nothing really changed.
The process is the key
When your organization wants to work agentically, the place to look are the processes as a whole: how a piece of work gets from the first person to the last. That is where the bigger changes are found. What each person has in their hands at any given time says very little about it. Pushing more tools onto people in the hopes of making this faster leaves the process exactly as it was. As inefficient as ever.
