Much gets written about AI agents: what they can do and whom they replace. Rarely does anyone write about what a company has to do before an agent can even start working. Aaron Levie, CEO of software company Box, recently listed that work and concluded that its volume will exceed anything we imagine today. He is talking about large corporations. The list applies just as well to a company with 80 employees.
From chat to agent
A chatbot answers questions. When it gets something wrong, you notice, because you read the answer. An agent does work: it prepares the quote, reconciles the order with inventory, answers the customer inquiry. When it gets something wrong, the result is already with the customer.
That difference is why the step from chat to agents is not a software update. A tool that works alongside you instead of just answering needs the same things a new employee needs: access to the right information, clear authority, an introduction to how things are done, and someone who reviews its work. None of that comes with the license.
Five work packages hardly anyone plans for
Levie's list boils down to five work packages. Each one is unglamorous. Together, they are the real effort of any agent rollout.
Data access. The context an agent needs rarely sits in one place. It is spread across the ERP, the inbox, the file share, a few spreadsheets, and the head of the person who has run the process for twelve years. Before an agent can work, that knowledge has to be reachable, securely and in order.
Permissions. A sales employee does not see payroll. The same applies to an agent, except nobody sets it up automatically. Which data may it read, which steps may it execute, and where can you trace what it did? Answering these questions before launch is far cheaper than answering them after the first incident.
Process knowledge. The biggest package. Agents work with what is documented, and in most companies, what is documented is how the process was designed, not how it runs. The exceptions, the special cases, the "for this customer we do it differently": exactly this knowledge decides whether an agent delivers usable work.
New workflows. Levie puts it in one sentence worth remembering: just replicating the old workflow will mute the gains. A process with an agent in it looks different from the same process without one. It takes a deliberate decision about who owns which part, where the human steps in, and how you measure whether the result is right.
Ongoing upkeep. Tools and methods in the agent space change month to month. You can swap a personal productivity tool in an afternoon. A business process with customers attached, you cannot. Someone has to stay on it permanently, not just until the project ends.
Who does this work?
Levie names four groups that will take this work on: large consulting firms, new specialized firms, the AI vendors' own deployment engineers, and new internal roles: the "agent engineer" as a job title.
For a corporation, those are four real options. For a Swiss company with 50 to 500 employees, the list shrinks: the vendors' deployment engineers serve large accounts, the big consulting firms only pay off at corporate scale, and a full-time internal hire for a job title that has existed for two years is hard to fill and hard to keep busy.
What remains are two paths: someone in-house builds the capability alongside their actual role, or an external partner takes on the work permanently, as part of operations rather than as a project. Both can work. What will not work is assigning the five packages to nobody and rolling out agents anyway.
What you can start this week
You do not need to buy an agent for this. The five work packages can be walked through on a single process: on paper, in an hour.
1 — Pick one processTake a workflow that eats time every week: quoting, order entry, reporting. Write down how it actually runs, including the exceptions. If that turns out to be hard, you have found work package three.
2 — Follow the informationFor each step, mark where the required information lives: a system, a file share, an inbox, or someone's head. Every stop in someone's head is a stop where any agent would fail today.
3 — Set the checkpointDecide at which point in this process a human would have to review the result before it leaves the building. You will need that answer for every automation, and it costs nothing.
The work is the actual rollout
The common picture of AI agents is a purchasing decision: pick a tool, buy a license, go. Levie's list shows why that picture does not hold. The tool is the smallest part. The rollout consists of data access, permissions, process knowledge, new workflows, and ongoing upkeep: work someone has to do before and after the agent runs.
That is not bad news. It just means: whoever takes this work seriously gains a lead that no tool purchase can catch up with. Anyone can buy the tools. The documented knowledge of your own operation, nobody else has.