For two years the conversation about AI agents revolved around developers. That has changed. In August, OpenAI published data from its enterprise customer base: since February, weekly use of its Codex agent grew 108 times in legal teams, 41 times in sales and 41 times in recruiting. In engineering, over the same period, it grew 5 times.
Another figure from the same report: companies that get the most out of AI (the top 10% by usage) now generate 8.3 times more tokens per user than the median. In January that gap was 2.6. It opened in five months, and newer models don't explain it. Everyone has access to the same models.
What sets the leaders apart
According to the report, these companies connect agents to three things: company context (documents, data, processes), real tools (systems where the agent can query and act), and clear permissions with human review.
An agent with no connection to your systems answers with generic knowledge. That has little value for your own work. One connected to a client's file, the catalogue or the CRM can draft an analysis, review a contract against your terms, or prepare an account report. That's the jump from "assistant that helps you think" to "agent that completes the work".
The uncomfortable part is that this isn't a purchase. It's integration work: decide what the agent can see, what it can do, where every action gets logged, and who reviews the result. Companies that do this work early compound the advantage, and those who wait don't make it up with a new license.
Why it starts in legal and sales, not in IT
These teams work with documents, repeatable processes and clear review criteria. That mix matches what an agent does well: long, delegable tasks with a person validating the output.
A legal team delegates the first pass of a contract review against a template. Sales asks the agent to prepare the prospect's dossier, pulling from the CRM and recent interactions. A recruiting team screens candidates against the role's criteria before anyone reads the first CV. In none of these cases does the agent decide: it prepares the work so a person decides with less effort.
If your company wants to start, the order that works is:
- Pick a repeatable process with a checkable definition of "correct".
- Connect the agent to that process's real data, with minimum read permissions.
- Define the human review step and the action log up front.
Three steps, a two or three week pilot, and you decide with results whether to scale it.
At Luxion we build this kind of integration: agents connected to your real data and processes, with control where it belongs. If you want to assess where this would fit in your company, we start with a small, measurable prototype before you commit to anything.