Coding agents are delivering real value, while knowledge work agents keep stalling. The difference usually isn't the model you picked. It's everything underneath it.
In a recent CXOTalk interview, Box CEO Aaron Levie laid out the core reason that coding agents have been useful ever since their introduction: coding has several properties that make it ideal for enterprise AI agents. These properties are:
Knowledge work is messier. Users are less technical, outputs such as contracts or analysis reports are harder to verify automatically, and the relevant data is scattered across various sources such as SharePoint, CRM systems, PDFs, spreadsheets, and email.
Levie gives an example, featuring Salesforce, that shows where agents are already useful. He connected Salesforce's MCP server to Claude Code and now runs customer and market intelligence queries he wouldn't have pulled up manually. The agent removes friction from a structured system of record. That pattern of agents augmenting access to well-organized enterprise data is an area where value is accruing right now.
Levie coined “AI psychosis” to describe executives who only see the happy path of AI demos. A CEO watches an agent generate a contract in seconds and assumes the hard work is done. What they don't see is the work that comes after: verifying the terms, pulling in past contracts for context, routing for approval, and handling exceptions.
That distance from frontline work creates a strategic blind spot. Leaders may fund ambitious agent roadmaps before the organization has the data access, review process, and measurement infrastructure to make those agents reliable.
Levie told Fast Company that years of data management fragmentation have become a real issue for enterprises trying to adopt agents. An agent is only as useful as the data and actions it can securely access.
The enterprises pulling ahead have done the unsexy work first: getting content and data into a state where agents can operate with the right security, permissioning, and compliance constraints built in. The model, and how it’s prompted, are usually the last step to optimize, not the first.
In our work with clients, the CIOs who get the most from enterprise AI agents invest in data governance and measurement infrastructure before scaling deployment. They want to know where agents are being used, what they cost, and whether they create business value.
The era of flat-rate AI pricing is ending. Built In reports that AI providers are moving toward usage-based billing, where every prompt, iteration, and multi-agent workflow carries a direct cost. GitHub Copilot has moved more plans toward usage-based billing. The same cost pressure shows up at scale: Built In also reported that Uber exhausted its entire 2026 AI tools budget within the first four months of the year, after deploying Claude Code to roughly 5,000 engineers.
That changes the ownership question. Levie noted that AI spend probably won't make sense as only an IT budget anymore. Finance, marketing, engineering, and operations teams using agents will need visibility into what those agents cost and what business value they create.
Before scaling enterprise AI agents, CIOs should answer three questions to determine whether conditions are right:
If the answer to any of these is no, agents will underperform and costs will outpace returns. Fix the underlying conditions first.
Larridin measures what enterprise AI agents are doing across your organization: which workflows they run, how many tokens they consume, and which teams generate value, versus just burning budget. Our AI Measurement Framework tracks utilization, proficiency, and business value across human and agentic work. It gives CIOs and CFOs the data to make confident scaling decisions.
If you're moving from agent pilots to enterprise deployment, Larridin gives you the visibility layer for AI measurement and optimization before usage sprawls across the business and spending rockets upward.
Enterprise AI agents are autonomous AI systems deployed inside organizations to complete multi-step tasks with limited human intervention. They can query databases, draft documents, trigger workflows, and connect to enterprise tools through protocols like MCP.
Coding has three agent-friendly properties: the work is text-based, the output is verifiable through automated tests, and the data is centralized in code repositories. Knowledge work usually lacks at least two of those conditions. Outputs such as contracts or reports are hard to verify automatically, and data is scattered across disconnected systems.
MCP is an open standard that lets AI applications connect to external systems, including data sources, tools, and workflows. Levie's Salesforce example shows why that matters: agents can surface intelligence from systems of record at a scale no human would query manually.
Measuring agent ROI requires three things: utilization, cost, and business outcomes. Without all three, you have either a usage report with no value story or a value claim with no cost accountability.
Ready to measure what your enterprise AI agents are actually delivering?
Book a Discovery Call and get full visibility into agent spend and value in days.