Key Takeaways

  • Only 5% of non-engineering enterprise AI agents make it into production, stalled by fragmented human workflows rather than model limitations.
  • CJ Hess warns that developers cannot automate a process that does not exist; if 40 workers run one role in 40 different ways, an agent cannot learn what correct behavior looks like.
  • Without operational standardization, engineering teams cannot build evaluation datasets or verify agent output quality.
  • Alex Lieberman notes that role-based access control creates a severe security bottleneck, forcing Tenex to build custom gateway infrastructure to govern enterprise data reads and writes.

Standardizing Processes Comes Before Building Agents

Software engineers love deploying autonomous agents to write pull requests, test code, and review diffs. That environment has rigid boundaries: code compiles or fails, test suites pass or break, and version control maintains an exact record of every change. Yet step outside the engineering department, and agent deployments hit an immediate brick wall. Only about 5% of non-engineering agents ever reach production inside large companies.

Founders usually blame the underlying models, arguing that context windows are too tight, hallucination rates are too high, or inference costs are too steep. CJ Hess and Alex Lieberman, co-founders of the engineering consultancy Tenex, argue that the real blocker has nothing to do with frontier models. It comes down to human inconsistency.

Operations, sales, and support departments do not run on strict logic. Most internal enterprise workflows are tribal knowledge wrapped in personal habit.

As Hess puts it: “I phrase it to the team as like you can't automate a process that does not exist.”

When Tenex audits corporate workflows, they frequently discover teams where dozens of staff handle identical roles through completely different habits. “When you have 40 people that are all roughly doing the same function, but they're all doing it a little bit of a different way, there's actually it's it's way harder to build an agent that's kind of proactively doing that in the background,” Hess explains.

If human operators execute a task in forty conflicting variations, you cannot build an evaluation benchmark. You cannot define ground truth. If the system has no definition of correct output, the agent cannot determine whether its actions succeeded or failed. Tenex now refuses to build autonomous agents until the client standardizes the underlying operation. Hess notes: “First we need to build a process around this organization within your company. And then here are the steps for us to actually hand off that work to an agent in the background.”

Missing RBAC Infrastructure Blocks Enterprise Deployment

Even when a company standardizes an operational workflow, deployment hits an administrative wall: security and permissioning.

Employees work with models like Claude and Codex across internal platforms. But traditional corporate software was never architected to expose role-based access control (RBAC) to automated agents reading and writing data across multiple internal silos.

Lieberman points out that this gap creates chaos for IT leadership: “Permissioning and uh like arbback basically as every employee in a company has clawed or they have codecs and they need access to certain data it is turned into an absolute nightmare for CIOS and CISOs and they don't have basically a good gateway for easily doing almost like a like a control panel for uh data permissioning.”

If an agent holds open read access to an enterprise database to answer routine queries, it risks exposing payroll records, executive discussions, or sensitive customer records to unauthorized staff. To unblock deployments, Tenex started building custom data gateways for enterprise clients, eventually productizing the layer as standalone infrastructure to govern model read and write access.

What to Do With This

Before writing agent code for any internal team this week, audit the manual process across five people who share the same job title. If their spreadsheets, ticket tags, and communication steps differ, pause software development and document an explicit standard operating procedure first. Once the human workflow produces identical steps, map the exact read and write permissions each step requires before opening model access to your internal data.