Let's Use AI for More than Paving Cowpaths

Vas Moza, the CEO at Varick Agents, posted an article that is worth a read if you're involved in enterprise AI transformation efforts. He argues that many AI initiatives have produced limited ROI for the same reason earlier technology transformations stalled… Companies use new technology to accelerate existing workflows rather than rethink them. Michael Hammer called this “paving the cow paths,” modernizing an existing circuitous route (instead of using new technology to establish a more efficient way to reach the destination).

Enterprise workflows are often not limited by the speed of executing any one step, but by the elapsed time between steps, which commonly relies on human input and responsiveness. That’s true in media operations. Over the past thirty years, media publishers have automated file transfer, ingest, metadata extraction, QC, segmentation, localization, compliance, transcoding, packaging and delivery. Yet a surprising amount of manual work and delay remains because the difficult part is often not the processing of any of these steps, but the handoffs between steps and deciding what should happen when something falls outside the expected path. 

In media processing, exceptions are not really exceptions so much as a steady state. A file may arrive without the metadata needed to process it. It may be the wrong edit altogether, requiring someone to find the correct version and retrieve it. A QC result may need to be interpreted against requirements of a specific production or distribution partner. Sidecars may need to be checked for alignment and conformance against an edit. Delivery may identify an issue with a title, far down the line. In these cases, the underlying systems usually work as designed. The delay comes from waiting for operators to assess the situation and assemble enough context to determine what should happen next.

That work happens across ticketing systems, MAMs, rights systems, bespoke platforms, email, Slack, spreadsheets… and the efforts of experienced operators. It’s definitely not efficient. Someone checks the MAM and metadata, someone else looks at QC reports, someone from a third team reviews the localization, another person checks the delivery specification, and then the distribution partner processes the asset. A meaningful share of the cost in media operations sits in these handoffs and the numerous exceptions at each juncture. And many of the decisions depend on operator judgment that is not clearly documented anywhere. (Though I am using media processing as the example here, the same pattern appears in workflows across most corporate functions.)

This is where AI can be more useful than simply accelerating individual steps in a process. The opportunity is not to make deterministic systems faster, but to streamline the human judgment and coordination required between them. AI can assemble the context required for a decision and learn from prior actions and outcomes across the operation.

An agent can gather information from several systems, understand the context surrounding an asset, invoke the appropriate tool and interpret the result. When a decision should remain with a person, an agent can assemble the evidence and present the issue in a form that allows the operator to decide quickly. As confidence builds in the agent's ability to determine the correct path, some of those approval steps can be reduced or removed, gradually straightening the cow path and unlocking real gains in a given workflow.

Media pipelines have the noted complication that exceptions are more steady state than you’d think. The work is not actually one standardized workflow, but many dynamic ones. Processing requirements legitimately vary by asset type, customer, territory, language, platform and rights position. Trying to capture every combination in a single workflow can simply create another kind of cow path.

A better approach is to give an agentic operating layer enough context to determine which path applies to the asset in front of it. Automated context assembly, combined with agents that can act on that context and learn from the results, is the opportunity, and that’s exactly what we’re building Mediafier to unlock. (I’ll explain a lot more about what we’re doing soon.)

There is a need for a context management layer that sits across the existing media technology stack and makes those systems available to agents through a governed operating layer. The goal is not to replace the infrastructure media companies already rely on, but to enable those systems to work together with the context, permissions and evidence required to coordinate agentic processing reliably.

For years, media automation has meant defining a workflow and automating each step as far as possible. That works well when the path is known in advance, but it can also leave us paving every bend and detour that developed over years of operating around system limitations. Agents give us a chance to bypass some of those turns altogether, using context to determine a more direct route through the work.

That is where much of the manual work still lives, and where the larger gains will be won.