Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

Why AI harnesses are essential to ROI

Companies are buying AI tools faster than they can manage them.

[Photo: Getty Images]

Profile photo of Sumeet Arora

Chief Product Officer · Teradata

Cost and the value derived have become the first questions in every AI conversation. Businesses are spending more on AI than ever, and many are still struggling to show a return on investment.

I recently spoke with a company that spent $10 million on AI so far but is failing to show clear ROI from those investments. Another told me that the bill for their AI coding tool jumped tenfold in a year. And of course, much attention has been paid to the Axios report about a company that reportedly spent half a billion dollars on Claude licenses for employees in a single month.

The common thread is clear: companies are acquiring AI tools faster than they can manage them, spending more than they planned, and producing less value than they promised.

The reason companies are in this scenario is because many have added AI tools one at a time. And now, many companies own a host of disconnected models, agents, and interfaces, but lack the crucial layer that’s needed to steer these models effectively. As models proliferate rapidly across a wide range of prices and capabilities, routing and governing them becomes even more difficult. For many companies, every model, agent, and AI application sits on its own island, with its own interface, costs, and risks. What companies really need is a way to manage and guide the models.

Enter the agent harness, the software layer around a large language model that lets it act on tasks rather than just respond to prompts. It runs the loop of calling the model, executing the tools it requests, and feeding results back, while managing context, memory, permissions, and error handling. It’s a term you’ll be hearing increasingly in enterprise AI conversations. 

Power needs a harness

Like a racehorse, AI has plenty of raw power and potential, but it takes a skillful rider to turn that power into a win. Similarly, your AI models need the right harness to connect them and ensure they are working most efficiently.

Most businesses own a stable of expensive AI models, but they haven’t invested in the layer that will guide the models.The agentic harness is what provides the models with context, routes tasks, manages execution, incorporates expert input, builds experience, and recovers from mistakes.

An agent harness connects the model to the outside world, enabling it to use tools, remember information between steps, and interact with complex environments. Done well, an agent harness sends each task to the right model for the job, including open-source models.

This makes AI more accurate and efficient, and therefore far less expensive to run. It’s essentially how companies move from having a bunch of disparate AI models to developing agentic AI that delivers ROI.

Four reasons to build a harness

A well-built harness pays off in four ways.

1. Control

    By relying on a harness instead of a bunch of discrete and disconnected tools, you can have a single workspace. That way, your business intelligence, data work, and agents all run through the same environment. 

    2. Choice

    With a shift in focus from just the models to the infrastructure that wraps around them, you can choose the right model for the right job and make new selections as prices and capabilities shift—something that’s happening constantly. 

    3. Context

    Agentic AI absolutely requires context to function. When models are personalized and informed by a company’s own data and context, you can stop burning tokens on rediscovering what the business already knows. Teradata estimates that this can cut token spend by 40% to 50%.

    4. Compliance

    As AI agents take on more consequential work, governance is rapidly becoming the second-biggest enterprise priority after economics. Business leaders need to know which agent did what task, when it happened, who authorized it, and whether it can be audited. A strong harness builds these controls in from the start, not as an afterthought.

    Choice is already reshaping how large enterprises think about procurement. Companies where AI tool costs have increased tenfold in a single year are now actively planning to build and own their own harness infrastructure to avoid being locked into any single vendor’s pricing or roadmap.

    Where to go from here

    The industry is already moving past the basic harness and toward the “meta harness,” a layer that sits above individual harnesses and routes intelligently across models, systems, and environments. For enterprise buyers, the question is no longer just which harness to choose, but whether it will work with the others you will need.

    Before committing to any AI project, ask one question: for every dollar you’re spending, what cost does it reduce, and where? If you can’t answer that concretely, the project isn’t ready.

    Strong models and strong harnesses are both necessary, but the harness is what makes models governable, portable, and more valuable over time. Leaders who treat the harness as an afterthought will keep accumulating tools. Those who build harness-first will find that the harness—not any single AI model—becomes their most durable AI asset.

    Before your next AI purchase, ask whether your harness lets you send tasks to the right model, open or proprietary, as the landscape shifts. That question will tell you more than any benchmark.

    Sumeet Arora is chief product officer at Teradata.