BLOG
To Avoid AI Lock-In, Companies Must Own The Workflows That Create Their Value
Why enterprise AI-agent strategy depends on flexibility, measurable impact, and owning proprietary workflows.
This article was originally published by The Intelligence Record.
“You cannot outsource how you do business. LLMs are generalized machines. If you do the hard work of figuring out your workflow and translating that into an agentic workflow, you’d want to own it.”
Chan Suh, Chief Digital Officer, Prophet
As companies move past their initial AI experiments, executive teams are taking a harder look at token limits, monthly bills, and the very real threat of vendor lock-in. While deploying agentic workflows is becoming a standard approach for corporate agility, it doesn’t mean a company should outsource its core operations. For organizations with highly specialized workflows, defaulting to off-the-shelf generalized models usually means handing over control of proprietary business logic. Scaling AI tools effectively hinges on the decision of exactly what to own versus what to rent.
Chan Suh has seen these cycles before. Back in 1995, he co-founded Agency.com with $80 and rode the early dot-com boom all the way to a NASDAQ IPO, growing the business to 1,800 employees and $200 million in profitable revenue. Today, as Chief Digital Officer at Prophet, he helps enterprises navigate the current tech wave. Drawing on decades of scaling business value, Suh advocates for definitive ownership over proprietary AI.
“Just because you can adopt AI doesn’t mean you should. It’s going to be very important for enterprises to understand how to use agents and what parts of their work, whether it’s in supply chain or marketing or anything else, can be helped by agentic workflows,” he says. Suh views the discipline of deciding where agents genuinely help before deploying them as the foundation for every other decision that follows.
Ownership of the Workflow is Ownership of the Business
Suh’s central argument is that a company can’t outsource its most intelligent piece. How it does business, creates opportunities, fulfills customer expectations, and runs the processes that make everything work isn’t something a generalized model should own. “You cannot outsource how you do business,” he asserts. “LLMs are generalized machines. If you do the hard work of figuring out your workflow and translating that into an agentic one, you’d want to own it. That ownership is super important.”
What has changed is the cost of that ownership. Building proprietary workflows into machine systems used to require translating the process, building the interface, testing it, and evolving it, all of which made ownership prohibitively expensive. Thanks to AI, that work is now within reach for companies of nearly any size. “If you make it, you get to keep it. The thing that makes your company do what it does super well, you get to keep it. And the cost of making it is not nearly as expensive as it used to be.”
The Dependency Risk Hiding in Subsidized Pricing
The strongest case for preserving optionality is economic. Suh warns that current token pricing for frontier models is subsidized, and companies building their futures on those prices are exposed to a cost structure outside of their control. “At some point somebody has to pay for it. If you’ve been modeling your company’s future on these LLMs, what happens when they come back and say, ‘Sorry, it’s 4x what you used to pay’? Is that the future you want to build for yourself? Do you at least want an escape hatch?”
The escape hatch is a proprietary or captive model. With open-source LLMs now available to be trained, Suh envisions a future where companies of any size run their own captive model for proprietary and non-frontier work, reserving the expensive frontier models only for the tasks that genuinely require them.
The provider-incentive problem is why he’s skeptical of letting the large model companies design an organization’s agentic infrastructure. “Asking large LLMs to create your agentic infrastructure is like asking oil companies to design your cars. It’s going to be super fast and super exciting and loud, but it’s also going to be a huge gas guzzler, because that’s how they get paid.” The performative version of AI adoption, announcing a marquee deal to satisfy the board for two quarters, eventually comes due when someone asks whether the company actually made progress.
Impact Before Tools
The measurement discipline Suh advocates inverts how most companies approach AI. The question is not how to become more AI-powered, but what AI can do for the specific business. “Then you figure out what flavor of AI you need. If you’re a straightforward manufacturing and distribution company, do you need something that can solve quantum physics? Maybe that’s overpowered,” he says.
He frames the principle as impact before tools: define the impact you want, then find the tools to create it, rather than adopting capability for its own sake. Adoption metrics are not enough because adoption doesn’t measure progress.
In Suh’s view, the most accessible starting point is high-volume data analysis that humans can’t realistically perform. A company with 100 salespeople can’t digest every weekly sales report, but an agent pointed at that information stream can find the patterns and surface what leadership needs. That work requires little infrastructure investment and delivers immediate value.
Embedding proprietary knowledge is the deeper layer, and it has also gotten dramatically cheaper. Suh estimates the technology to run the basics of a custom AI deployment now costs below a million dollars a year, putting it within reach of companies that need to decide where their proprietary knowledge lives and what they want to keep.
Stop Planning, Start Executing
Suh’s guidance for companies beginning this transformation is to compress the time horizon and act. Because no one can predict the AI landscape five or ten years out, the right question is what a company can do in the next 24 months to maximize optionality while accelerating efficiency. “Companies need to understand what they can do right now, how they’ll measure it, and build on those successes,” he advises. “Stop doing pilots. Pilots are a great way to avoid accountability. Do the things you know you can do.”
Suh’s warning against ten-year infrastructure plans is pointed. The old model of cementing a technology architecture for a decade is, in his view, a fool’s errand in a landscape moving this fast. The better path is measurable progress in six-month increments, delivered through partners who preserve independence rather than trapping the client inside a single ecosystem. “Find the partners who can help you do it at a reasonable cost while preserving technology independence, so you don’t get captured. Do some things right. Stop drawing plans.”
The proliferation of frameworks, Suh notes, is itself part of the problem. Anyone can now generate a strategy document in minutes, but that doesn’t make it good. The advantage goes to the companies that execute against measurable impact while keeping their options, and their proprietary workflows, firmly in their own hands.
Don’t Miss a Beat
Sign up to receive our latest insights.
FINAL THOUGHTS
Let your people (and your agents) do the work they were meant to. Prophet gives you both the marketing and AI expertise that can help you create business impact. Learn more about our AI Solutions and how they can help drive uncommon growth for your organization.
