Enterprises don't need clean data to start building AI. The foundation a company lays today, even on a decade of disorganized files, decides whether it keeps control of its own presence as AI agents begin mediating discovery and transactions. As agentic commerce matures, a brand's own agent stands to replace the website as the front door, and the companies organizing their unstructured data now hold the advantage when that shift lands.
Carver Anderson, CEO of Suggestic, a health-technology company that builds AI-driven personalization for wellness and clinical brands, works from that premise. An electrical engineer by training who still writes software, he has spent much of his career developing digital products across the healthcare sector. As a product lead within UnitedHealth Group's UHG Ventures, he helped bring Level2, a type 2 diabetes reversal platform, to market, and his earlier work spanned finance and consumer wellness, including a stint at mindbodygreen. The hesitation he hears from many leaders strikes him as unnecessary.
"You don't have to wait until your data is in perfect shape to build a pipeline. You can get started immediately," Anderson says. Many executives hold off until the next model lands, a habit that misreads how enterprise value gets created. Modern tools ingest unstructured information almost immediately, so the barrier to starting has all but disappeared.
Few organizations have data in anything close to perfect shape. Every large company carries some version of the same disorder, the residue of acquisitions, divestitures, and private-equity ownership, and none of that is disqualifying.
The condition of those files is what most often stalls an AI effort before it starts, so leaders feel they have to clean everything up first. Anderson rejects that sequence. "We have a client that was concerned because they had a decade's worth of disorganized files in Dropbox. That's fine. It takes a couple of discovery sessions to figure out what's relevant and come up with a naming structure. We can just build an ingestion tool to find the right files and build some intelligence into the pipeline," he says.
Underneath sits ordinary infrastructure. Identify the data, process it, and load it into a knowledge base, whether a vector database or a traditional store, and the rest comes down to designing semantic search that surfaces the right context at the right moment. Companies sidestep expensive legacy overhauls because the internal IT skills usually already exist.
If those skills already sit in-house, the remaining obstacle is perception. Capable models are now broadly available and largely commoditized, and the advantage goes to teams that know how to apply them. Anderson spends a good part of his pitch deflating the mystique around that work. His team brings more than a decade in machine learning that predates today's large language models, and treats the current moment as a straightforward extension of enterprise architecture.
"This isn't magic. It's really complex math that can feel like magic, but it's just another tool in the stack. Like all good traditional IT, you need competence in system architecture and tool choice to match the use case. But it's really not that hard," he says. That competence is also less rarefied than the hype implies. The tooling shifts week to week, which keeps even specialists in a state of constant learning and puts a capable in-house team closer to the people building the models than most leaders assume.
Most corporate AI pilots stall before reaching production. A promising demo lingers, then gets orphaned when the multi-year plan behind it slips, or the priorities shift. Anderson's answer is to abandon the long roadmap and start from the friction people feel every day, skipping the technology conversation entirely. A handful of those conversations surface the highest-return opportunities, which he carries back to leadership. The most valuable use case is often too heavy for a first build, so the team scopes down to something a short pilot can prove in six to eight weeks, well short of the twelve-to-eighteen-month program few leaders have the appetite for anymore.
"It demonstrates that these things can be implemented quickly and without spending a fortune. People see that it works, they can manage it, and then you build from there," he says. That early win generates the momentum that turns a single proof of concept into something a department adopts on its own.
That playbook applies in any sector. In a regulated field like healthcare, the constraints are heavier, and the returns are larger. The clearest near-term payoff is personalization that older systems can't deliver affordably. Legacy care models sort people into broad demographic buckets and tweak a plan at the margins, treating lookalikes as interchangeable. Processing many individual signals at once changes what a care plan can account for, and capabilities that stalled a decade ago, when this kind of tailoring took armies of people, now move forward. "Even if people look the same on paper, they're individuals, and that can happen through natural conversation alongside a multitude of signals like claims data and wearable devices," Anderson says.
The constraint side carries the lesson most leaders in regulated industries miss. Heavy compliance looks like pure cost, but built in early, it becomes a barrier few rivals can clear. The finer discipline is knowing exactly where lifestyle guidance ends, and a regulated medical device begins, a line that is rarely crisp and shifts between the United States, the European Union, and other markets. Anderson's teams settle it case by case through deliberate user-experience design, scoped with each client. "For the wellness companies we work with, the AI is obviously not diagnosing, and it's not directly asking users to act. That must come from a practitioner. It's saying, based on our information and past data, here are some options you might consider," he says.
Every piece of that approach builds toward one end. The pipeline drawn from messy files, the pilot that proves itself in weeks, the compliance that fences off a market, each one lays the groundwork an autonomous brand agent will eventually stand on. The infrastructure that tailors a patient's care plan today is largely the infrastructure that lets a company transact through its own agent tomorrow, and the agent-to-agent protocols that connect them grow more capable by the week.
Timing is the open question, and the answer rests with the incumbents. Full autonomy depends on the largest platforms loosening their grip, and the companies with the most traffic to protect have the least reason to move first. Newer entrants without that legacy may push the shift forward faster. "True agent-to-agent interaction is going to require releasing a little bit of control. An LLM agent speaking to a brand agent just makes sense as a future state," he says.
