Companies have embraced generative artificial intelligence, examining where software could help employees. The truly agentic enterprise must examine what happens when the software can do the work itself.
An agent can take information from one part of a company, decide what needs to happen next and perform the task, allowing work to move across functions without an employee initiating each step.
“Think about autonomy rather than assistance,” i2c CEO and Founder Amir Wain told PYMNTS CEO Karen Webster.
Wain described agentic AI as a progression from the familiar large language model. An LLM can take a prompt, interpret information and generate a response. An agent uses the model’s capabilities as part of a broader system that can perceive information, decide what to do, and act. Wain added a fourth component: a learning loop that allows the system to improve based on what happens after it acts.
Put enough of those capabilities inside a company and the operating model starts to change. Agents can perform work that once passed among employees, departments and software applications. For businesses created around AI from the outset, that can also change what is required to get a company off the ground.
Webster pointed to one-person businesses built around agents as an early example.
“These companies aren’t building big teams,” Webster observed. “They’re building businesses around agents that do the things that lots of people used to do.”
Wain said software development demonstrates how far that can go. The time and money required to turn an idea into a prototype have fallen sharply because AI can perform much of the development work. He said an idea can potentially become a live product on the same day.
That doesn’t make the resulting product a viable business. Customers still have to want what’s on offer, and a company still has to find a way to scale it. The more immediate lesson for established enterprises is how much of the work that preceded a product launch can now be reconsidered rather than merely accelerated.
Wain said companies should start with processes they perform at scale. Automating something done rarely offers limited value. The larger gains can come when a company discards assumptions embedded in an old workflow and designs the process around what agents can now do.
At i2c, Wain said, that approach led the company to redesign its software development life cycle rather than add AI tools to individual stages of the existing process. Yet he doesn’t consider i2c a fully agentic enterprise.
“I would say we are piloting. I wouldn’t say we are fully there yet,” Wain told Webster. He described the company as being in a transition in which roles are being reconsidered as autonomy expands.
The Accountability Question
Wain expects AI capabilities to advance rapidly enough that companies shouldn’t build permanent operating assumptions around current limitations. His argument is that many operational activities that still require human intervention today may eventually be performed autonomously.
Repetitive processes performed at scale are obvious candidates, but he said companies shouldn’t simply reproduce existing workflows with agents substituting for employees. They should reconsider which steps remain necessary when an agent can gather information, make a decision and execute the resulting task.
Governance becomes more demanding once an agent can act.
Wain said companies should evaluate autonomy along two dimensions: whether an action can be reversed, and the potential impact if the agent gets it wrong. A decision that can be readily undone permits more latitude than an irreversible action involving money, customers or regulatory obligations.
“You can’t blame it on the agent,” Wain said. “You have to take the responsibility.”
Webster connected that issue directly to financial services.
“It’s who’s liable when something goes wrong,” she said, pointing to uncertainty across regulators and jurisdictions. “Unless some of those things become clearer, it’s hard to imagine giving agents the wheel.”
Fraud carries grave consequences. Wain said a company considering an irreversible autonomous action has to determine whether potential losses fall within a range it can absorb. An agent authorized to conduct activity on a vastly larger scale creates a correspondingly larger exposure, making limits on its authority essential.
Companies also have to protect themselves from agents that operate outside their intended parameters. Wain raised the prospect of rogue agents and called controls essential. He separately predicted that AI will create new forms of fraud, putting enterprises in the position of governing their own agents while defending against AI deployed by others.
How far that progresses over the next several years is difficult to forecast. Wain declined to make a precise prediction for 2030 because of the speed at which the technology is developing. He does expect interfaces, businesses and jobs to change substantially.
The nearer-term work is already identifiable. Companies can determine which processes merit redesign, establish how much authority agents receive and make certain that someone remains responsible for every decision the software is allowed to make.
Watch the full interview with Amir Wain to learn more about:
- How an entrepreneur could build and test an AI-native business with far fewer resources than previously required.
- What Wain would do differently if he were creating i2c from scratch with today’s AI capabilities.
- Why he expects familiar apps and interfaces to change as agents become capable of handling more activity on a user’s behalf.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
