How Businessolver Is Looking to Move AI Beyond the Chatbot

AI has become increasingly effective at answering questions, but 2026 AI Tech Ascension Award Winner Businessolver is betting that the next major shift will be systems that act before users know they need help. We spoke with Sony SungChu, Chief AI Officer at Businessolver, to discuss how Sofia is evolving beyond the chatbot model, what responsible AI looks like in highly regulated environments, and why anticipatory intelligence could reshape how employees interact with benefits and workplace technology.

Many AI solutions in the market focus on answering questions and reducing headcount. Why is anticipatory intelligence the next evolution?

In general, chat-based AI still waits for a person to provide a question as input. It’s a natural approach if someone knows what they need and where to find it. But benefits often create the opposite situation. People encounter eligibility rules, coverage decisions, enrollment deadlines, and unfamiliar terminology at the same time they are dealing with a change in their lives. They might not know what to ask or that they’ve missed an important question.

Anticipatory intelligence looks at everything we know about a person so that it has enough context to offer guidance before the person engages. For example, Sofia knows that an employee’s child will age out of the employee’s plan in 60 days. It checks the plan rules, confirms when coverage ends, and prompts the employee early enough to review other coverage options. The employee does not need to know what question to ask or remember the deadline.

What makes Sofia an intelligence layer instead of a chatbot?

One way to look at this is that a chatbot is generally the conversational surface. For Sofia, her role is broader in the sense that she sits behind that surface and the whole system is made up of the data it can access, the services it can invoke, the rules that constrain its behavior, and the workflows it can help complete. When we put that together along with reasoning and memory, we get a full intelligence layer.

For example, an employee chats Sofia: “I got married. Can I add my spouse to my health plan, and what will it cost?”

In this case, Sofia needs to identify the employee and their plan, check the qualifying life event rules and timing, retrieve the applicable coverage options and costs, and explain this all to the employee in plain language. If the employee is eligible, Sofia can help start the enrollment workflow and manage the end-to-end communication with the employee. If information is missing or the situation falls outside the rules, she can route the case to an advocate with the relevant context attached.

We’re seeing organizations place significant value on immediate AI transformation. What does effective transformation look like?

I think organizations often mistake giving people a new tool for transformation. Giving people a tool changes access to technology, but transformation happens when work moves differently through the organization, and we can measure the improvements.

Any transformation depends on the same practical discipline. Someone has to own the use case, and the system has to fit into a real workflow with reliable data and a clear standard for quality. People need to know when to trust the output, when to review it, and what to do when it fails. You have to understand the failure modes before you put the system into production.

Without that foundation, we’ve found that an AI pilot can look successful in a demo but not as useful when it meets real work.

And while the tech creates the opportunity, adoption and training determine whether it becomes real improvement. Businessolver’s 2026 State of Workplace Empathy AI Special Report helps illustrate this. Although 61% of employees are optimistic about AI’s impact on their future, nearly 40% remain concerned, and only about half say their employer has provided adequate training. Those who received that training were more likely to say AI was improving their confidence and the value of their work.

How can organizations apply AI responsibly in highly regulated environments such as HR and benefits?

At the heart of it, it’s knowing what information was used, what rules constrained the result, who is accountable for the outcome, and what happens when the system is wrong.

For our Ai applications we require data classification, access controls, testing, monitoring, auditability, and an escalation path. We also require clear ownership after launch. An AI system can change as its data, prompts, connected tools, or surrounding workflow changes, so it’s important to keep up.

We also impart that human review is useful but only when the human has the context, authority, and time to evaluate the output. Sending an opaque recommendation to an employee and calling it human oversight doesn’t create meaningful accountability. In benefits, plan rules and eligibility systems must remain authoritative. An AI system can help interpret a request, select relevant information, or explain a result, but it shouldn’t invent a rule or quietly substitute its judgment for the system of record.

As engineers, we should know what parts of a system work better as deterministic components and which parts are better suited for generative AI and people should know when AI is involved, what role it plays, and how to challenge or correct an outcome.

Why invest in a purpose-built AI system instead of a generic or custom labeled model?

Models, whether they are generic or custom labeled provide general intelligence and the purpose-built AI system gives that model context, which in turn, makes the intelligence actually useful.

For example, foundational models don’t inherently know which source is authoritative, whether a user is allowed to see a record, or which business rule applies, or when a response should be escalated. The nuance comes from the surrounding architecture.

So really the best architecture can continue to use general models as they improve, but it surrounds them with the controls required by the domain.

How do you measure whether AI is improving outcomes or driving optimal results for your organization?

We strongly believe that a system can produce fluent answers, attract heavy usage, and still fail to resolve the underlying problem. This is something we learned from building Sofia back in 2015.

So we look at measurement in layers. The first layer is model performance, such as accuracy, retrieval quality, and the rate of unsupported answers. The second is system performance, including permissions, latency, tool execution, escalation, and the ability to recover from an error. The third is the business outcome: resolution quality, repeat contacts, time to complete a task, cost, decision quality, or user confidence.

Those layers need a baseline. If an organization claims that AI reduced service effort, it should be able to compare the new process with the prior process and account for changes. If the system offers recommendations, the organization should track whether people acted on them and whether the intended result followed.

How do you see AI changing the way we work over the next decade?

In general, systems will handle more of the routine cognitive effort, and people will spend more time on judgment, relationships, direction, accountability.

Employees will have more access to knowledge, but it will also raise the standard for human judgment. People will need to understand the context behind a problem, assess whether an answer is trustworthy, and recognize when the situation requires escalation. When to use AI and when not to, because not every task is suitable for AI.

The risk is using AI to produce more decisions without producing better ones. A good reason to keep humans around and in the loop.

What separates organizations that are creating real business value with AI from those that are still experimenting?

I think many believe AI is different in this regard. It is different in some ways, but the same change management principles still apply. We create value when we connect a real business problem to a defined workflow, know who owns it and make it measurable.

It’s also important to understand the gap between a demo and a production system. A demo can show what is possible, but the last mile is usually where the real work begins. 

Finally, feedback loops are still important. Organizations need to keep learning what is working, what is not, and what needs to change.