The numbers from Liberty Mutual’s 2026 Independent Agency Growth Study tell a clear story. Of nearly 1,200 US agency principals and staff surveyed, 65% of agents said they had used AI for work in the past year, up from 37% in 2025. Yet agency-level oversight has barely moved in the same period: only 18% of principals say their agency has a well-defined AI policy, and just 22% trust AI with client and business data.
That gap - agents running ahead of the institutions they work for - is a significant issue in the independent agency market right now. Jonathan Patel (pictured left), co-founder and COO of Cara, a New York-based AI operating platform built for insurance agencies, spends most of his time in it. His view is that agencies closing this gap fastest are not the ones deploying the most AI. They are the ones that have worked out exactly where AI stops and the licensed professional takes over.
“When a task affects client coverage, that’s when a human obviously will review the work, stand behind the final decision,” Patel told Insurance Business. “Technology is great at getting you from zero to a certain percentage of the way there. But obviously with the risk and the type of business that we operate in, a professional needs to be able to stand by that.”
Read more: Agents are using AI. Their agencies mostly aren’t - why?
The starting point for understanding why agencies are slow to move is understanding what agency staff actually do with their days. Cara, whose founders previously built and sold Oyster - a digital insurance brokerage acquired by McGowan Companies in 2025 - puts the proportion of a broker’s time spent on administrative tasks at up to 70%, consistent with figures cited in its own published materials. That means rekeying data across carrier websites, generating ACORD forms, building coverage comparisons, checking policies, producing proposals - work that does not require licensed judgment at every step but currently consumes the majority of the working day.
The document-heavy, rule-based workflows are where AI is delivering the most measurable value right now, Patel said. Policy checking, coverage comparisons, certificate of insurance issuance. According to Cara, workflows that traditionally take 90 minutes of manual work can be completed in approximately two minutes on the platform. The company raised $8 million in seed funding in March 2026 and reached seven-figure ARR in seven months.
The harder question is not whether AI can do this work. It is whether agency staff will trust it, review it properly, and stand behind it with clients. “AI can take something from one data system, can analyze it, can produce a proposal in that workflow,” Patel said. “But the output will sit there for an account manager to review, to review the audit trail, see where the data has come from and say, okay, I’m comfortable with this, I agree with this, and I’m happy to stand behind this.”
That last clause matters more in insurance than in most industries. An account manager who cannot explain the basis of a coverage comparison to a client - who has simply relied on output without understanding it - is carrying E&O exposure, not saving time. As the E&O gap around AI-driven professional services continues to widen, the question of who reviewed what and when is becoming a live coverage issue for agencies themselves, not just their clients.
The case for domain-specific tools over general AI comes down to what gets missed. A 400-page commercial policy is not a document that general AI reads the way an experienced account manager does. If a form attached to that policy affects a provision on page 10, a general AI tool may produce a confident answer that overlooks it entirely.
“General AI tools can produce answers that sound convincing but often miss important details,” Patel said. “And when you’re talking about real E&O exposure, overlooking a policy exclusion or a coverage provision for an agency can create exposure. So I think we’re in a unique position where that is of extreme importance.”
For agency principals evaluating vendors, Patel’s checklist starts with data: where does client and policy information go, and is it used to train models? SOC 2 and HIPAA compliance are baseline requirements, and he recommends reading the full reports rather than simply asking whether a vendor holds them. Beyond data security, the operational questions matter as much. Can the tool trace its reasoning back to the original policy document? Does it flag uncertainty rather than guess? How does it integrate with the agency management system, and what access controls are available across teams?
These are not abstract concerns. Zywave’s 2026 Broker Services Survey found, for the first time, that failure to adopt modern technology and AI has entered the list of top reasons employers give for switching brokers. It means the question for agency principals is no longer whether to adopt AI but how to do it in a way that is auditable, defensible, and genuinely integrated - rather than layered awkwardly on top of existing systems.
Read more: Reclaiming broker time: The technology reshaping insurance servicing
Agencies that treat AI as a one-time implementation will fall behind those that treat it as an ongoing operating discipline. “You might have implemented AI six months ago and said, great, I’m ready for the next two to three years,” Patel said. “And that’s just not the case.”
The practical upside for agencies that get this right is tangible. Account managers who are not rekeying data and reformatting proposals can support larger books of business without additional headcount - a significant advantage in a market where talent acquisition remains a persistent constraint. The agencies best positioned to capture that upside, in Patel’s view, are not necessarily the largest or best-resourced. They are the ones willing to identify the three to five workflows causing the most friction, start there, and build buy-in from staff alongside the technology rather than hoping the technology drives the change on its own.
“Skills like coverage expertise, client communication, thinking from first principles, sound judgment become even more valuable,” Patel said. “Because if you just produce output and don’t understand what’s gone into it - the data sources - it’s just not valuable.”