Before diving into the risks, it's worth stating plainly: Trinity Pension Consultants is not anti-technology. Far from it. We're always evaluating how AI, automation, and better portals can help us serve clients more accurately and efficiently. The goal of this article isn't to scare advisors away from AI. It's to help them use it with open eyes. 

As Michael puts it, "The issue really lies with taking AI and not adding the human element and fact-checking the information it gives you. It's a great starting point, but it should never be the ending point for advisors." 

That distinction, between AI as a starting point and AI as a substitute for expertise, is the thread that runs through everything below. 

AI has found a comfortable, low-risk home in several corners of retirement plan work: 

• • Participant communications, simplifying enrollment materials or explaining plan features in plain language 

• • Census data processing and reconciliation 

• • Meeting assistants that capture detailed notes during client or committee meetings 

• • Summarizing documents or regulatory updates for internal research 

• • Prospecting and CRM support, drafting outreach emails, building target lists, and keeping sales funnels on track 

Advisors are also using AI to prepare for new business meetings by pulling together plan-level information ahead of time, helping them walk in better informed. Used this way, AI is a productivity multiplier. The trouble starts when that same confident, plausible-sounding output gets applied to something far more consequential like plan compliance. 

Every retirement plan document is its own animal. Eligibility requirements, vesting schedules, testing methods, and distribution provisions are all specific to that plan's language. AI tools, by contrast, are trained on general patterns across the industry. They can give you a confident answer, but confidence is not the same as accuracy. 

"AI will just take a lot of generalities and give a confident answer," Michael notes. "At the end of the day, if a mistake is made or bad advice is given, liability doesn't pass to the technology. It's still assumed by the person giving the advice." 

That last point deserves its own moment: whatever AI tells you, you're the one who owns the outcome. 

AI models are trained up to a certain date, but regulations don't stop moving just because a model's training data did. This is especially true right now, in the middle of the SECURE 2.0 rollout. 

"There's just under 100 regulatory changes, and not all of them are finalized," Michael explains. "You could get a confident answer from AI about a SECURE 2.0 change, but the final regulations and rules haven't even been rolled out." Add the fact that many recordkeepers have not yet built the infrastructure to support new provisions, and you get a real gap between what AI says is possible and what is actually available to implement. 

Even something as concrete as contribution limits or the new Roth catch-up requirements can trip up a tool that isn't tracking real-time guidance. 

One subtle risk is assuming that because an AI-generated test result or summary looks polished, the underlying work was done correctly. Michael is careful to separate two things that often get blurred together: AI's ability to produce plausible-sounding text and its accuracy on plan-specific regulatory details. Those are not the same skill. 

Retirement plan data is sensitive, including Social Security numbers, compensation details, and dates of birth. Before feeding any of that into an AI tool, advisors should know exactly where the data goes, whether it is used to train the underlying model, and how long it is retained. 

This isn't hypothetical. Michael points to a real case where a plan advisor was sued after participant data ended up in the public domain via a general-purpose AI tool. Even "anonymized" data can sometimes be re-identified, particularly for smaller plan populations. Before adopting any AI tool, it's worth asking the vendor directly:

• • Where is our data stored, and who has access to it? 

• • Is our data used to train your models, and can we opt out? 

• • What happens to our data if we stop using the tool? 

• • What human oversight exists for anything compliance-related? 

• • Can you provide documentation we could show during a DOL or IRS audit? 

Under ERISA, the duty of prudence is about how a decision was reached, not just whether it turned out fine. Advisors leaning on AI-driven recommendations should treat the output as a research starting point, document the human review behind any decision, and validate anything AI-generated against the actual plan document, ideally with a TPA or ERISA counsel involved. 

Some things simply can't be automated away: 

• • Interpreting ambiguous or conflicting plan document language 

• • Determining the right correction method under EPCRS, when more than one path is available 

• • Handling coverage and nondiscrimination testing nuances when a workforce changes mid-year 

• • Navigating a DOL or IRS audit, which requires a real conversation, not a generated document 

Michael shares a case that illustrates this well: a long-standing Trinity client failed ADP/ACP nondiscrimination testing for the first time in years. Because Trinity knew the client's history, including its consistent practice of making a large year-end profit sharing contribution, the team recommended retroactively implementing a 4% safe harbor non-elective contribution. The plan passed testing, highly compensated employees avoided refunds, and the client still made the contribution it had already planned to make. 

"That kind of solution doesn't come from running a test and getting a pass or fail," Michael says. "It comes from years of institutional knowledge about a specific plan and the plan design expertise to recognize an opportunity that a generic tool would never surface." 

As AI becomes more accessible, it becomes easier to skip expert review altogether. That makes the TPA relationship more valuable, not less. A good TPA brings the plan-specific knowledge AI fundamentally lacks, tracks evolving guidance as SECURE 2.0 continues to roll out, and understands whether recordkeepers and vendors have the infrastructure in place to support a given change. 

Michael's advice for advisors: bring questions to your TPA before implementing new AI tools that touch plan administration, ask your TPA to review AI-generated compliance content before it reaches a client, and use that relationship as a sounding board whenever something novel comes up. 

If there's a single takeaway, it's this: the question isn't whether AI gets it right most of the time. It's what happens the one time it gets it wrong, and who's left holding the liability when it does. 

Used well, AI frees up time for the parts of the job that actually require a human: building relationships, catching the thing that "doesn't look quite right," and standing behind a recommendation with real accountability. Used carelessly, it can quietly introduce risk into one of the most highly regulated corners of financial services. 

As Michael puts it, trust with clients "is built in drops, and lost in buckets." AI can help fill those drops faster than ever. Just make sure a knowledgeable human is still the one signing off before anything goes out the door.