Why The Next Competitive Advantage Lies Beneath The Interface

From An Old Debate To Generated Screens

In 1997, Ben Shneiderman and Pattie Maes debated whether people would eventually delegate to software agents or keep direct, predictable control. Two years later, Eric Horvitz proposed a middle path – pair automation with direct manipulation, and let the system act alone only when acting is worth more to the person than doing nothing.

Direct control won by default for decades, mostly because agents weren’t good enough (that is, until now), and I think the mindset is changing very quickly.

Jakob Nielsen has called AI the “first new interaction model in more than 60 years,” where people state the outcome they want rather than the steps. Researchers have since shown models can generate an entire interface for a single request, and work presented at CHI 2025 built interfaces that reshape as the task changes.

If that holds up, the screen looks less like a destination and more like a receipt: generated for a moment, read once and gone.

Designing For The Agent And The Person Behind It

When an agent does the clicking, your most frequent user is software somebody else built. The first thing I would worry about is legibility, since clean APIs, clear permissions and predictable results will matter more than a polished dashboard.

The second is identity. Researchers at MIT and Stanford have proposed extending today’s login and authorization standards so a person can hand an agent scoped, auditable authority. Retirement already handles a version of this through powers of attorney, just not at software speed.

I don’t think agents remove the person from the process so much as move her within it. I keep coming back to Lisanne Bainbridge’s 1983 paper “Ironies of Automation“ (well worth reading if you haven’t yet), which argued that automating most of a job can make the part left for the human harder because she is now watching over something she rarely touches. That is our account holder with the three old 401(k)s, who sees one screen and has to get that one right.

So, for every flow, I would ask how much of it she should still touch and sort them into three groups:

• Delegable work the agent can simply do, like balance checks, statements and last activity

• Confirmable work where the agent proposes and the person approves on one screen, like contribution changes, address updates, rebalancing and routine rollovers

• Human-only decisions like beneficiary designations, taxable distributions and anything that cannot be undone

That approval screen will need some deliberate friction. One study found that designs that made people stop and think cut blind trust in the AI the most, even though people liked them least. I spend my days removing friction, so that stung a bit, but I would build it anyway and keep it short.

Disclosure And Loyalty When An Agent Sits In The Middle

Most financial regulation assumes a person read something. Scholars have long argued that mandated disclosure rarely does what it promises and that a reliable intermediary works better as a substitute for it.

A personal agent may be the first intermediary that reads every notice, every time. That shifts the question from whether the person saw the disclosure to whether her agent understood it and acted in her interest, and I don’t think our rules are written for that yet.

AI agents raise a classic principal-agent problem (information asymmetry, discretionary authority and loyalty), with the usual fixes struggling to keep up at machine speed. Retirement has had a word for the answer since ERISA was enacted in 1974—“fiduciary”—and the open question is who owes that duty when an agent sits in the middle.

What I find most interesting is that ERISA already defines a fiduciary by function; if you exercise discretion over the money or give advice for a fee, the duty is yours. An agent that picks where her three 401(k)s go, chooses the funds and moves the money is doing exactly that, yet it’s not a person and cannot be held to anything. I suspect the duty lands on whoever built it, deployed it or put it in front of her, and I would rather the industry help answer that than wait for the first lawsuit to answer it for us.

The System Of Record And The Value Shift Occurring

Many tech conversations right now are about software going headless: agents calling services directly and whole categories of applications collapsing into a single conversation. Some go as far as saying IP in software has effectively become worthless. I think that’s mostly right about the experience layer, yet it remains to be seen for others.

If the screen gets generated on demand and the workflow in the middle can be rebuilt by an agent, a lot of what software companies have traditionally charged for gets squeezed. My answer to what doesn’t is the system of record: the authoritative, reconciled record of who owns what, how it’s taxed and who is accountable when it’s wrong. In retirement, that record carries custody, regulatory reporting and activity that has to tie out every day, and none of that goes away just because the interface did.

So that record has to become agent-ready—real-time, permissioned and logged access, and rules that live in or link to the record itself. Much of wealth and retirement still runs on batch-era systems that settle overnight, and I suspect that gap gets exposed quickly once an agent asks where a rollover stands at 9 p.m.

Going back to our account holder with three old 401(k)s, she may never see most of what we build, and that’s fine. What she will notice is whether the one screen she sees was worth stopping for and whether the record underneath was right when the money moved.

So if the experience layer is heading toward zero, I suspect the value moves down into the platforms and rails that have to be right and out to the people and institutions willing to stand behind them when an agent gets it wrong.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?