I Gave an AI Agent a Real Job

Last week I wrote about AI agents and asked a fairly basic question:

Why would someone like me want one?

I understand why businesses are interested. An agent that can process invoices, answer customer questions, write software or manage inventory might save a company considerable time and money.

But I don’t run or try to run a company anymore.

I wanted to know whether an AI agent could do something useful for me.

So instead of asking another question about agents, I decided to give one a real job.

It turns out I had a pretty good one waiting.

Where Should We Live?

My wife and I live in a house in northwest Reno. At some point, whether soon or several years from now, we will probably have to decide whether staying in the house continues to make sense.

There are plenty of alternatives.

• Stay where we are.
• Buy a smaller house or condo.
• Rent an apartment.
• Move into a 55-plus community.
• Consider independent living.

That sounds simple until you start defining what you actually want.

We would want roughly 950 to 1,200 square feet, two bedrooms and preferably two bathrooms. No stairs. A balcony or patio would be nice. Little or no yard maintenance. Covered parking. Fitness facilities either included or nearby. Reasonable access to shopping, restaurants and medical services.

Then there is cost.

I established different limits for rental housing, purchasing a condo or house, and independent living. Independent living is particularly difficult to compare because a monthly charge might include meals, housekeeping, utilities, transportation, activities and other things we currently pay for separately.

Suddenly, “Where should we live?” isn’t a question.

It’s a research project.

Which sounded like a pretty good job for an AI agent.

Here’s the Job

I gave the agent the objective, our requirements and the financial limits.

I told it not simply to find properties. I wanted it to compare the alternatives, eliminate those that clearly didn’t work and identify the strongest candidates for further investigation.

I also gave it some boundaries.

Don’t contact anyone. Don’t make appointments. Don’t submit applications. Don’t give anyone personal information.

And don’t make decisions for me.

If it ran into something requiring my judgment, come back and ask.

Then I turned it loose.

Twenty Pages Later

The result was a report of more than 20 pages.

That wasn’t quite what I expected.

But there was a lot of useful information in those pages.

The agent determined that conventional 55-plus apartments probably weren’t worth pursuing. The ones it could verify were generally too small and often had only one bathroom.

It found several conventional apartments that came reasonably close to our requirements and reduced those to two or three worth further investigation, explaining the rationale for its recommendations.

It found that condos were a more practical ownership alternative than either houses or townhomes under our requirements. Townhomes frequently introduced the very thing we were trying to eliminate: stairs.

Independent living was more difficult. One facility appears potentially to fit our requirements and budget, while several others need actual pricing before they can be seriously evaluated.

We already knew enough about one of the better-known independent-living facilities to eliminate it. Friends who live there have told us that the size apartment we would want costs considerably more than I am willing to spend.

The agent didn’t waste time trying to convince me otherwise.

So did it solve the problem?

No.

But something interesting had happened.

It Gave Me Better Questions

I still have some legwork to do.

For an apartment, I need to know whether the advertised rent is really the total rent after mandatory fees. Is covered parking actually available for the particular unit? Can we reach it without stairs? What utilities are included?

For independent living, the questions become even more specific.

• What does an actual two-bedroom unit of the size we want cost—not what does the least expensive studio “start at”?
• What is the total monthly charge for two people?
• How many meals are included?
• What about housekeeping, utilities, Internet, transportation and parking?
• Are there entrance or move-in fees?
• And, perhaps most important, is such a unit actually available?

The agent couldn’t answer all of those questions because much of the information simply isn’t published.

Someone is going to have to make some phone calls or visit the facilities.

That someone apparently is still me.

But there is a substantial difference.

Instead of calling 15 places and asking each one a collection of questions as I think of them, I can concentrate on a handful of realistic alternatives with a fairly precise checklist of what I need to know.

The agent didn’t eliminate the work.

It eliminated a lot of unnecessary work.

That may be a more useful result than I expected.

Then I Discovered Something the Agent Didn’t Know

There was another lesson.

The agent compared the alternatives financially and concluded that staying in our present house was the cash-flow benchmark to beat.

On the information I had given it, that made sense.

We don’t have a conventional mortgage payment. The agent added property taxes, insurance, HOA fees, maintenance, utilities, yard work and other identifiable housing expenses and came up with a monthly cost.

There was only one problem.

I had forgotten to tell it about a cost we don’t actually pay each month.

We have a reverse mortgage.

With a reverse mortgage, we don’t write a mortgage check every month. Instead, interest and fees are added to the mortgage balance.

Next month’s interest is then calculated on a larger balance.

It compounds.

The cost is real. It just doesn’t show up as money leaving our checking account.

That means staying in our house could be considerably more expensive economically than the agent’s comparison indicated.

The agent hadn’t made a mathematical mistake.

I had given it an incomplete problem.

That may have been the most important thing I learned from the entire exercise.

The Agent Doesn’t Know What I Know and Doesn’t Know What It Doesn’t Know

I’ve been delegating work to people forever.

If I had hired a research assistant and neglected to mention the reverse mortgage, that person might have reached exactly the same conclusion.

An AI agent is no different.

It can research what I ask it to research. It can organize the information. It can compare alternatives and point out inconsistencies.

But it doesn’t automatically know the things I haven’t told it.

Worse, it may have no way of knowing that an important piece of the problem is missing.

It doesn’t know what it doesn’t know.

And I have to know enough about the problem to recognize when something important is missing.

That puts an interesting twist on the idea that AI will allow us to hand jobs to software and forget about them.

At least from this experiment, I don’t think we’re there.

Then There’s the Other Decision Maker

There was one more test.

My wife read the report.

She mostly agreed with the way the problem had been structured, but she didn’t necessarily agree with all of the agent’s conclusions.

Several apartment choices disappeared immediately. One was in a location she didn’t like. Others simply weren’t attractive enough or were below our minimum size to pursue.

She also noticed something neither the agent nor I had focused on.

Most of the rental search seemed to be finding larger apartment complexes. What about smaller properties, individually owned condos for rent and other possibilities that don’t show up as prominently in the usual searches?

Good question.

Another question for the agent—or perhaps for me.

That’s exactly what should happen.

The agent wasn’t deciding where I should live. It was helping two people think about where we might live.

And those two people don’t necessarily assign the same importance to every criterion.

There’s an even more fundamental question we haven’t resolved.

I have been thinking about alternative living arrangements for some time. I’m still not convinced my wife is as interested in changing our living arrangement as I am.

So the real question may not yet be:

Where should we move?

It may be:

Should we be considering moving at all?

No amount of Internet research can answer that question for us.

And I’m pretty sure I don’t want an AI agent trying.

And the Financial Question Got Bigger

There was one more complication.

My original instructions assumed that if we bought a condo, we might finance up to a certain amount.

But that’s not necessarily what we would do.

We could conceivably buy a moderately priced condo for cash and have no mortgage payment at all.

That sounds attractive.

Except the money has to come from somewhere.

If we take money out of our investments to buy the condo, how much investment income do we give up? Would the remaining investments still produce enough income to cover our expenses?

And which of our current expenses would actually disappear if we moved?

Suddenly I’m not just comparing the monthly cost of a house, apartment and condo. I’m comparing what happens to our entire financial picture under each alternative.

That’s going to require more work.

So Did the Agent Work?

I think so.

But not quite the way I expected.

It didn’t tell me where to live.

It didn’t eliminate the need for phone calls, visits or decisions.

What it did was take a messy problem involving dozens of possibilities and turn it into a much smaller problem involving a handful of alternatives and a much better set of questions.

It probably saved me hours of searching.

More importantly, it helped me understand what information I still need.

And some of its information has an expiration date.

If we decide to sell our house, next spring may be a more sensible time. By then the apartments, condos and independent-living units available today may be gone and entirely different ones may be available.

The research will have to be updated.

But the structure of the problem won’t change nearly as much.

I’ll know what I’m looking for and what questions to ask.

There is still more work to do. We need real prices for the independent-living units we would actually consider. We need better numbers for the true economic cost of staying in our house. We need to investigate some different kinds of rentals and perhaps a few condos.

And eventually my wife and I need to decide whether we actually want to do anything.

So this experiment isn’t finished.

There may be another article—or two—before it is.

The experience also made me think about other jobs where an agent might actually be useful.

Comparing insurance alternatives. Evaluating health-care choices. Planning a road trip involving routes, hotels and stops.

They have something in common with the housing problem: lots of alternatives, information scattered across different sources, changing prices and availability, and ultimately a decision that I still want to make myself.

So after giving an AI agent its first real job, I have a somewhat better answer to the question I asked last week.

Why would I want an AI agent?

Not necessarily to make my decisions.

Maybe I want one to do enough of the grunt work that I can spend my time making the decisions that actually require me.

That seems like a pretty good division of labor.