Why Would I Want an AI Agent?

For the last couple of years, I have been using artificial intelligence almost every day.

I ask ChatGPT questions. I use it to research subjects, help develop articles, compare alternatives, find information I would otherwise spend hours searching for, and occasionally figure out what Microsoft has done to my computer this time.

But lately I keep hearing about the next big thing in artificial intelligence: AI agents.

The technology industry seems very excited about them. Companies are developing agents to write software, handle customer service, process invoices, manage inventories and perform all sorts of other business functions.

That’s interesting.

But I don’t help run a company anymore or provide advice to those that do..

So, I have a much simpler question:

Why would someone like me want an AI agent?

From Answering Questions to Doing Jobs

The easiest way I have found to understand an agent is to compare it with the AI I already use.

Normally, I ask ChatGPT a question and it gives me an answer. I ask it to research something, and it researches it. I ask it to help write something and it helps me write it.

An agent goes a step further.

Instead of giving it a question, I give it a job to accomplish.

The difference sounds small, but it isn’t.

Suppose I am planning a trip. I can ask ChatGPT to find hotels along our route and suggest the best alternatives based on price, location and amenities. Then I can ask about restaurants near those hotels. If we plan to fly, I can ask it to find flights offering the best combination of price and convenient times and assess the alternatives. Then I can ask about transportation options at our destination. With all that information, I can compare the choices and make the reservations.

An agent could potentially be given the larger assignment: plan the trip within these dates, this budget and these preferences. It could determine what needs to be researched, gather the information, compare the alternatives and come back when it needs a decision from me.

I specify the objective rather than every individual step.

Of course, the next step would be to let the agent actually make the choices and complete the reservations. I’m not ready for that last step yet.

But the idea started to get my attention.

But Do I Actually Need an Agent?

This is where I think much of the discussion about AI agents gets things backward.

The fact that a computer can do something for me doesn’t mean I need it to.

If I want to know when to prune my roses, I can ask ChatGPT and have an answer in thirty seconds. Creating an agent to manage the problem would probably take longer than pruning the roses.

The agent has to solve a real problem.

As I started thinking about things I actually do, several possibilities emerged.

One is insurance. Comparing insurance policies can involve different companies, different coverages, different deductibles and different exclusions. The cheapest policy isn’t necessarily the best one. An agent could collect the alternatives, put them on a comparable basis, identify important differences and help narrow the choices.

Another is housing. At some point, many older people face a much more complicated question than simply whether to sell the house.

Stay where we are? Buy a condo? Rent an apartment? Move into a 55-plus community? Consider an independent-living facility?

Now add costs, square footage, number of bedrooms and bathrooms, parking, maintenance, meals, fitness facilities, transportation, medical access, shopping and location.

Suddenly this isn’t a Google search. It’s a research project.

And while thinking about this article, another very real example appeared.

I needed help finding a doctor for a family member whose insurance relationships are changing. Finding doctors is easy. Finding doctors who take the specific insurance plan, are accepting new patients, are reasonably close, have the appropriate specialty and actually have appointments available is something else entirely.

That’s another research project.

Now the idea of an agent doesn’t seem quite so theoretical.

Where an Agent Might Earn Its Keep

Agents are becoming surprisingly easy to set up requiring no special knowledge or tools. You do need to describe your objectives and identify constraints in ordinary language. Access to relevant accounts or tools will need to be authorized. If the agent runs into obstacles, it normally will come back and ask for clarification or more direction.

There is nothing very difficult or technical in this set up.

Here are some examples which  have something in common.

They involve multiple sources, multiple criteria and information that changes. They may also require repeated checking.

That’s very different from asking a question and getting an answer.

Perhaps that’s where personal agents begin to make sense.

If something requires an hour of searching through websites, building a spreadsheet, checking several alternatives and then doing it again next week because the information changed, I would happily hand much of that work to someone else.

If that someone else happens to be software, I don’t particularly care.

But it has to save me work and provide good and accurate answers..

If I save an hour doing the research but spend 45 minutes telling the agent what to do and checking whether it did it correctly, we haven’t invented a particularly impressive labor-saving device.

Then Come the Uncomfortable Questions

Once I started imagining actually using an agent, my questions changed.

• How difficult is it to set one up?
• Do I simply describe what I want done, or do I need to learn another piece of software?
• What information does it need from me?
• Can I watch what it is doing?
• How do I know when it has finished?
• What happens when it finds something it doesn’t understand?
• When does it have to ask my permission?
• Can it send an email? Make an appointment? Buy something?

And perhaps most important:

Where is the OFF switch?

That question becomes increasingly important as the agent gets more authority.

I might be perfectly comfortable allowing an agent to search apartment listings without asking me.

I would probably allow it to check my calendar to determine when appointments could be made’.

It could send an email in my name. Maybe I want to see it first.

Right now I would hesitate letting it make a $5,000 purchase or deposit?

Right now I would not let it move money between accounts anticipating the deposit.

With the right instructions, an agent could manage my files and potentially delete some files?

After some of my recent adventures with Microsoft, I am not sure I want anyone deleting files.

There is a big difference between having artificial intelligence advise me and giving artificial intelligence authority to act for me.

Maybe That’s What an Agent Really Is

The more I look at this, the less I think agents are primarily about artificial intelligence.

They are about delegation.

We’ve been delegating work to other people all our lives. We tell someone what we want accomplished, give them enough authority to do it, establish limits, check the results and intervene when necessary.

Now the “someone” may be software.

That doesn’t eliminate the old questions about delegation. It makes them more important.

• What am I asking you to accomplish?
• What are you allowed to do?
• What aren’t you allowed to do?
• When do you have to come back and ask me?

And how do I stop you?

Those sound less like computer questions than management questions.

Which may be why, after initially wondering what on earth I would ever do with an AI agent, I am becoming interested in trying one.

I have a couple of real jobs waiting for it. The next question is whether an agent can actually do them — and whether using one makes my life easier rather than simply giving me one more piece of technology to manage