Your company will issue you an AI agent like it issued you a laptop. Except this one will learn how you think and start taking work off your plate.

The cliché is Jarvis from Iron Man, but honestly, more like Donna from Suits. Part chief of staff, part minion. And yes, they literally tried making a digital Donna in the show.

The part that interests me is how well Donna understands the person she works with. She knows what matters, sees what’s coming, and handles things without needing the whole situation explained every time. That’s a pretty useful starting point for thinking about a personal agent.

As my team builds toward a more AI-native org, I keep coming back to two types of agents: personal agents and shared experts. I think the relationship between them will shape a lot of knowledge work beyond developers.

Your agent will learn the job with you

Your personal agent will know how you think, what you’re working on, and how to get things done for you. It’ll learn which details you care about, who you work with, and when you need to be pulled into a decision.

It could help you onboard by walking you through real work in context. As you settle in, it could prepare responses, follow up on things you’ve delegated, and work in systems on your behalf. Eventually, you might realize you haven’t personally logged into half of them since last year. It might even cover a meeting for you, if the job is to gather information and bring it back.

Imagine asking it to get a vendor proposal ready for a decision. It already knows why you need the vendor, what you’ve discussed, and what’s still missing. It gathers the current information, gets help reviewing the terms and costs, and comes back with the questions you actually need to settle. Once you decide, it handles the follow-through within the authority you’ve given it.

You’d still be involved. But you wouldn’t need to personally move every piece of information between every system and person along the way.

Shared experts are an organizational responsibility

Your personal agent shouldn’t have to become the company’s best contract reviewer just because you need a contract reviewed.

That’s where shared expert agents come in. Legal, finance, design, engineering. Digital employees the organization manages, with deep company context and internal experts responsible for testing and continuously improving them. Your personal agent would call on them like you would.

In that vendor example, your agent could ask the shared contract expert to review the terms against the company’s current guidance. A finance expert could check the commercial assumptions. Your agent brings the results together in the context of your decision, including anything that still needs a person’s judgment.

Calling something a legal expert doesn’t give it authority to approve a contract. The organization still has to decide what it can do, what needs review, and who owns the result. The same is true when a human employee joins a team. A job title is the beginning of that conversation.

Consistency is why this matters

Give everyone the same data and tools, and you still get wildly different results. Everyone builds differently. Someone writes a great prompt. Someone else connects the wrong folder. Another person makes something useful for themselves that falls apart when anyone else tries it.

Some setups expose information, confidently get things wrong, or feed slop back into company knowledge. Then the next agent retrieves that slop and treats it as context. Pretty quickly, it gets hard to tell where the original information ended and someone’s generated interpretation began.

Shared experts give the organization a standard it can actually manage. The people responsible for the work can test the same kinds of requests, inspect failures, and improve the expert everyone relies on. When company guidance changes, there’s a clear place to update the behavior and check whether it still holds up.

That concentrates responsibility too. A shared expert that gets something wrong can spread that mistake much further. Someone needs to own it, know what good looks like, and be willing to pull it back when it stops meeting that standard. We should expect ongoing work here.

There’s still room for people to build and experiment. I want that. But a useful personal setup and something the whole company depends on need different levels of care.

Most people shouldn’t have to become AI hobbyists

I don’t think most knowledge workers will need to know about context windows, token spend, prompt engineering, or which model to pick. They’ll work with their agent like a reliable remote coworker. Explain what they need, give feedback, and get on with the job.

People will still need to learn how to delegate, check work, and recognize when something needs their attention. Those are familiar skills. We already expect them from people working with other people.

The technical decisions still matter. Somebody has to make them. I just don’t think that somebody needs to be every employee, every time they want help getting through the day.

Trust will have to build through use. An agent that drafts a message well hasn’t automatically earned permission to send it. One that can explain a budget hasn’t earned permission to change it. Responsibility should expand as the work holds up and the person delegating it gets comfortable.

If you build SaaS, think about who will use it

A lot of your interface exists so humans can move information around. Find the record, check another screen, copy something over, update a status, tell somebody it’s done. Agents won’t need all those screens.

People will still need ways to understand what happened, make decisions, and correct mistakes. But I think a lot more of the routine work will happen through personal agents and shared experts calling on the capabilities underneath the interface.

For a founder, I’d think hard about where the product fits. A personal agent might want broad access to a system so it can carry a task through. A shared expert might need one very specific capability that your product does exceptionally well. Either way, it needs to get the work done and return something the agent can use, including a clear failure when it can’t.

That’s what I see us building toward: a personal agent that understands you, shared experts the organization can stand behind, and products useful enough for both to call on.

Go wide or go deep. Just be really useful to at least one instead of bad at both.