Love it or hate it, AI has entered the workplace in a big way. As an engineer, it fascinates me, and as a result I’ve experimented with it a lot. Here’s some of my key high-level takeaways for getting the most out of it.

Key takeaways:

1. Treat it like a ‘drunk professor’. Sounds smart. Mostly correct. But every now and again slips in something untrue and it’s hard to tell.

2. Think about your interactions with it like a barbell. Put your weights (time) on the ends. Clearly defining the work and carefully reviewing it. Let it handle the middle.

3. Adjust the weight distribution between defining the work and reviewing the work, depending on the task.

4. For long tasks, “define the destination, not the route”. It’s good at rerouting if you give it a clear end point.

5. Share the larger reason behind the work. The “why” helps it make decisions as it works.

6. Apply the same thinking to automating workflows with AI. Maintain human definition at the beginning and review at the end.

7. Human connection, thoughtfulness, and creativity have never been more important.

People at both ends

Working with AI is a lot like working with a drunk professor. It sounds smart and it’s mostly right. But now and then it slips in something untrue, in the same confident, intelligent tone as everything else, so the mistakes are hard to spot. It will also talk at length about whatever you ask, whether or not you asked the right question. This is why I find it’s critical to keep people at both ends of the work. Defining very clearly what you want before AI starts and checking what comes back before you ever say it’s done and share it. AI is good at lots of things, but critical thinking and judgment still need to be owned by the person using it.

I picture this as a barbell, with the weight, meaning your time and attention, at both ends, with AI in between. As a rough split, the first 20% of the work is human. Work out what you’re trying to do, who it’s for, what a good result looks like and what context matters. The middle 60% is where AI helps most, because it’s good at sifting a lot of material, gathering it, structuring it and filling gaps. The last 20% is human again. Read what came back and ask whether it captured what you meant, used the right context and suits the audience who’ll ultimately read it or use it.

I came to the barbell idea on my own. When I looked into it, I found that others had, unsurprisingly, been talking about it for a while. Julie Holmes, who speaks and writes about AI at work, calls it the 20-60-20 rule and wrote it up in February 2025. She notes that the split can move toward 10-80-10 as your prompts and templates improve, but I’d propose something a bit different…

More of a seesaw

A barbell carries the same weight on both ends, and an even split is a good place to start. Holmes’s 10-80-10 keeps the ends even and makes both lighter. The change I make more often is moving time from one end to the other, a bit more like a seesaw. For a given task, I keep the total time invested the same, and move it between the two ends to suit the task.

For example, for a task where I’m not fully sure of the end shape, I spend less up front and more at the end (or step through it, more on that later). For a task I know it’s done well before, I put more effort into defining a reusable skill and less time into reviews.

Long tasks are long levers

The length of the task changes the balance too, and as the models improve, AI can take on longer tasks. METR, a nonprofit that tests AI models, measured how big a task an agent could finish on its own about half the time. It sized each task by how long a skilled person would take. That size doubled about every seven months from 2019, and faster from 2024. It’s only accelerated from there. I frequently prompt 2+ hour tasks now, where in the past I rarely went above an hour. Or if I did, the quality degraded significantly.

A seesaw is a lever, and a long agent run makes the board very long. You set the brief at one end, and the result lands at the other, where you review it. The longer the board, the further a small tilt moves the far end. A slightly wrong assumption in the brief can land the result a long way from where you meant. A close review can still catch the miss, but by then the work has gone a long way in the wrong direction.

So the length of the task sets the total time, and as the task gets longer, tip more of that total toward the start. A short task needs a short amount of time on either end, while a research project an agent works on all afternoon needs much more, and most of it belongs in the brief. A quick brief with a close review works best on short tasks.

You can also shorten the lever by checking in. Over several sessions, the work becomes a chain of shorter levers, and each check-in is a chance to steer. On a big piece of work, the first session might go entirely on defining the work. Even with check-ins, keep the whole journey in mind, because small turns in each session add up. There are pros and cons to steering, and it requires other tricks like using documents, project management tools, or GitHub issues to track the overall to-dos. But it can be helpful if you are still trying to figure out the best path to a solution.

Give it a destination

No brief gets every assumption right, so on a long run some small miss at the start is likely. Careful aiming only gets you so far, so put your time up front into the destination. Describe where you want to end up in enough detail that the agent would recognize it. In my experience, agents are good at rerouting around a problem as long as they have a clear end point, so let them find their own way there, with guardrails where they need them.

Think of the difference between giving someone directions and giving them a destination. Directions say go up, then right, then up, then right again. They work until the driver hits traffic or a road closure. After a small detour, a driver can usually find their way back to the route. A big detour that you didn’t plan for is much harder to recover from. A destination, with the address and a description of what the place looks like, lets them work around whatever comes up and still arrive.

Tell it why

A destination still isn’t the whole brief. An agent will follow the task you give it, but if you don’t explain why the task matters, it can misread what you want. The reason helps it make the many small decisions along the way.

Take the drive example again. Should the driver go slowly and carefully, or get there as fast as they can? Are they mapping a route that lots of people will drive many times, so they should look for the most efficient one? Or is this a quick trip to see if the place can be reached at all?

Say you’re asking AI to help draft a proposal. You can do all the right things to define that proposal. But if, in the back of your mind, you’re also debating whether to make this format your new template, the output can be very different depending on whether you mention it. If you don’t, AI won’t know. If you do, it might build the proposal in a way that’s easier to reuse for other kinds of projects. Then, when you review it, you may come to a different conclusion on whether it should be the new template.

Automations are the same idea

When you automate a workflow so AI does that kind of work again and again, the same thinking applies, with people defining the work at the start and reviewing it at the end. How that looks depends on the workflow.

Sometimes the people sit inside every run. A person makes the judgment call at the start, the automation does the middle, and a person reviews what comes out. In outreach, a person might pick which accounts to contact and say why, the automation gathers research and drafts a first note, and a person reads and edits each note before it’s sent.

Other times, most of the thought goes into the system before it ever runs. You work out what it should do, who it’s for and what good looks like, and then you build it. Someone still reviews what it produces to make sure it’s right, especially while the system is new and whenever its output goes to another person. As the system proves itself, you can spend less time on that review. I think the people and teams who find that balance for each task and each workflow will do well, and they’ll be using AI responsibly while they do it.

What’s left for people

The two ends matter most when the work reaches another person, and in sales and marketing it usually does. Anyone can now send cold emails and LinkedIn messages in bulk for very little, and both channels have become saturated. As someone with an inbox and a LinkedIn profile, I find that annoying, and buyers seem to agree. In a Gartner survey, 73% of B2B buyers said they actively avoid suppliers who send irrelevant outreach.

Surprisingly, we’re seeing cold calling make a comeback. As someone with a phone, I find a call that has nothing to do with me annoying too, but when it’s relevant and the caller is respectful, I actually don’t mind it. But only to a degree, and I’d expect many people would feel differently. The same is true for conferences and webinars where we are seeing something similar. For higher price points, people still want that personal relationship.

So when you do reach out, don’t just send a template, and don’t rely on AI for the personal part. Consider picking up the phone (respectfully!), going to an event, or hosting a webinar (that's useful not just a pitch). People can tell the time and effort you put in, and their likelihood of responding and buying is linked to that effort.

People bring what the two ends need: creativity, a sense of where things are going, empathy, thoughtfulness and the human connection that builds relationships. So yes, AI has reached the workplace and is likely here to stay. But I think those who use it responsibly, and who lean in on the human elements, will ultimately do better than those who simply "token max".