The State of Productivity AI in 2025

Episode 14 explores how to use AI to help you with general productivity tasks, without needing a computer science degree.

RAG. Slop. MCP. Soulless. MOE. Autocomplete. Nano Banana. 10x. Coming for your job. AGI.

If you’re lost in a sea of acronyms and clickbait as you’re trying to figure out AI these days, join the club! At the risk of losing out on precious clicks, I’m not going to speculate on whether AI will deliver nirvana or extinction in the future. Instead, let’s talk about what AI can help you with and do for you today, in December 2025. And even more importantly, we’ll share how to get started so that you can get 90% of the benefits with tools you already have, without having to spend a week connecting things together, buying new subscriptions, and troubleshooting all of it.

Want a TL;DR? Gemini and Microsoft Copilot are built in to Google Workspace and Microsoft Office, they can already access your data, and get reasonably close to the best performance you’ll find even if you’re willing to do a bunch of work. If you have specialized needs, you may get some benefit around the edges from using the best-suited model, but for most things we do day-to-day, the easiest path is the best one. Try asking AI to help you brainstorm, summarize, critique, and make pictures!

What’s AI good for anyway?

AI is decidedly not human, so the tasks it can handle don’t map to our intuitive sense of how difficult they would be for us. It’s worse at counting than my five year old, but it can answer a question about a statistic buried in a document faster than a team of 20 library science PhDs.

AI geeks refer to this counterintuitive performance as “the jagged fronter” because it’s uneven and because the limits of AI are continuously pushing further out. So to find the places AI can help you in your workflow, you’ll have to be a little bit experimental. Some things will work great, others not so much. And the scope of what will work great is increasing quarterly.

The easiest starting point for using AI is brainstorming. AI is tireless, and it doesn’t (yet) get offended if you don’t like one of its ideas. You can ask for 20 taglines, expecting 10 will be terrible, 8 will be mediocre, but 2 might be great. The best thing about this kind of “breadth-first” search is that you don’t have to worry about hurting anyone’s feelings. You don’t owe it an explanation when you say “none of these work, try a different angle.” It will just start again, no feedback sandwich required.

AI is also excellent at shortening and simplifying dense information. In November, we used it for reviewing carbon capture research proposals. Every year, our company makes a donation to an academic research lab working on some kind of bleeding edge technology that could lead to economically-valuable carbon capture (this year, photosynthetic cement!) The proposals are only 1-2 pages, but they are hard to read for people who aren’t specialists in the field.

We asked Gemini to explain each proposal for a target audience of “smart college freshman,” and the dense technical papers became instantly accessible. It wasn’t perfect. There were some proposals where I read the summary and then read the actual paper and thought “that’s not quite the gist of it,” but it was pretty good, and plenty useful enough for deciding where to dive deeper!

Another strength is qualitative analysis. Earlier this year, we asked AI for help taking hundreds of survey responses and organizing them into themes. This used to require someone sitting in front of a spreadsheet for hours, reading each response and manually categorizing them, because users never describe the same problem with the same words. AI did it in less than a minute.

Again, at this stage, AI can do that initial sorting, but it might miss the nuanced cases where someone uses a word that seems like it belongs in one category but actually refers to something else entirely. You’ll still want human eyes on the results. But what used to take a couple days of laborious transcribing now takes only half an hour of human review.

And don’t forget pictures and videos and audio. Modern AI can work with visual content too, both in terms of “what might be wrong with this mechanical thing” or “what is this plant?” type questions. It can transcribe audio as well as most humans can, and read handwriting even better. It can also generate new pictures and videos, like the fun cat gifs that I generate as a way of motivating myself to write these blog posts!

One of my personal favorite use cases is using AI to write spreadsheet formulas. I don’t work in spreadsheets every day, so I’m never in the flow enough to remember the exact syntax for a complex lookup. The morning we recorded this episode, I needed to create a graph with a seven-day rolling average. Instead of spending ten minutes Googling formulas and syntax, I just told Gemini what I wanted. To get something that actually worked, I had to break it down into steps myself - first create a column with the rolling average, then verify it looked right, then create the graph, but it would have taken me a lot longer to do it the old way.

Where AI struggles is anything requiring nuance and subtext. A BCG study provides a quantitative example. 758 Boston Consulting Group consultants were asked to use AI to help them given realistic consulting work, like idea generation, problem‑solving, and writing. The consultants who used AI were more productive overall, but when the “right answer” required reading between the lines of what people said in interviews, AI led people astray.

AI also has trouble with things like precision math, even for things we consider basic, like counting. AI models semi-famously fail on tasks like counting the number of B’s in “blueberry.” But here’s a trick: if you ask it to write a script that counts letters instead of asking it to count them directly, it’ll produce working code in 30 seconds that you can run yourself. The model is better at coding the logic than performing the logic. Very different from us!

Incidentally, this blog post is not the work of AI (though the ones for Episodes 2-4 are). I am too persnickety about my tone and about making sure the ideas are represented on the page as they are in my mind to accept AI as a ghostwriter!

The Resume Screener

One of my favorite things we’ve done with AI came out of a hiring challenge.

As a 15-year-old company, we’ve come to value hiring people with a long enough tenure somewhere to have the experience of asking “what idiot did this work…” only to find that it was you, and you’d forgotten about it. Folks who’ve stuck around somewhere long enough to have a couple of those moments tend to have Chesterton’s fence embedded in their bones. They also tend to grant grace to other people’s work, since they can remember cases where their own fell a bit short.

In that spirit, we were looking for someone with at least three years of experience at the same company in a similar role. We posted the job to LinkedIn and got a flood of resumes, but very few candidates actually met that requirement. In the old days, sorting through the resumes one by one was a bottleneck in the hiring process.

Every morning, I’d open the folder, click through PDFs one by one, scan for dates, and do mental math. Nope, not this one. Nope, not this one. It drained energy before the actual evaluation work even began. Some (okay, fine, maybe most) mornings, I couldn’t face it at all.

So we wrote a script. Using Google Apps Script and Gemini, we processed all the resumes in our submissions folder and asked one simple question: “Does this person have three years of experience at the same company in this specific department? Yes or No.” The results went into a spreadsheet.

For the AI ethics folks reading this: we did not use AI to decide whether to hire anyone or to rank candidates. We only asked a binary, factual question. The AI was a filter to remove the obvious “nos” so we could focus our human attention on candidates who met our requirements.

It got it right about 95% of the time, which was better than our intern a few years ago did! It occasionally missed people who had been promoted within the same company (a good signal; we prioritized those people!) if they listed the two roles separately on their resume, so we spot checked it. But overall, it was remarkably good, and it made the mental burden so much lighter, all for free.

If you want to look at the script, we have a version here that you can copy and use! Feel free to ask Gemini for help making changes to it.

Who Benefits Most? The Novice Advantage

From our experience using it, it seems like AI helps less experienced workers more than experts. I get a lot more value out of Gemini in Google Sheets than our Director of Finance, who can write a formula as fast as he can type.

Some rigorous studies support this intuition! A study of 5,172 customer support agents at a Fortune 500 software company found that newer and lower-skilled agents were able to handle cases 30-34% faster to when given AI assistance. But the top-tier, most experienced agents saw almost no productivity gains, bundled with a slight decline in quality.

Much like my spreadsheet example, pairing the agents who were experts with a “B-player” assistant resulted in the agents taking as long to check and correct lower-quality work as it would have taken them to just do the work themselves.

Another research group called METR looked at AI’s effects on open source developers and found the same thing. They found that experienced engineers with five-plus years on a project were actually slowed down by AI assistance. The newcomers, though, got a significant boost.

Also interesting: the developers using AI thought they were going faster than they were without it. Their self-reported productivity was higher than their measured productivity. Much like keyboard shortcuts (users of keyboard shortcuts are slower than mouse users, but perceive themselves as being faster), there’s something about the type of cognitive load that working with AI produces that confuses our sense of time.

AI is like a great equalizer that pulls below-average performers up toward average, but doesn’t help (and may even hinder) people who are already operating at a high level. If you’re a great writer, ChatGPT might make your writing worse–and fill it with em-dashes! If your support team is super experienced and deeply knowledgeable about your product, they won’t see the gains that a new hire would.

Prompting Tips That Actually Help

Through our own experiments, we’ve found a few approaches that make AI more useful.

Breaking tasks into steps is probably the single most useful technique. Instead of asking for a final product all at once, ask the AI to do one piece, verify it looks right, then move to the next piece. This lets you catch errors along the way. Some people call working like this the “Centaur” model, where you’re the torso directing the work, and the AI is the legs making it go faster.

Asking the AI to act as a persona can help focus its responses. Instead of just asking “critique this podcast,” we’ve prompted AI to evaluate our podcast episodes with “You’re the producer of Freakonomics Radio—what are your notes for improving this episode?” (we ignored most of its advice; if you don’t like tangents, don’t listen!) These models are trained on the entire internet, so specifying a persona is like asking “which section of the internet should you draw from?”

Asking for emotional reads is underrated. “How does this email come across to a potential buyer?” gives you an objective tone check that’s hard to replicate, even from another person on your team. AI doesn’t have the background you have, or even the same shared experiences you’d have with a colleague.

For teams, setting up shared context projects is a game-changer. Using features like Gems in Gemini or Projects in ChatGPT, you can upload all your brand context, product details, and use cases one time, and use it over and over again. You can also create a shared “parent prompt” that you and your team can iterate on and improve. Then anyone on the team can run tasks without re-explaining everything from scratch. We built a custom GPT for sales discovery calls that knows all our features and use cases. You just type in a prospect’s email domain and it identifies their likely needs and generates a personalized demo script. No more generic “test email, test event” demos.

Finally, one note for early adopters. If you tried AI back in 2023 and weren’t impressed, it’s worth trying again. OpenAI’s data shows that people who started using their tools earlier use them less and trust them less than people who came later. That makes sense! If you showed up at the bleeding edge, tried tasks the models couldn’t handle yet, and got burned, you might never have looked back enough to see how far things have come. The models today are an order of magnitude more capable than the bleeding edge two years ago.

Tip of the Week: Connect Your Data

Our tip this week is simple: go beyond the chat window.

Go into your Google Workspace or Microsoft Office settings and enable the AI extensions for Gemini (Google, go to the gear menu –> Apps and toggle on Google Workspace) or Copilot (Microsoft, go to the gear menu –> Connectors and toggle on Outlook/etc.). You can grant permissions to both tools to access your actual documents, spreadsheets, and slides. This unlocks the ability to say things like “look at the project brief in my Drive and draft an email to the team based on the timeline.”

Your Workspace/Office subscription probably already includes some level of access to these AI tools. They already have well-tried paths to access your data, and because they’re part of the platform, they don’t present compliance/regulatory or security risks like hooking up a random new AI tool into your company’s data would.

You don’t need to know what RAG or MCP or MOE means, or spend a week connecting systems together to start getting value from AI today. The built-in options are good enough for 95% of the stuff you’d want to do, with only a minute or two of setup required.

Wrapping Up

You can get a lot of productivity benefits from AI you’re already paying for, with almost no up-front time investment, today.

AI is brilliant at some things and clumsy at others, and the line between them doesn’t follow human intuition about what’s “hard” or “easy.” So you’ll have to try a few things out to find the parts of your workflow where AI can really make a difference. If you don’t know where to start, try brainstorming ideas or shortening and aggregating information from documents.

Don’t wait for the perfect AI. The current models are powerful enough to be plenty useful already. And if you tried it a couple years ago and gave up, give it another shot. Things are different now.

To learn more, listen to the full podcast episode.