S2 Ep 34: Botsitting - Where Your AI Time Savings Actually Go

Matt Wilkinson and Jasmine Gruia-Gray examine botsitting, the hidden time teams spend steering and fixing AI output, and what life science marketers can do to keep more of the time AI saves. Speaker turns have been lightly cleaned for clear transcription errors and false starts.

Introduction

Speaker: Matt Wilkinson [00:02]

Hi, Jasmine.

Speaker: Jasmine Gruia-Gray [00:03]

Hello, hello, what's going on, Matt?

Speaker: Matt Wilkinson [00:09]

Well, what isn't going on? There's so much, so much exciting stuff in the world right now. But I guess to jump straight into today's episode, I've been spending a lot of time working with my bots, making sure they stay on track.

What botsitting is and what the Glean data shows

Speaker: Jasmine Gruia-Gray [00:25]

Okay, well let's kick it off then. Picture the most efficient person on your team right now. They just used AI to draft an entire campaign in about four minutes. And they feel like a genius for roughly the next 60 seconds, right up until they start cleaning it up. Because as we always say, AI for direction, but humans for decision. There's a name for what happens next. And this is what Matt and others call botsitting. And the Glean Work AI Index 2026, a survey of 6,000 knowledge workers, found that AI saves people about 11 hours a week. That's a lot. Then they hand 6.4 of those hours straight back, just sitting there. Babysitting the thing that was supposed to save them the time in the first place. That's not a rounding error. That's more than half the gift returned to the register. And it gets worse once the same index zooms into the US specifically. More than a third of all the time anyone spends using AI at all, not just the saved hours, the total time goes straight to fixing what AI just handed them. Over a third admit they've shipped work they never actually checked. So somewhere out there, a confidently wrong sentence about your own product is already live and nobody caught it. Here's the part that should actually worry you, and it's the same report. 87% of workers say they're using AI. 75% say they feel more productive, and only 30%.

Speaker: Matt Wilkinson [02:37]

Yeah.

Speaker: Jasmine Gruia-Gray [02:46]

15% of organisations are performing measurably better because of it. Everybody feels faster, but almost nobody is. Everybody feels that they're getting lots done, but the accuracy of what they're getting done may not be what you think it is. So, Matt, if 11 hours go out the door and 6.4 comes straight back as cleanup. Where's the rest of that math falling apart?

Speaker: Matt Wilkinson [03:19]

Well, so it's really interesting and I think there's two parts to this. One is actually working with the AI and making sure it's staying on track. So the challenge is making sure that as you're working with it, it's great moving in a certain direction, but then you've got to engage with it so it's not just saving you that time in totality. But the other is going exactly where I think you expressed already, it's going into checking the work, correcting the work. And that's really because AI doesn't have taste. You know, it's read the entirety of the internet and nearly every book ever written. And yet it doesn't know when to stop. So it will hand you back something that's fluent and confident, and that's the dangerous part. And then somebody in your team becomes the person dragging it back to reality one paragraph at a time. You know, I find myself doing this as I expand our MCP server capabilities for delivering synthetic customers. The bots want everything to be perfect and over-engineer for edge cases that really don't need to be served, you know, solved for. And so you end up wasting time and tokens fixing problems that really don't exist.

Workslop and the real cost of cleanup

Speaker: Jasmine Gruia-Gray [04:40]

Okay, but how much is this actually costing somebody? Do you have real numbers?

Speaker: Matt Wilkinson [04:47]

Well, it's a lot more than it just being annoying. There's a Harvard Business Review piece on what researchers called workslop. You know, AI output that looks finished and isn't. And the number attached to that is brutal. Almost two hours per piece to clean up. That works out roughly nine million dollars a year lost in productivity for a 10,000-person company. Now that's not a rounding error. That's an entire budget line no one's tracking, sitting right there in plain sight. When you add that to the cost of the creation in tokens, you've got to be really careful in terms of looking at AI and seeing how you can use AI in a sensible way that really allows you to create real value rather than workslop that just ends up costing the business money.

The Sisyphus problem

Speaker: Jasmine Gruia-Gray [05:42]

So there's a Sisyphus analogy in here somewhere. Sisyphus being the person who pushed the boulder up the hill, pushing the AI output up the hill, if you will. It looks finished at the top and then it rolls right back down into someone's inbox as a correction job or a review job. Is that right?

Speaker: Matt Wilkinson [06:05]

Yeah. Yeah, and the cruel part of the myth holds up too. Sisyphus doesn't get stuck pushing up the hill because he's bad at it. He actually gets really, really good at pushing boulders up hills. The boulder just doesn't get over the top. Every time he reaches the top, the gods sent the boulder back down again. No matter how good he gets. And that's kind of botsitting in one sentence. You don't graduate out of it just with better prompting. You have to get better at creating prompts, prompt engineering, context engineering, and really understanding when and where to use AI for maximum impact.

Speaker: Jasmine Gruia-Gray [06:46]

So I have this crazy visual in my head right now of a whole team using AI and all these boulders being pushed up, never getting to the top, and just like this mass of boulders tumbling down onto the team.

Speaker: Matt Wilkinson [07:01]

Absolutely.

Creating for volume and the AI summary loop

Speaker: Jasmine Gruia-Gray [07:05]

So let sorry, come on.

Speaker: Matt Wilkinson [07:05]

It's I was gonna say, and I think one of the things that's really important here is that you know, the whole you know, nobody planned for that traffic jam. And one of the things that AI is really good at is creating these really big in-depth documents, but we're getting to a point where in many cases people are creating lots and lots of you know, really, you know, convincing sounding arguments that maybe need some work. And then people are taking those same AI outputs and giving it to their chatbots and asking their chatbots to summarise it for them because they don't have time to read all of the information in there. So we're actually getting to the point where we're no longer creating for the recipient. We're creating because it looks great. And then we, you know, the recipient's going, I haven't got time to read all of this. Let me let the AI summarise that. Now you take that to the extreme. Let's have a look at tools like Fyxer that promise to be able to look at, you know, handle your email inbox for you. Not something I've ever used and not something I would ever trust to an AI in that way, at least not yet. But can you imagine that we're creating so many emails, so much input that actually, you know, we're being bombarded by so much information and noise now that we actually we no longer want to worry about, you know, you know, that information. yeah, there's that there's that great saying that sort of said I'm sorry if I'd had more time, I would have written a shorter letter.

Speaker: Jasmine Gruia-Gray [08:50]

No.

Speaker: Matt Wilkinson [08:51]

And I think that's really where we need to get to in terms of AI and our usage. you know, I just managed to create a PowerPoint slide for somebody in a couple of minutes. I gave it a great prompt. It had a lot of context. It knew what it was doing. It was a fantastic way of delivering value and getting an idea out of my head and onto a and onto a slide. That sort of thing is a perfect example of. having a really, really clear brief and understanding what does somebody want to receive. It's a one slide, you know, it's one slide to help an executive make a decision. That's all we need to deliver. We need to make sure that we're getting away from trying to create volume with AI and we need to really focus in on the quality. And I think that's something that a lot of particularly people that are early on in their the adoption curve are, you know, are struggling with a little bit.

Will botsitting fade as teams mature?

Speaker: Jasmine Gruia-Gray [09:50]

So do you think that we're in this age of botsitting because most people are just cutting their teeth, so to speak, on AI. And their workflows are still clunky, even if the even in the case that they have a workflow, many just use AI in isolation. If we give it 18 months or so. do you think that the baby sitting burden will drop?

Speaker: Matt Wilkinson [10:25]

to an extent, yes. I mean, as the tools get better, as they learn what we're looking at, and as people learn how to use them, of course, things will improve. But the data says the core problem doesn't really fade with experience. It it seems to actually be going in the opposite way right now. And that is that the highest performing organisations bots it the hardest. They spend a bigger share of their AI time checking and correcting, not a smaller one. Now, the difference isn't that they need less supervision. It's that they aim the supervision better and they know exactly when to keep AI off a task entirely. Time doesn't make this go away. Skill just makes it more targeted. And maybe it gets us to the point where yes, we can absolutely deliver more, but the way that we're delivering and the quality that we're delivering, you know, we still need to have that human AI symbiosis to really deliver quality work.

Is this just good editing?

Speaker: Jasmine Gruia-Gray [11:27]

Okay, but to be sceptical, aren't you just describing a good editor? Writers have always reviewed drafts. is it really judgment or discernment that we're talking about now?

Speaker: Matt Wilkinson [11:47]

Well, it's both of those, but I think it there's also something a little bit deeper here. If we start looking at the use of AI and some of the promises that the frontier labs are making about, you know, the end of jobs and you know, we don't need to, you know, people will be, you know, we won't need as many people in the workforce. that seems to become a fallacy because we still need people to Be prompting the AIs to be working with them and to be delivering things with them. And there is this sort of structural tax quietly eating away at the return on investment that the CFO was promised by these AI investments because we still need to be making sure that we're not just creating workslop. You know, normal editing doesn't run $9 million a year in cleanup. You know, that's that's something that you know, really does impact the bottom line for any of these AI implementations. So when we say that, you know, when organisations say, hey, yeah, we're do we're doing really well on implementing AI, we've given people Copilot licenses, that's great. But actually if people don't know how to use them and they haven't, you know, they haven't got the skills to use them, they're likely contributing to that workslop rather than actually say, you know, seeing the benefits and getting ROI from the investment of the extra copilot licenses.

Context, guidelines and the brief

Speaker: Jasmine Gruia-Gray [13:20]

Okay, so a little moment ago you had said that really clear prompt engineering, really clear definitions of what you want the output to look like is super important to maybe reducing botsitting. What else is involved?

Speaker: Matt Wilkinson [13:44]

Yeah, so there are, you know, there's a there's a survey that suggests that organisations with better context that have architectured that in a way that allows the AI to better understand the environment and the tasks it's being asked to do are 52% less likely to ship work nobody can stand behind. So there is definitely a quality improvement in terms of Understanding not just prompting, not just how the AI works, but how the data that we have and being able to sh to share data in a really sensible and organized way with the AI. But it is tricky to do. And so one of the things that we really need to do as anybody that's you know involved in trying to help organisations adopt AI and use AI. Is to really give them an understanding of what does the AI really need to do a great job. It's a bit like trying to brief an agency. If you imagine trying to brief an agency on how to write an article, you might need to know who's going to be reading this, what's the persona that we're targeting. That's our grounded synthetic buyers. You'd need to be able to tell them what information do I need to do I need to share? What information can I share? What information can't I share? What's my brand tone of voice? What are the brand guidelines? There's a lot of information just in there to be able to bring this in. Then we've got to think about things like what's the business strategy, the portfolio strategy, what's the context in the outside world? Has there been a big shift in the last couple of weeks? That means the questions that our, you know, our buyers are going to have shifted. So there's a lot of that sort of context that needs to get that the AI needs to be able to do a really good job. But even with that, we do still really need to be able to create great guidelines that are able to the AI can then use to check its outputs against to say, okay, so if we're talking about this page, these are the answers. And now we're just giving an example of content creation because it's a lot, a lot easier than code creation.

Speaker: Matt Wilkinson [15:59]

But if we think about that in terms of trying to develop code for new products, we really have to be very, very careful as to the guidelines that we specified for the AIs to use.

Speaker: Jasmine Gruia-Gray [16:13]

You know, it's interesting that you're bringing up the brief because I think that a lot of marketers and even product managers have sort of shelved their briefs because they think, well, my AI is connected to my internal sources, and so I don't need to go to that extent anymore. And I think that the idea of bringing out your brief, looking through all the elements of the brief, making sure that you have the right persona, that you have a clear understanding of their pain points, that you understand what their buyer's journey is like. Doing that review up front, feeding that context to the AI can potentially save you time on the back end and reduce the amount of botsitting.

Speaker: Matt Wilkinson [17:11]

Absolutely. In fact, there's some pieces of work that I've been doing where creating an input brief allows me to get a lot, lot closer or receiving an input brief, allows me to get much, much closer to a you know, to an approved piece of work. Yes, there's a lot of context in there. Yes, there's a lot of really powerful information that's already been developed, but the brief is so important because the brief tells us where are we going. What direction do we need to travel in? And without that information, really we're just, you know, we're just saying hey to an AI, hey, create me a blog post about X. Create me whatever about X. It can do it. It'll do a convincing job and it will sound confident, but it's not going to do something that you know meets the brief in terms of what we're really looking for. And I think that's the big difference. We really need to put ourselves in the shoes of who's receiving the information. So if we're writing for our buyers, creating synthetic customers that represent them is a fantastic way of being able to make sure our messaging is landing with what we believe we understand about our buyers. But the same can be said from our internal stakeholders, the people that we're, you know, trying to communicate with internally. If we really understand what are the needs of finance, of you know, of product marketing, of R and D, of all of these different groups, we're in so much of a stronger place to be able to actually create something that's gonna help them.

Practical advice: quantify the tax and triage the work

Speaker: Jasmine Gruia-Gray [18:50]

Super. So if you could give three pieces of advice to someone listening to this episode, what would they be?

Speaker: Matt Wilkinson [18:59]

Well, I think the first thing I'd look at is actually start to try and quantify, and I mean giving a rough estimate, of actually how much time do they spend fixing AI output every week. No fancy tracker, just a number on a sticky note. Maybe use a you know, time tracking tool. But look at every time they've, you know, what are they doing with AI and how much time does it start do they spend fixing that things? Yes, of course, what you'd want to do is have a baseline of being able to compare well, how much did this task take before AI and then afterwards? But I think it gives us an interesting way of comparing, you know, contrasting, you know, what are we actually trying to fix? Because many times, you know, if there is a tax, if there is a specific set of challenges, there's likely to be a specific solution that will allow us to remove 80% of that problem, you know, with 20% of the effort. You know, but you can't fix an invisible tax. And so being able to name what the challenge is, is probably the first thing that I would advise anybody to do.

Speaker: Jasmine Gruia-Gray [20:08]

Anything else?

Speaker: Matt Wilkinson [20:10]

Yeah, so I would also sort tasks by how important they are. You know, if there are things that touch, you know, regulated claims or actual buyer relationships, then you need to make sure up front that you're spending, that you're really budgeting in time the time of approval. It's not just, hey, I can give this to AI and get the article or I can get the you know the piece of content or whatever it is. We need to make sure that we're budgeting the real review time up front instead of just discovering it the hard way just before launch. Yeah, we actually need to do this. So I think that's a really important thing that we need to be looking to do.

When fixing AI takes longer than writing it

Speaker: Jasmine Gruia-Gray [20:49]

So what's the one thing you handed to AI this month where fixing it took longer than it would have taken to just write it yourself?

Speaker: Matt Wilkinson [21:01]

Well, that's a great question because I don't think that I've found one of those recently. But then I've been so busy building context and building things that I think I've probably got past a lot of the time where fixing things takes longer than just writing it myself. But I will agree that sometimes the creative process can go way off the rails when you're working with AI. And so sometimes bringing that back can be a challenge.

Speaker: Jasmine Gruia-Gray [21:35]

I absolutely never have AI write emails. I find that's where I found that it takes way longer to rewrite the email after it's taken a stab at it. And so I've learned the hard way that I may as well just take it myself.

Speaker: Matt Wilkinson [21:54]

Yeah, I mean I sometimes will dictate an email to the AI, get it to try to tidy it up for me, and I'll give it some idea about what's important there. But I'll also give it a sense of why I'm struggling with it. So when I'm dealing with a high-stakes email, I will absolutely use the, you know, I'll consult with AI, I'll work with AI. But often it's more to avoid the blank page and also and sometimes as well to make sure that I'm not looking at something that's from a blinkered position. So I think that's where I would say, yeah, the email took a lot longer to create because I was using AI. But I typically I'll have ended up getting to a better result because of that collaboration rather than just relying on the first output that I receive.

Closing thoughts

Speaker: Jasmine Gruia-Gray [22:47]

Yeah, that's a great tip to tell it why you're struggling with the email so that it can avoid that and also explain how it's overcoming that in the message it's creating. Well, as always, Matt, this has been a spectacular discussion with lots of great best practices. Thank you so much.

Speaker: Matt Wilkinson [23:09]

Well, thank you. And if anybody's there listening to this, whether they're at work or home or just gently rocking the AI baby in the crib, let us know if you need some help with botsitting and we'll be happy to help.

Speaker: Jasmine Gruia-Gray [23:23]

And thank you again to all our listeners for taking their time to listen to another episode of A Splice of Life Science Marketing. Bye for now.

Speaker: Matt Wilkinson [23:33]

Bye.

Q&A

How do I find out how much botsitting my team is actually doing?

Start with one person for one week. Ask them to log every AI task in a shared sheet with two numbers: minutes spent generating and minutes spent fixing or re-prompting. No new software needed. At the end of the week, sort by fix time. The top three tasks usually account for most of the tax, and each tends to have one specific cause, such as a missing brief or no brand guidelines. Fix that cause first and measure again the following week.

Which tasks should I keep away from AI, or budget extra review time for?

Sort your AI work by consequence. Anything touching regulated claims or a live buyer relationship goes in the high-stakes pile, and its review time gets scheduled before drafting starts, with a named approver. Low-stakes work such as internal summaries or first-pass outlines can run with light checks. If a task in the high-stakes pile keeps needing heavy rewrites, take AI off it for now and use the model to critique a human draft instead.

What should go into an AI brief so the first draft lands closer to approval?

Treat the model like an agency you are briefing for the first time. Include the target persona and their pain points, where they sit in the buying journey, what you can and cannot share, brand tone of voice, and any market shift in the last few weeks that changes buyer questions. Add a short checklist the output must pass. Build this once as a reusable template for your most common asset type, then refine it each time a draft comes back wrong.

We have connected Copilot to our internal sources. Why is the output still generic?

A connected model can find your data. It still has to guess who the piece is for, what decision it should drive and when it is finished. That guessing produces confident, generic copy that someone then has to drag back to reality. Keep writing a brief for every significant asset, even with connected sources. Next week, run one asset both ways and compare how much editing each version needed.

How do I stop AI output creating more reading work for the people who receive it?

Define the recipient and the decision before you generate anything. If an executive needs to choose between two options, ask for one slide with the options and a recommendation, then cut anything that does not serve that choice. Length is where AI inflates effort for everyone downstream. A useful test before sending: if the recipient would paste it into a chatbot to summarise it, it is too long. Rewrite it at the length they would actually read.