Introduction

Hey everyone, I am Dan Charles. I'm here to talk today about some of the discovery that I've gone on whilst learning with AI and being a part of a community like this.

So obviously you can see here what the title of the talk is.

What I'm going to do is I'm going to go through sort of who am I, why am I here, and also then looking at talking through sort of what I've experienced, some of the science that seems to support my observations um and then moving through to maybe some practical

framework and examples of how you might be to make some good decisions about what you're doing with ai and um and how it's going to impact your processes your workflows and um everything you

Who I Am and Why I’m Here

do which i believe it's gonna be a big a big part of for a lot of people so who am i uh i am a i'm just a tech nerd really generally I'm a solutions architect engineer by trade at the moment I am incubating a startup business in the health tech space mental well -being it's called nurture

well -being so if you guys anyone wants to talk about sort of what the therapeutic relationship looks like how it might look like in the future we're currently doing feasibility and studies around that so absolutely love to talk to you guys over pizza and then favorite tech so as

Favorite AI Tools

As Andrew's covered off, we've got Claude Code there, big favorite for me, as it is for a lot of people at the moment. And that was before the whole anthropic Department of Justice piece around the war. So, yeah, good stuff there.

And Gemini as well. So Gemini is mainly because it's a part of my workspace. I'm Google Workspace. So it comes with it.

And also it seems that a lot of the hallmarks of Google's search engine have made their way into it. so in terms of research and all that good stuff and brainstorming really really powerful powerful

Building Agentic Systems—and Noticing Unintended Effects

powerful tool right so what do i notice so as i've been building with a lot of this technology i've sort of produced a number of agentic systems the first of which is called admin os effectively all of them are just insert name here operating system afterwards but admin os specifically was something that I produced and it's been a massive unlock for me with my startup

so I work full -time and I was thinking well how can I possibly manage all of the context calendar meetings invites notes follow -ups all that good stuff how can I follow that and and also hold down a full -time job and this has been a big unlock for me because I've effectively augmented a PA and it's it's worked it's

When Automation Reduces Learning and Recall

fantastic 1but what I noticed was some unintended consequences of that so I I could get everything I wanted from it but I found that often I was unintended I was I was learning maybe less and my recall was less uh and actually sometimes I was reading more than actually generating the content so although it's taking parts of that job it actually didn't make

me always the most effective at what I wanted to get out of it or the recall just wasn't there so I thought well let's look at the science is it me or is it is it just me uh or other people as

well and I started to look through or I stumbled upon effectively generator versus reader which I'll cover off in a little bit but what what that

What the Research Suggests: Cognitive Offloading and “Reader vs. Generator”

effectively looks like to start with is cognitive offloading so in short or simple terms most people will understand they store information other people so

for example bit of a simple one you'd go to Dave in IT to get your computer sorted if you had a problem with it then google came along so we offloaded the storage so it's no longer just in people but it's also on the computer so i can go to google now so that same journey might look more like i'm going to go to google to before i go to dave to sort my computer

maybe google's got some answers got some sources that are useful for me and then ai sort of shifts that again and instead of any of that processing that i'd normally do in terms of ingesting those sources and synthesizing an output or deciding what to do next ai is going to potentially reference sources for me and just tell me how to fix the computer and i wondered what does that

mean so in that sense offloading generation which is what i feel is is happening is very well documented so over the past sort of 50 odd years there's a lot of sort of science white papers as studies suggest that effectively once you become a reader of information and not a generator of it every metric by which we measure sort of recall and knowledge and and information in that

sense it reduces now so you can see it's not like it's not a crazy reduction but it's noticeable you know 20 a fifth is noticeable and that's what i felt i was experiencing so it's it's not just me i felt that you know i i there's something to this so i i sort of dug a little deeper and i and I thought about it from another perspective.

A Second Lens: What Happens When You Automate Other People’s Work?

I thought about it from what I'd done for others. So when I'd automated work in other spaces for other people that weren't myself, what did that do?

And you can see here like all the classics before weeks, after minutes, all the productivity hacking you could possibly hope for, but it seemed even more consequential,

this sort of lack of recall that I was experiencing for others and not in the same way.

Case Study: Automating an Art Trail’s Volunteer Processes

So for a little bit of context, context, I am on the board and advisory member of a committee for an art trail for my hometown. And as part of that, I'm sort of head of all things tech.

So it's my job this year to automate as much as possible, and bring things into line with perhaps sort of 2025, maybe even just 2010 would do.

So I digitized the onboarding form, workflows and processes. And then what I did was use agentic AI and direct replacement of volunteers.

So before, when we're putting content up and onboarding the artists that would all be manual even putting onto the sort of the content onto social medias onto the website and I automated all of it but what I didn't realize

The Hidden Benefits of Doing the Task Manually

was that although it's all better and the automation didn't fail what did happen was the person or the people the volunteers that were doing this still did the work or some of it because there was more than just the output the value wasn't just in what was being done in the content

being online it was also in the activity of using a browser going to a certain website logging in to places they might not often using it maybe even using a computer logging to social media for older people some older people that's not something they do very often so there was a lot of benefits there that were being missed by me and that that led me to sort of what what could

A Practical Decision Framework: Keep, Give, Share

i suggest about this so i came up with well hopefully as few syllables as possible but but keep give share as a way of splitting down tasks that you wish to automate or you think AI could augment or help and understanding where the value is or starting to break down what that value might look like.

So each question, I'm going to cover some examples, the same ones I've just gone through there, but each question we ask ourselves as part of critical evaluation should hopefully lean towards one of these categories is my hope. It certainly works for me.

It may even help you you split down the tasks further so you'll start to think about things in a more granular fashion again i'll cover that in the examples in a second um and then there's a little nice little piece

“Be a Centaur, Not a Cyborg”

down there be a centaur be a centaur not a cyborg um that's from a harvard study called the jagged frontier it's a great read um and that's an exact quote it's basically saying that 1cyborgs integrate ai at human and human capabilities at a granular subtask level while centaurs strategically strategically delegate between human and AI subtasks.

And that's just something that I personally align to.

Putting It Into Practice: AdminOS as a Worked Example

So what does that look like in practice?

So I spoke a little bit at the start about AdminOS. And there's a specific part of AdminOS I'm going to talk through now, which I'm kind of excited about.

But it's something a lot of people I think are coming to at the moment.

So it's self -learning, which means that it can implement new features for itself on my behalf, test them QA them and then push them and that's that's fantastic follows all the sort of software development lifecycle stuff the goodness that you want and I

Deliberate Tradeoffs and Human-in-the-Loop Quality

saw asked myself these questions before doing that and as a consequence it turned out quite well so I what I think anyway the first question you can see they're sort of did I learn some by doing this and that's where you can see

it sort of challenges a little bit because the active meta delivering features and functionality is a learning in itself so I'm sort of sharing if we If we go out to keep give -share there, I'm sharing.

And then you can see sort of would I notice if the output was wrong? Well, I wouldn't necessarily, in total honesty. But testing and QA should catch that. So I'm giving that away. And it's a conscious decision to give it away

because I understand that, like Andrew was talking about, with evals and stuff and iterative testing, you can get to a point where it might not work the first time, but it certainly will after three or four attempts, given the right testing and given the right involvement, engagement from from the human in the loop i won't cover all of these but you can sort of

you can see the point so i haven't got the time for all of them but um effectively if you think about it it goes quite well um it's still reading versus generating that's still a problem sure thank you um it's still reading not generating um so what i covered before still still applies but I've split it down so I at least understand the impacts of that as opposed to sort of walking into it as I'll cover in a second without necessarily having that insight

Revisiting the Volunteer Automation With the Framework

now I covered obviously a little bit of the the unintended consequences of the automation I did for the art trailer part of and you know it it would I think it would have made a lot more sense if I'd done and asked these questions beforehand and I think in future I will but you can see for

For number one, I didn't really question that enough, and as a result, learning from this experience, I would question and interrogate that a lot more, thinking about cognitive decline or thinking about some of the mitigating factors of these pieces, especially for an ageing population.

Then the second question, sort of ordinarily, no, but in this context, there were clear benefits that I didn't realise, um and quite like the same you can see the same pattern emerging which is that i didn't consider a lot of this um including the sort of what replaces the cognitive demand um and i didn't consider that at all meaning that the chair was still doing superfluous work but it was it was out of habit and want not out of value necessarily perceived value at that time um and then finally

Cognitive Exercise, Aging, and What We Might Accidentally Remove

I'll just cover off number five. Who needs the judgment in five years? Well, for this, it's cognitive exercise, right?

So what I'm taking away is maybe part of this person's life that they actually, although it's monotonous, it's repetitive, it's actually interesting to them in some way and useful and has those purposes to exercise them and mitigate things like Alzheimer's and dementia.

This is something that I found really interesting. I'm still sort

Implications for the Future of Work

tackling or wrestling with is an application of the same thing so my feeling is that at the moment we see a little bit of an issue in some companies whereby we see that some of the lower experience roles you can augment or replace them with ai now i personally feel that ai assistance will then mean that some of those some of the cognitive exercise you'd normally do and some of our seniors

Juniors, Skill Development, and the Risk of Over-Reliance

have today won't exist in a future of work so a recent Microsoft study shows that although Jenny Jenny I can improve worker efficiency it can also inhibit critical engagement and lead to over -reliance and then diminish skill I think that's especially important to consider for juniors and I hope that by

Conclusion

using a framework like keep give share and breaking down some of these tasks or some of these job roles early on it will mean that you can keep the education and cognitive exercise in some of those roles rather than having it be lost to time creating as you can see here a gap in that ladder up towards seniority that's me thank you very much for listening