Latest / In The Loop / How To Decide What To Automate With AI For Your Team & Customers

Transcript

How To Decide What To Automate With AI For Your Team & Customers · In The Loop

• 0:00
  So every single week, it feels like there are new AI features,
• 0:04
  automations and transformations cropping up all of the time.
• 0:08
  But many of the conversations that I've been having week on
• 0:11
  week is how to even decide on where to start.
• 0:15
  We have the pleasure of speaking to hundreds of different product
• 0:17
  teams, teams who are trying to simultaneously transform their
• 0:21
  entire product process and the products themselves.
• 0:26
  And I just say it's a really big question of how do we decide and
• 0:29
  where do we start? And there's so much pressure.
• 0:32
  Your CEO wants things to change, your team wants things to
• 0:35
  change, your customers want things to change.
• 0:37
  The board wants things to change.
• 0:39
  Marketing needs things to change, so organisations right
• 0:42
  now are under so much pressure to change.
• 0:44
  And the problem is that most advice is either incredibly
• 0:47
  abstract or is just pure hype. So today I wanted to do
• 0:51
  something a little bit different.
• 0:52
  I wanted to talk through a fantastic piece that was written
• 0:55
  by one of my favourite substance called the Department of Product
• 0:59
  and inside it, it proposed a really interesting framework for
• 1:02
  helping people decide on what classifieds as a good automation
• 1:08
  opportunity for either their users, inside their product or
• 1:11
  for their team. And as I discussed this article,
• 1:13
  I'll be sharing as many insights as possible based on what we're
• 1:16
  seeing and hearing day in and day out.
• 1:19
  This is in the loop of Jack Hart.
• 1:21
  I hope you enjoy the show. See this This week's going to be
• 1:35
  a little bit different because I've been really hoping to do an
• 1:39
  episode like this for some time. I have to say, the amount of
• 1:42
  conversations we have every single week is quite staggering,
• 1:45
  and it's not an easy thing to decide on.
• 1:49
  How or where do we automate first?
• 1:52
  And that can be, as I said, for a user or for your team.
• 1:55
  We're in the very fortunate position to have a broadly AI
• 2:00
  first roduct development Rocess all the way from product all
• 2:03
  through to engineering and roduct marketing, which has been
• 2:06
  a heck of a lot of work to get in place.
• 2:08
  And our entire product is about enabling technology companies to
• 2:14
  have a conversational interface into their customers and their
• 2:17
  users. So we're we're quite in a in a
• 2:19
  really unique position to kind of sit at the intersection of
• 2:23
  how to both understand this opportunity and actually do
• 2:25
  something about it. And so I guess this piece by the
• 2:28
  Department of product I found perfectly timed and really
• 2:31
  interesting. And it opens up to a reference
• 2:33
  to Benedict Evans, which is he he's a tech analyst and his
• 2:37
  latest presentation they gave this tech investor, Benedict
• 2:40
  Evans essentially said that right now automation is a is
• 2:44
  biggest superpower. And he suggests that whilst it's
• 2:47
  not as exciting as AGI, that automation alone is a big, big
• 2:52
  deal. It's the boring cousin of AGI to
• 2:55
  artificial general intelligence. And he makes a good point, which
• 2:58
  is barcodes and databases will not sexy, but they completely
• 3:02
  transform retail. And so I'd argue we're seeing
• 3:05
  almost the exact same thing happening with AI and AI
• 3:08
  features in products right now. The winds aren't necessarily
• 3:12
  coming through big model breakthroughs every six months.
• 3:15
  They're coming from the application layer.
• 3:18
  So the top ends are well above the model the the technology
• 3:21
  people are interacting with. That takes away frustrating and
• 3:25
  tedious tasks from people's day-to-day lives.
• 3:27
  But with that comes a really key tension.
• 3:30
  And as the Department of Products said in the article,
• 3:32
  just because you can't automate something doesn't mean you
• 3:35
  should, because the outputs from these automation efforts is
• 3:40
  often far from perfect in many, many cases.
• 3:42
  That judgement call on what to automate, maybe for your users
• 3:45
  or for your team, is what this article discusses and what this
• 3:48
  episode will cover in depth today.
• 3:50
  And I see this is a major problem to solve.
• 3:52
  We're actually building an agent in 2026 that helps people with
• 3:55
  this because a lot of the time these are really new ways of
• 3:59
  looking at technology and process.
• 4:03
  Because somehow people have to also understand how AI and
• 4:06
  agents work, what they're really good at, what they're not good
• 4:08
  at, and then understand how to RIP apart an entire process or
• 4:12
  product to improve it with those strengths and weaknesses in
• 4:17
  mind. So the article starts with
• 4:18
  something called the verifiability principle.
• 4:21
  So before we get into the framework itself, the article
• 4:24
  she started by sharing a really interesting insight from an open
• 4:27
  AI Co founder. His name is Ilya Sutskieva, So
• 4:32
  very famous online, got a massive YouTube account, very
• 4:35
  big on X. And he said, in his view, a good
• 4:37
  candidate for AI automation is a task that is easily verifiable.
• 4:42
  You know, the more a task or a job is verifiable, the more
• 4:46
  amenable it will be to automation.
• 4:49
  So it's a very simple principle, but it's very powerful when you
• 4:51
  think about it, because the core question is, can you tell if
• 4:55
  it's worked? Now, there are many tasks for an
• 4:57
  individual that they can themselves take that judgement
• 5:00
  call and make that judgement call.
• 5:01
  But for some big enterprise organisations, they're going to
• 5:05
  be very obsessed with that verifiable or verifiability
• 5:09
  principle. So if you think about it from a
• 5:10
  user's perspective, if AI say auto categorises all their
• 5:14
  expenses, they can verify that really quickly and instantly.
• 5:18
  You know, they can literally say, was this right or wrong?
• 5:21
  But if AI writes, let's say, personalized recommendations,
• 5:24
  it's very hard to know if that was going to be right or wrong.
• 5:27
  They're in very subjective territory.
• 5:29
  So that part of verifiability is really important because if the
• 5:32
  AI is not good at doing something that is easy to say
• 5:35
  that's right or wrong, then sometimes it will not be a good
• 5:40
  candidate for automation. Now the new ones here is how
• 5:43
  good the process is around that AI.
• 5:46
  However, this does also create attention because most of the
• 5:50
  things that are hard to verify is also where you're going to
• 5:54
  get the greatest gains in many way from automation.
• 5:58
  So something that we learn at mindset is you have to really
• 6:01
  build process around these things because ultimately
• 6:04
  judgement becomes the key success criteria of a human in a
• 6:08
  modern day process. So we need people that can make
• 6:11
  really good judgement calls on the quality of an output from an
• 6:14
  AI. And that really is going to give
• 6:17
  you an A good indication of how much human in the loop is needed
• 6:20
  here. Because most of the time an
• 6:22
  agent with a bunch of, let's call it tools, so things it can
• 6:25
  do in the world and the ability to do good reasoning can do most
• 6:29
  tasks to a pretty good standard. Now, if you want to fully
• 6:32
  automate it, so you just don't want to look at that again,
• 6:35
  you're happy to say go, never speak to me, go and do that
• 6:38
  thing. Then the verifiability principle
• 6:41
  becomes really important. Otherwise you need that human in
• 6:44
  the loop part of that process involved.
• 6:47
  So it's a lot simpler a lot of the time for your team because
• 6:49
  you are a member of that team and a member of that process.
• 6:52
  Now this gets even harder when you're trying to think about
• 6:55
  what type of automation experience should I offer to my
• 6:57
  users or my customers. And so this article then goes a
• 7:00
  bit deeper into whether this should be a product feature or
• 7:03
  something just for an internal process automation.
• 7:05
  Now to help decide whether you want to create AI guess a
• 7:08
  product for your users or your customers inside your existing
• 7:11
  product. What you should be thinking
• 7:12
  about is first, what triggers this AI?
• 7:15
  Is it a user action? Is it an event that happens in
• 7:18
  the platform? So someone hovers over
• 7:20
  something, clicks on something, is an event?
• 7:24
  Is it a certain signal? Second, what relevant data is
• 7:28
  going to be there for collected and what processing is the AI
• 7:33
  going to do? So is it going to be classifying
• 7:35
  it? Is it going to be generating
• 7:36
  something or writing or creating something?
• 7:38
  Is it going to be predicting something?
• 7:41
  And then 4th is some form of when I say logic or instructions
• 7:44
  on how it decides what happens next.
• 7:48
  Should it auto execute that task?
• 7:50
  Should it come back and ask the user?
• 7:52
  And finally, therefore, what is the end action that happens or
• 7:55
  what is that end outcome that happens?
• 7:57
  We actually run sessions with companies that are coming
• 8:00
  through our sales process on essentially trying to create a
• 8:02
  job description for this AI idea.
• 8:04
  You know, what is, what are the inputs or what are the user
• 8:07
  questions? What situations is it going to
• 8:09
  trigger in or appear or suddenly step forward and say, I'm, I'm
• 8:14
  here to do something. Therefore, what magical skills
• 8:18
  does this AI thing need? So it needs access to user
• 8:21
  memory or data. It needs access to some other
• 8:23
  system in a third party application.
• 8:26
  A good example here is taking predictive shopping carts.
• 8:29
  You could auto fill a shopping cart when a user opens an
• 8:32
  application and then send the items are already in there ready
• 8:36
  to be checked out based on some sort of behaviour or action they
• 8:39
  were doing elsewhere in the application.
• 8:42
  Or you could just simply show suggestions to them and let them
• 8:45
  approve each one one by one. Or a middle ground auto add high
• 8:49
  confidence items and then suggest medium to lower
• 8:53
  confidence items to the user to say, do you want this in your
• 8:55
  cart, yes or no? All of that could have happened
• 8:58
  based on something triggering in the application.
• 9:01
  The AI using data processing and generating or suggesting things
• 9:06
  and then either doing a full human in a loop or just doing
• 9:08
  itself. And those design decisions is
• 9:11
  going to dictate whether the user feels helped or bulldozed.
• 9:15
  With that in mind, how do we understand whether this
• 9:16
  candidate or idea is a good opportunity to automate
• 9:20
  something for maybe my users on my team?
• 9:22
  Now, the framework suggested in the Department of Product
• 9:25
  article essentially rates opportunities across 3
• 9:29
  dimensions and scores them from one to three.
• 9:31
  You've got safety, volume and verifiability.
• 9:35
  So safety is about user in part when or if you are wrong.
• 9:39
  So if the AI suggests the wrong product substitution and the
• 9:43
  user catches it before check out, then it's annoying but it's
• 9:47
  fine. So therefore it's got a simple
• 9:49
  safety score. If it also approves something
• 9:53
  and just takes an action by itself and cost people lots of
• 9:56
  time or money or potential risk to their health or safety, then
• 10:01
  therefore you've damaged trust and it's not the the safest of
• 10:04
  ideas. So I'd score 3 for anything
• 10:07
  that's just a simple easy to catch low in part low stake
• 10:10
  error or and score one for anything that is going to hurt a
• 10:15
  user before they can intervene. The second part of that
• 10:19
  framework is volume. So volume is about whether this
• 10:21
  matters enough to build. If users do this task
• 10:24
  constantly, every session, every day, and they take lots of time
• 10:29
  doing it, then it's a fantastic opportunity to automate this
• 10:31
  thing for your user or customer or team.
• 10:34
  If it's a once a month thing, maybe you know, it's not really
• 10:37
  worth the investment. So if it's I would score 3 for
• 10:41
  something that's constant daily actions and then I'd score one
• 10:45
  for rare ones. And 3rd and final is
• 10:47
  verifiability. So this is about feedback loops.
• 10:51
  Can users easily tell if the AI got it right?
• 10:55
  So maths for example is a simple yes or no, and if it's objective
• 11:00
  and simple to say yes or no, I'd score it a three.
• 11:04
  However, personalized recommendations are subjective.
• 11:07
  Therefore I might score it a little bit lower to A1 or A2,
• 11:11
  and the maximum score of these things might be 9.
• 11:13
  Maybe you just want to not Add all the scores up and you just
• 11:16
  want to assess each one of them within their own dimension.
• 11:20
  I'm always not. I'm not always a massive fan of
• 11:22
  trying to add different ideas and scores up and somehow
• 11:26
  finding, you know, the best scorer out of all of them.
• 11:28
  Because some ideas you'll think that's actually just such a
• 11:30
  great opportunity, even if it is not that safe or even if it's
• 11:34
  not that verifiable. But because of the volume, I was
• 11:37
  going to go for it anyway. So the output of this for you
• 11:39
  and your team is going to be a classification of something of
• 11:44
  automated. Now make it just an assistance.
• 11:47
  So AI assistance versus full automation or just something
• 11:50
  that's labeled as not ready yet or unsure for a user facing
• 11:55
  features. Most will land in the middle to
• 11:57
  be honest. You know it'll be an AI
• 11:59
  assistant zone and this is where design really matters.
• 12:02
  AI assistant I would describe as AI doing the work but the user
• 12:06
  having control before anything is finally committed.
• 12:09
  You know, if it books my entire holiday, I'd rather it check in
• 12:13
  with me first. So the article gives examples
• 12:16
  that score of things that scored six out of nine.
• 12:18
  So things are AI meal planning based on my health objectives or
• 12:22
  AI generated meal plans based on the things that are in my
• 12:26
  shopping cart. These require taste in a
• 12:29
  preference judgement and the AI could probably do a good job of
• 12:31
  recommending it, but probably not just doing it all for me.
• 12:35
  And this leads on to thinking about therefore, why will things
• 12:37
  fail? If we now know why they succeed,
• 12:39
  why will they fail? Well, if you create an AI
• 12:42
  automation process change or something for your customers and
• 12:45
  users, well, it's going to fail because of a lack of trust.
• 12:49
  So early mistakes are often very visible.
• 12:51
  So maybe users will just start ignoring the AI suggestions or
• 12:55
  the AI actions that it can do for them, and just therefore
• 12:59
  adoption will just flatline. Maybe there's going to be
• 13:01
  frustration with edge cases. So it might work great 80% of
• 13:05
  the time, but that 20% where it's wrong is infuriating.
• 13:09
  So users therefore feel like they have to double check
• 13:11
  everything all the time or it's just really creepy.
• 13:15
  It knows a little bit too much. Or if it's too aggressive and
• 13:18
  users feel like it's in their face or they're being surveyed
• 13:21
  all the time. And these three things that
• 13:24
  could go wrong with, I wouldn't see this as a we shouldn't do
• 13:27
  this at all. These are just things that you
• 13:29
  have to design for and think through again, whether you're
• 13:32
  going to make this full automation and AI assisted thing
• 13:35
  and make sure that the user's very careful and aware of the
• 13:38
  things that this AI can and can't do, or just labeling
• 13:42
  something is not ready yet. So yeah, to wrap up, this is a
• 13:56
  little bit different. We went through really practical
• 13:58
  tools. We'll also share the article in
• 14:02
  the notes to the Department of Product because it's a really
• 14:05
  great framework. But right now, I think the only
• 14:08
  thing that you can't do is doing nothing at all.
• 14:11
  Because right now the world is moving so quickly and every
• 14:14
  single leader must be completely ripping apart their processes
• 14:17
  and thinking about every opportunity to automate inside
• 14:20
  their product for their users 24/7.
• 14:23
  And this should be the only thing you're focusing on.
• 14:25
  So I'm recommending you have a read of that article and start
• 14:28
  to reflect on what you're going to automate in 2026.
• 14:31
  Anyway, that's it for me this week.
• 14:33
  I hope you enjoy this episode and I'll see you next week.