Latest / In The Loop / EP16 - The Top Five AI Features SaaS Companies Are Shipping In 2025 (And Why They Work)

Transcript

EP16 - The Top Five AI Features SaaS Companies Are Shipping In 2025 (And Why They Work) · In The Loop

• 0:00
  By the end of this episode, you'll understand the five
• 0:02
  categories of AI features that technology companies are
• 0:05
  actually launching in 2025, from small improvements to AI native
• 0:10
  applications to AI first products.
• 0:13
  As inspired to do this episode after a great piece by Rich
• 0:15
  Holmes on the Department of Products, Substack.
• 0:17
  It was published last week and analysed 25 big product launches
• 0:21
  by massive companies like Google, Notion, Microsoft,
• 0:24
  Stripe, LinkedIn, and many, many more.
• 0:26
  I've then shaped this thinking a little bit based on what we're
• 0:29
  seeing in the market today because we're working with many
• 0:32
  technology companies and helping them build AI features faster.
• 0:35
  This is In the Loop with Jack Horton and hope you enjoy the
• 0:38
  show. The first category that I'm
• 0:50
  going to discuss is called UX enhancers or embedded UX
• 0:54
  enhancers. And I guess what distinguishes
• 0:56
  this category from other types of features is that they're
• 0:59
  either often completely invisible in the background and
• 1:02
  just pop up and appear at the right moment.
• 1:04
  So they'll appear contextually with small pop ups and widgets
• 1:08
  that just help me do something or solve a bit of a pain point
• 1:11
  that I would normally experience when using that type of
• 1:13
  technology solution. And as Rich Holmes broke down in
• 1:16
  his piece, these features are not trying to get users to
• 1:19
  engage with the products in a completely new way.
• 1:21
  They're just solving smaller jobs and tasks for people that
• 1:25
  are really just annoying and I guess work as a layer on top of
• 1:29
  that user interface that's already been built.
• 1:33
  So let's discuss some examples that will help you understand
• 1:36
  what I mean here. Kindle launched A recap feature
• 1:39
  which perfectly exemplifies what we're talking about.
• 1:43
  It aimed at solving the challenge and the problem of
• 1:46
  narrative continuity between books.
• 1:48
  So that moment maybe when you pick up a book in a series many,
• 1:52
  many months after reading the last one.
• 1:53
  And so then you kind of forget a lot about what's happened and
• 1:57
  maybe therefore just don't even pick up that book and you stop
• 1:59
  reading it all together. Rather than forcing people to,
• 2:02
  say, go online, read a summary or try and recap what they
• 2:06
  actually read previously, or maybe even abandoned the series
• 2:09
  altogether, Kindle now generates automatically this concise
• 2:12
  little summary of the previous book as soon as you open a new
• 2:15
  one. What makes this quite clever is
• 2:16
  that it's contextually relevant. It only really appears when it's
• 2:20
  actually needed and doesn't interrupt the core reading
• 2:23
  experience of a person, doesn't ask readers to change how they
• 2:26
  use the product. It just removes a simple pain
• 2:29
  point that's annoying and does so that adds very little new
• 2:33
  user interface or buttons to press or screens to visit.
• 2:36
  Another great example mentioned by Rich Holmes was Google Chrome
• 2:40
  who recently launched an on device AI system using their
• 2:43
  Gemini model. And essentially what it's doing
• 2:45
  is trying to catch security threats in real time.
• 2:47
  So when it detects something suspicious, maybe like a support
• 2:50
  scam or phishing attempt, it just pops up and tells you and
• 2:54
  it doesn't interrupt your browsing experience.
• 2:57
  It's not trying to create a separate feature area or product
• 3:00
  area that makes you actively have to engage with it.
• 3:03
  It's just a widget that is directly integrated into that
• 3:06
  normal browsing experience. Another great example was
• 3:09
  referenced by a fantastic product thinker by a Cash Gupta
• 3:13
  and he talks about a company called Ateo.
• 3:15
  So Ateo, for those that don't know it is ACRM.
• 3:18
  So helping people manage sales deals.
• 3:20
  And what they've done really cleverly is implement AI in a
• 3:23
  way that just super helpful and they're growing really fast in
• 3:27
  what is an extremely crowded market and they're up against
• 3:30
  the likes of Salesforce and Deals and winning.
• 3:32
  An example of this enhanced UX embedded UXAI feature is that
• 3:36
  when a user creates maybe a new automation on their platform,
• 3:41
  the system automatically summarises the automation and
• 3:45
  gives it a description and a name and a summary based on what
• 3:47
  it's supposed to be doing. The user isn't having to do
• 3:49
  anything, it's just a small thing at the top left of the
• 3:51
  screen. There's no flashing AI badge, no
• 3:54
  special interface you have to learn.
• 3:56
  It's just a small amount of magic and delight when users
• 3:59
  realize that what was previously a very tedious step is now just
• 4:02
  done for them. Another example is by Atio
• 4:05
  again. So when someone responds to a
• 4:08
  e-mail coming through their system and says I'm busy all
• 4:11
  week to the salesperson at your system, automatically adjust the
• 4:15
  scheduling. So immediately sending out
• 4:18
  another time and another option for them.
• 4:20
  Again, super simple, but really helpful.
• 4:22
  And so to bring it back to my point, this first category of
• 4:24
  enhanced UX and embedded UX often feels invisible in the
• 4:28
  background and pops up with a widget on your screen to give
• 4:31
  you that help you need. And as Akash got to put it,
• 4:35
  these companies instead of saying to themselves how do we
• 4:37
  add AI to everything, it's actually them asking how come we
• 4:41
  add AI that makes this experience just feel a bit
• 4:44
  magical. The former leads to random
• 4:47
  bolted on capabilities that feel misjointed and us as users feel
• 4:52
  like why have they done this, this is pointless.
• 4:55
  Whereas the latter question drives a complete re imagination
• 4:58
  of a products experience for people.
• 5:00
  And the common thread here is that all of these examples are
• 5:04
  an integration with existing experiences.
• 5:06
  They don't ask a user to fundamentally change their
• 5:08
  behaviour. And these implementations often
• 5:10
  represent one of the lowest technical barriers to entry for
• 5:13
  new AI features for a company because they're simply improving
• 5:17
  what exists already. But meanwhile, it's not a low
• 5:19
  impact change. These are really nice touches
• 5:21
  that just feel magical to a person.
• 5:23
  And the beauty of this is that in essence, these are just an AI
• 5:27
  agent, a workflow, a set of tools and ability to trigger
• 5:30
  based on a thing that happens in their product.
• 5:33
  Instead of that output from an agent and a workflow being text
• 5:37
  in a chat component or in a chat widget like ChatGPT or another
• 5:41
  chat interface, it just appears in their existing user
• 5:44
  interface. So if I take that down a level
• 5:46
  and make this really tangible by explaining how an Ato does
• 5:49
  something like that, first an agent is created which loves to
• 5:53
  summarize things. That's its purpose.
• 5:55
  It also will respond. Its text responses will be
• 5:59
  structured in very concise ways with clear bullet points for
• 6:03
  example, because it loves to have very concise summaries.
• 6:05
  It then has a set of skills too, which are basically workflows.
• 6:09
  These workflows will trigger based on a user in their
• 6:11
  platform doing a specific thing. In this case, it's them creating
• 6:15
  an automation in RTO's platform. The agent then recognizes that a
• 6:18
  user's just done this automation and immediately triggers It
• 6:22
  fires off maybe a signal to the agent and says time to do your
• 6:25
  summarization. And then the agent will decide
• 6:27
  to use the workflow it's got access to.
• 6:29
  So the summarization workflow and step one of that workflow
• 6:32
  will be to trigger based on a user doing something.
• 6:34
  As I said, Step 2 might be to summarize the automation.
• 6:38
  Step 3 was to reorganize it and rewrite it.
• 6:41
  Step 4 was to publish it. And again, that summary doesn't
• 6:44
  appear in a conversation or a chat window.
• 6:46
  It appears in a user interface of Attio or a little widget pops
• 6:49
  up and says here's the summary. I understand this so intimately
• 6:52
  because we're actually launching something that does this type of
• 6:55
  thing next month. So workflow builder plus new
• 6:58
  widgets that can appear and be embedded on SAS Company's
• 7:01
  platforms anyway, so let's move on to another category of
• 7:03
  feature. This is agents that control
• 7:05
  existing functionality on a platform.
• 7:08
  So this second category really represents a very fundamental
• 7:11
  shift in how users will interact with software that's already
• 7:15
  been built. So agents that don't just, you
• 7:17
  know, answer a user's question, but actively controls and
• 7:21
  operates an interface based on what a user needs to complete.
• 7:25
  So rather than a user navigating very complex user interface,
• 7:28
  it's quite frustrating. These agents allow people to
• 7:33
  speak to a, well, what is a chat bot essentially in natural
• 7:36
  language and the agent just does it for them by controlling the
• 7:39
  features on that application. So a great example here is what
• 7:43
  Stripe just launched, which is their dashboard assistant.
• 7:45
  Again, this was discussed by Rich Holmes in the Department of
• 7:47
  Products sub stack article. So this dashboard assistant or
• 7:50
  agent enables someone like a merchant using Stripe, so a
• 7:54
  vendor that sells things to simply type refund transaction
• 7:57
  ABC 12345 and then the agent will just navigate to the
• 8:02
  appropriate section of the dashboard, locate the
• 8:03
  transaction, kick off the refund process.
• 8:06
  Maybe in the conversation we'll come back to the human and say,
• 8:09
  do you want me to process this? Press confirm and it's done.
• 8:12
  So suddenly creating products, updating pricing and managing
• 8:15
  admin tasks using Stripes existing interface is really
• 8:19
  simple. And a key take away here is that
• 8:22
  I guess for decades companies have built products around user
• 8:26
  interface with buttons and forms and drop downs and things like
• 8:30
  that. And as a user, we have to learn
• 8:33
  where something's located on their app, how to complete
• 8:36
  certain tasks using the forms they provide.
• 8:39
  And what often starts as an extremely innovative solution
• 8:44
  often becomes incredibly frustrating for us as users
• 8:47
  because we get so used to it. We now expect something better.
• 8:50
  And that's why agents controlling existing features
• 8:52
  that a company's already built is so, so magical.
• 8:55
  Users don't always have to go and find out how to create and
• 8:58
  send a financial report which is located somewhere in Stripes
• 9:02
  application. They just need to have an agent
• 9:04
  on the side ask it a question and say where is this thing gets
• 9:09
  taken to that area and says now complete and send that financial
• 9:12
  report to X people. Because the features are already
• 9:16
  built to do that. The agent can control those
• 9:18
  features and execute that task. So for SAS companies that are
• 9:21
  really big and have been around for a while, this approach
• 9:24
  offers a really compelling strategy to compete with these
• 9:28
  agent for SAS applications and SAS startups that are launching
• 9:31
  pretty much every week. And rather than having to
• 9:33
  rebuild every inch of their experience from scratch, they
• 9:36
  can layer on these agent interfaces on top of their
• 9:40
  existing functionality, which leverages their core IP and
• 9:44
  technological advantage that they've built up over many
• 9:47
  years, whilst delivering a conversational experience and
• 9:51
  interface that users now want. And This is why I'm so excited
• 9:55
  to launch a closed invite only beta programme of this
• 9:58
  functionality in late June, early July.
• 10:00
  It's something that I think is going to be the future of many
• 10:02
  SAS applications. And I think this is going to be
• 10:05
  one of the core important strategies our SAS customers are
• 10:08
  going to have to take over this next 12 months as they start to
• 10:10
  launch brand new products and AI experiences throughout the
• 10:14
  course of the next few years. And I'll come on to what I mean
• 10:16
  by this in just a few minutes. Now if I was in the product team
• 10:19
  at one of these companies, I'd be asking myself and the team
• 10:22
  which part of our product requires the most training,
• 10:25
  which parts of our product gets the most questions asked about
• 10:27
  it. And how can we take away that
• 10:30
  complexity with natural language agents that control that
• 10:33
  experience. So it represents an extremely
• 10:36
  high quality opportunity for agents that control features.
• 10:39
  Anyway, moving on to the third type of AI feature that we're
• 10:43
  seeing getting launched. This is task or use case or role
• 10:47
  based agents, which are able to automate specific tasks in
• 10:50
  separate systems. What distinguishes this category
• 10:53
  from others is that it's very laser focused on specific jobs
• 10:56
  to be done or job function, job role or a task.
• 11:00
  For example, the sales outreach agent.
• 11:02
  So I'll dig into a couple of examples here.
• 11:04
  A sales software called Salesloft recently launched a
• 11:07
  bunch of agents and essentially this tool, as it's called
• 11:10
  product offering, help sales reps make calls to people to
• 11:13
  sell things. And now they launched these
• 11:15
  agents to help these salespeople do jobs that currently sales off
• 11:20
  don't really support. For example, the agents they've
• 11:23
  launched can read into Salesforce or HubSpot and do
• 11:27
  things in there. It can also push things into
• 11:30
  communication channels. It can also write new emails for
• 11:32
  them. So an agent that does many of
• 11:34
  the specific tasks that they would have to do at work.
• 11:37
  Linkedin's approach is also another good example.
• 11:39
  They've launched multiple specialized agents for different
• 11:42
  user segments of their product. Their latest is a dedicated
• 11:45
  interview preparation agent on their platform that helps
• 11:48
  basically job seekers practice for interviews.
• 11:51
  And what's particularly interesting about these use case
• 11:54
  or domain specific agents is that they enable these
• 11:58
  technology companies to extend their value proposition and
• 12:02
  their offering beyond their core product functionality.
• 12:07
  So for example, Microsoft launched their deep research
• 12:10
  agent that allows people to research all of their internal
• 12:13
  SharePoint information, for example, but also allows them to
• 12:17
  search externally through web sources and the Internet.
• 12:21
  This would have been the domain of Google, not Microsoft.
• 12:24
  So what's really interesting here is companies are suddenly
• 12:26
  able to expand into other market segments, jobs to be done that
• 12:31
  they previously would have had to build all of this user
• 12:33
  interface around and would have taken years to deliver.
• 12:36
  But now they can just build a single agent, launch it, sell
• 12:39
  it, and deliver value quite fast.
• 12:40
  Now good product teams at these companies should really start by
• 12:43
  identifying different user segments with specific recurring
• 12:47
  needs. So ask themselves what are the
• 12:49
  highest value jobs to be done that someone would love to give
• 12:53
  away? And importantly, what do we
• 12:54
  think AI could do? Well so certain things AI will
• 12:58
  always be or might be for a long time just rubbish up.
• 13:02
  So AI might be able to produce lots of content, but it often
• 13:05
  lacks context. So really think through what
• 13:07
  they can be really good at. And this really would help
• 13:11
  prioritize worthy investments in AI agents for specific tasks or
• 13:16
  job functions. So we've had over say 200 AI
• 13:19
  coaches and tutors built on our product and launched into other
• 13:23
  technology company solutions. So Type 4 I'm going to spend
• 13:25
  very little time on because it kind of blends into other
• 13:27
  categories. But Type 4 really is Gen.
• 13:30
  AI first features. This was covered again by Rich
• 13:33
  Holmes in the Department of Product article.
• 13:35
  And a really good example here is how Google or Google Hangouts
• 13:39
  should I say, has just launched their Generate with AI feature.
• 13:42
  So you can now change your background but generate it with
• 13:45
  AI language models. So you can ask it to create a
• 13:48
  pirate background or a football field background.
• 13:51
  Another great example is Spotify, which launched the AIDJ
• 13:53
  feature and now you can speak to it and ask it to produce a
• 13:57
  playlist for certain moods or certain types of situations.
• 14:01
  Again, you could argue this blends into other categories and
• 14:03
  hence I'm not going to spend too much time in this episode on
• 14:06
  that. I'm now going to move on to the
• 14:07
  final category we're seeing getting launched in the market
• 14:10
  today, and that's AI native stand alone applications.
• 14:13
  Now, these aren't upgrades to existing products.
• 14:15
  They are entirely new experiences built from the
• 14:19
  ground up with AI capabilities. And what distinguishes these
• 14:22
  products from other types of features is that they're a
• 14:26
  fundamental RE architecture and rethinking about how an
• 14:29
  application can work. These are often reimagined
• 14:32
  entire product categories based on what becomes possible from
• 14:35
  AI. And a great example of this is
• 14:37
  Google's Notebook LM, which is discussed by Richard Holmes.
• 14:41
  And it isn't just an enhancement by Google to, let's say, the
• 14:45
  Google drives features or a note taking up.
• 14:48
  It's a fundamentally new way of looking and experiencing
• 14:51
  information. You can quite literally upload
• 14:53
  books, videos, documents and turn it into a podcast.
• 14:57
  You can extract insights, identify patterns, whatever you
• 15:00
  need to do, and really enables you to have a conversation with
• 15:03
  that piece of information. You also see this pattern in
• 15:06
  marketing tools like Jasper and Copy AI, which are marketing
• 15:09
  tools built around AI experiences.
• 15:11
  And they're not just adding AI to a Google doc or or a
• 15:15
  Microsoft Word document. They're creating entirely new
• 15:18
  experiences and solutions to create content optimized for AI
• 15:22
  workflows. These applications aren't
• 15:25
  constrained by legacy approaches or user expectations.
• 15:29
  They're completely free to reimagine how something should
• 15:33
  work now. For big and established
• 15:34
  technology companies with 10s of thousands of customers, these
• 15:38
  stand alone applications represent both the probably
• 15:42
  greatest opportunity long term, but also the greatest risk.
• 15:44
  And that's because an AI company or an AI startup that launches
• 15:48
  one of these types of new products that are reimagined, it
• 15:51
  could make your entire category or even set of products
• 15:53
  obsolete. But meanwhile, it can also
• 15:56
  present a huge opportunity for you as an established player to
• 16:00
  launch a new product into the market and completely dominate.
• 16:03
  But they also open up new possibilities for big companies
• 16:06
  to launch offerings that previously weren't very
• 16:09
  feasible. Fundamentally, this is a long
• 16:11
  term play that some companies will do so many companies won't
• 16:14
  be able to do because of resource constraint, because of
• 16:17
  how much risk is involved and eventually will probably acquire
• 16:20
  other start-ups. Anyway, that's it for today.
• 16:23
  I've gone through those top five categories.
• 16:25
  I hope that was interesting. Thanks for listening and I'm
• 16:28
  going to see you next week.