Shiny features and hard problems

Today, technology vendors are rushing to introduce AI features in all possible corners of their existing products. There doesn’t seem to be any hardware or software that you couldn’t spray a lil’ AI magic dust on top of and instantly make it shine much brighter on the stage at industry events.

Who can blame these professionals for just doing their job, though? You really can build pretty amazing experiences with AI today and catch the world’s attention. Those who are incorporating publicly available GenAI services from platform vendors into their existing products through an API can create their first iterations of new AI features blazingly fast. If you’re not tapping into Azure OpenAI services and the likes, are you even competing with the vendors in your field?

The biggest buzz is around the providers of those AI models that make this development possible for all the rest. When is the next GPT model becoming available? Will Claude beat Gemini? What will Apple do to replace Siri with proper AI capabilities that meet user expectations? These are the kind of questions that unite tech geeks across ecosystems and industries.

The reported deal between Apple and OpenAI made me think about the path of technology evolution that has lead to this moment. If the public launch of ChatGPT has been described as the “iPhone moment” for AI, then the release events arranged by OpenAI are today’s equivalent of what Apple product announcements used to be sometime in the not too distant past.

These two players give us a way to reflect on a universal phenomenon that is also very much present in the Microsoft business applications space (the underlying theme here at Perspectives on Power Platform). The role of shiny new features that grab the public’s attention - and the hard problems that persist from one year to another.

From the iPhone to HAL9000

Since the only Apple devices I’ve every purchased for myself have been iPads (being a PC + Android guy), I haven’t bothered to actively follow the Cupertino events for many years now. Based on what I’ve seen reported on social media from the latest May 7th event for the iPad product line-up refresh, they keep delivering exactly what I would expect. The biggest focus is on A) how thin the device is and B) how much the theoretical performance has increased. There is no “one more thing” anymore.

Despite of the product development done on A + B, the ability for iPad to grow up to be an alternative (let alone replacement) for my PC isn’t getting any closer. Features like Stage Manager are just adding up to the cognitive load of a user, based on my experiences, rather than making me want to pursue a new way of working with just an iPad. The “post-PC era” claimed by Steve Jobs back in the original iPad launch still hasn’t materialized in the broad sense it was talked about back in 2010. Instead, the list of iPadOS limitations just grows as people expect more from their tablets.

It’s not just hype, of course. Mobile devices have introduced an undeniable paradigm shift into how we think about applications and usage scenarios for computing. Yet they didn’t replace everything that already existed. We got tons of new tools and abilities, while the big industry players gathered incredible piles of money from the new market. And still, the traditional computers like laptop PCs remain relevant for many of the computing tasks that were already in place before smartphones arrived. New tech didn’t solve all the old problems. The rally behind iPad was largely about conflating the shine factor from the iPhone with the harder problems that a PC replacement needed to address somehow.

Compared to the matured mobile device market where Apple dominates, OpenAI is now living somewhere in between the original iPhone and iPad releases in terms of its shiny new technology. It’s only been 17 months since ChatGPT became available. It means the world is still just getting into grips with what is (and isn’t) possible with an advanced LLM. These new three-letter acronyms represent the idea of a revolution: not just a more efficient way to do old things (evolution), but a completely new way to involve technology into our lives.

In 2007, being able to browse a website or pinch-to-zoom a map on an internet connected touch screen that’s available in your pocket at all times was a revolution. The previous evolutionary era consisted of smartphones that had more polished features for traditional phone use cases, yet failed to evolve into usable pocket computers. Today, ChatGPT is to Google search what iPhone was to Nokia Symbian smartphones.

This week when OpenAI launched the latest GPT-4o model, the “omni” capabilities were demonstrated in several very impressive videos of how you can now talk with the AI. As in, not just use voice recognition to turn your spoken words into a text prompt. This is no longer about talking at it - rather the AI now invites users into engaging in a human-like dialogue with it. When I got to test it on my own phone and ChatGPT Plus subscription, it really startled me how my dramatically perception of the tool shifted as a result of this.

I already knew it was basically a similar GPT-4 level service in terms of knowledge and text processing capabilities. From a raw information work perspective, there was no major leap to be found here. The big leap that happened on the human-computer interaction level, though, was profound. It changed the mental model from a chatbot with great text processing skills to a virtual someone. Instead of operating a smartphone type of computer, it felt that I was talking with an entity like HAL 9000 (from 2001: A Space Odyssey).

The new interaction pattern made me use the tool differently and also affected my own behaviour. I now was naturally greeting the AI at first, like a person who had joined a phone call. I was giving information to it that wasn’t a carefully planned prompt with a huge payload to carry the necessary context. Instead I talked out loud and just let my mind work out the intent as we both went along. Finally, when coming to the point where I wanted to move on to other things, I couldn’t just hang up on the AI without saying “thanks, bye”.

This is definitely an example of a shiny new capability that I wasn’t explicitly asking for - yet it caught my attention and made me interested in exploring its implications on an article like this. I don’t yet know exactly what problem the new omni-aware, virtually present AI entity on my phone will solve for me. In that sense, it is not yet a tool. It’s a technology representing a possibility to be used as a new kind of a tool, once the users get to discover what it feels like to them. Thanks to the new initiative of OpenAI to democratize the prior premium models to all ChatGPT users, anyone interested can soon participate in the collective experimentation phase.

“I’m sorry, Dave, I’m afraid I can’t track that email”

In many ways, “with AI” is today’s variation of “on mobile” trend we saw at around 2010. As the world became aware of the fact that smartphone users can install new apps on their devices (a capability not invented by Apple but rather done right by them), every software vendor had to respond to this. There needed to be a mobile app that did something for your system, or you weren’t perceived to be in the competition anymore. Just like AI is treated today.

Computing paradigm shifts like cloud computing or mobile phones or generative AI make it possible for vendors and users to do things that weren’t feasible before. At the same time, these new technological waves have a tendency of forcing vendors to reinvent the wheels already used in their earlier carriages. Often the result may not be a wheel that is objectively any better at its core function. We see this all the time in software products like Dynamics 365 that now have a legacy spanning over two decades already.

Let’s look at one example. The evolution of the contact management part about CRM has gone through many technological stages. If we start from the paper based days and work our way towards 2024, we can identify at least these six steps:

• On a Rolodex

• On a personal computer

• On a shared server

• In the cloud

• On mobile

• With AI

Each step of course makes it possible to advance to the next step. We couldn’t jump directly from a rotating card file device on the office desk into the Copilots of today. Migrating from one system to another and adopting new tools along the way is often the price you must be willing to pay for moving forward in your business practices and unlocking new opportunities.

Now, let’s put aside our technology and platform strategy envisioning hat for a moment and assume the role of a normal business user. As an entrepreneur who now runs his own business, I’ve recently had the pleasure to again be primarily the end user of a CRM system (in addition to being the sysadmin, of course). What I have quickly observed is that contact management is still not easy - despite of the massive technological progress.

From the Rolodex invented in 1956 to the Microsoft Copilot for Sales launched in 2022 / Copilotized in March 2023 / [renamed more times than I can keep track of], the core user need for establishing contact records for the business to manage has remained intact. It’s pretty simple: something happens in the real world, and since I expect there to be relevant interactions ahead with the parties involved in the event, I’d like to capture this data in a structured format for managing future events.

Microsoft CRM was originally born from the big vision of bringing CRM features to where the users already were: in Microsoft Outlook. The product team went as far as implementing an actual lightweight CRM server and database inside the Outlook client to offer offline capabilities, which was a the source of endless grief for anyone having to support such a deployment.

Aside from viewing your entire CRM database via the frame of Outlook, which was always an awkward idea, the one unquestionably valuable capability that Dynamics CRM did offer was the “Track in CRM” button. Clicking it from an email in your inbox or a meeting invite on your calendar would promote the item into the shared CRM database as an activity, alongside its activity parties. Sure, the new contacts created as a result of this click needed some manual data editing to be complete, but it is still one of the clearest & most obvious productivity features of CRM software I’ve come across.

Given the origin story of the product, it’s a bit discouraging to find that in the year 2024, getting a customer email reliably tracked from your Outlook inbox into your Dynamics 365 CRM is still hard. I’ve now witnessed this several times when using New Outlook on my Windows 11 PC and attempting to use the latest Copilot for Sales feature to track emails and contacts into my CRM database. Instead of the classic “Track in CRM” and “Set Regarding” buttons, I now get to watch the Copilot for Sales add-in trying to load itself:

And then fail to load. Ah, reminds me of the “fat client” days of Dynamics CRM for Outlook where the buttons kept disappearing or the add-on getting disabled. Only this time it’s supposed to be all in the cloud, not on your local PC.

The troubleshooting tips from the “why am I seeing this” link above provide a deep-dive into SSO error messages from Microsoft 365 Office add-ins. Right at the start the article talks about getting Fiddler, error handling in getAccessToken API and other similar dev stuff. Not helpful for a low-code developer like me, let alone any user that will encounter such errors when trying to use Copilot.

Does it work on Outlook for the web? Thankfully, it usually does. How about on the mobile then? Ahem. Sorry to burst the bubble, but tracking emails and contacts from a mobile phone has largely remained merely a theoretical possibility for as long as I’ve worked in the Microsoft ecosystem. Once again, on the 2024 release wave 1 there is a coming feature called “support sellers with Copilot for Sales mobile experience in Outlook”. I’ve seen far too many features like this come and go to believe that it would suddenly “just work”.

Client for Outlook.
Email router.
Forward mailbox.
Server-side sync.
App for Outlook.
OWA & MOWA.
Folder-level tracking.
Copilot for Sales.

So many partially great technical solutions for one core requirement in CRM. No final solution in sight. Not even from the software giant that provides a major share of the servers, clients, infrastructure and user experiences for business emails globally. Think about that for a moment. Even when you own 100% of the stack, getting emails and contacts linked to one another 100% of the time is impossible.

Note that I did not write this article only thanks to one irritating bug/feature in Copilot for Sales. Four years ago I was in a similar situation when setting up another blank Dynamics 365 CRM environment, with basic functionality to manage the contacts of a newly established company. “Track in CRM” soon started to fail for all appointments, in a practically uncustomized environment. Months later I discovered that due to a dependency from a Universal Resource Scheduling solution that MS had automatically deployed to our CRM, there was a plugin error that didn’t get a fix deployed because MS failed to deploy any updates to URS after the environment provisioning. The root cause at that time wasn’t AI then, but rather the Dynamics 365 product bloat that was messing up 1st party solutions.

With all the perspective I’ve gained from working with CRM in different stages of the product’s lifecycle, I’ve come up with the following prophecy:

It may not be better, but at least it’s different

Adding AI features on top of your existing business applications is a prime example of a shiny feature. Being able to reliably connect emails sent to/from customers into your CRM system is a persistent hard problem. When Microsoft pushes ad banners inside your Outlook inbox to promote a new product to do a thing that their previous product already did, which category do you suspect it to fall into?

Just like with any regular tech company, who can blame Microsoft representatives for pushing people to use the latest AI powered tools? The business incentives within all these organizations are by default aligned to reward shiny new features. They are what dominates airspace in official channels, because it’s a lot easier to talk about new technical possibilities than addressing existing technical problems (especially within your own product suite).

For those of us who do not work for Microsoft, we can make a choice. How much of our attention do we focus on the endless list of new & shiny features announced, which may further build up the anxiety of struggling to keep up with Power Platform? Are we sure we’re not using them as our own escape mechanism for not having to deal with the old & hard problems that exist in the day-to-day reality of business users working with these tools?

As tech professionals, it’s only natural to get excited about the shiny stuff. We do need to be aware of what’s happening in our ecosystems, but perhaps we should stress a lil’ less about taking all of it into use. Especially the V1 products that receive biggest fanfare out there in the internet are more likely to be a distraction rather than a perfect fit to solve an existing problem your business has identified.

Think of it this way: do you remember the last time you regretted not adopting a new Power Platform / Dynamics 365 feature sooner? I sure don’t. The reverse is probably easier for many of us to relate with. Meaning, putting your hopes on something shiny and new that never matured enough to be able to solve more problems than it generated. Such is life on the technological bleeding edge.

Keeping your eyes open while not getting blinded by the shine

Some tech always is the next iPhone, though. Disruption happens gradually, then suddenly. How can we know when it’s actually worthwhile to invest our time and attention into the shiny stuff announced by tech vendors? I think it comes down to identifying the difference of what I mentioned a bit earlier in this article:

Directing your energy towards something that you can explain to yourself as being revolutionary rather than evolutionary is one principle that can help. Using myself as an example, having observed the MS BizApps market for a long time, I saw Microsoft’s launch of Dynamics 365 (to replace earlier CRM & ERP offering) as a logical evolutionary step. Safe and sensible. Whereas the birth of Power Platform represented the potential for a citizen developer led revolution through the rise of low-code as a new mainstream concept. Hell yeah that sounded a lot more exciting! Which is why I decided to jump all-in there in 2020, before many MS partners even realized it’s an actual market that exists.

So, why didn’t the shine of iPhone directly convert into the death of the PC? Because getting a mobile-first device like an iPad to properly cover all the common scenarios that a PC/Mac does is a hard problem. Based on many recent comments in the tech media, it feels like the consensus is now that Microsoft’s Surface ended up being the more practical solution for tablet based serious computing after all (at least if you’re not immersed in the Apple ecosystem for all your services/data/devices). Some out there still are waiting for the magical solution for iPad to overthrow the PC or Mac - 14 years after the initial launch. It starts to remind me of a certain email related functionality in CRM…

Finally, back to the AI topic. We’re at a point in time where it’s kind of obvious that something revolutionary will come out of the GenAI boom started by ChatGPT. It’s simply too shiny to ignore. Whether this something will be specifically the kind of chat based experiences we see now added into existing software through implementations like Microsoft’s Copilot - that’s still up in the air, the way I see it. What we do know for sure is that LLMs can do magical things when given the right data, acting as a calculator for words. A brand new thing we didn’t even know we needed just a moment ago.

Assistants that help the individual worker get things done are the most likely avenue where the AI revolution will gradually become the new normal. As the tools evolve to be more HAL 9000 like and present themselves as an on-demand virtual colleague, people can discover they can now solve problems they didn’t even dare to approach before. Just like there was a time before we used Google search, new mainstream activities for information processing are bound to emerge. The non-deterministic nature of LLM based AI tools is something that will cause many fingers to get burnt, though. We don’t know for sure how they respond each time and if what they say is accurate. The machine actively lying to you is now a feature of generally available, paid software in the year 2024. From OpenAI, Microsoft and everyone else.

Organizational business applications are a different frontier. Whether LLM based solutions can ever become reliable enough to autonomously run existing business processes, that remains to be seen. Protecting your business from prompt injection is a hard problem. It’s unlikely to get solved in some future version update of your favorite GenAI API provider. A more likely near term solution will be restructuring the processes and tools in a way that allows leveraging AI capabilities inside them, while acknowledging that humans can’t let go of the wheel just yet.