It’s been almost two years since ChatGPT stormed into our lives.

In a post I wrote back then (here), I shared my thoughts on this, claiming that ChatGPT was the most disruptive technology I’ve seen since the introduction of the iPhone. The time that has passed since then just convinced me more that my observation from that time was correct.

The introduction of ChatGPT and the Gen-AI technologies that followed created a clear inflection point in the market, resulting in a wave of startups aiming to revolutionize the world using GenAI. But it wasn’t limited to startups. Almost any decent company, which offers digital services to its customers, started advertising that it will power its products with AI.

Did they all live up to their promise?

Definitely not.

However, as with any emerging technology, the excitement surrounding GenAI also brought its shortcomings to light. Here are some examples of these shortcomings:

• GenAI is not as affordable as originally thought. High API usage, such as OpenAI’s, quickly becomes costly.
• GPUs became scarce, resulting in an availability problem for data science teams. Just in case you weren’t aware – GPUs (graphics processing units) are the fuel of the GenAI models, as they need extreme processing power. The data science team I worked with at the time tried to secure a machine with a GPU for continuous training sessions with some models, but the machine was repeatedly unavailable. (Incidentally, these shortages led me to invest in NVIDIA stocks – no complaints there, but that’s another story!). It was the first time I witnessed machine reservations fail with a major cloud provider.
• For some of the domains, and some of the use cases – GenAI is still not there. For example, image generation is good, but not great (see my post about this here). Code generation, in some use cases, is produced with built-in bugs. Generating business related texts, such as posts for social media, sales emails, etc.. are ok, but definitely lacks the ‘spark’ to make them stand out, even after applying proper prompt engineering. As of today, you can relatively easily spot presentations, posts and general texts that were generated by AI.
• Despite its impressive abilities, GenAI can still produce wildly incorrect results with unwavering confidence. Here are two examples: ‘how many ‘r’s are in the word ‘strawberry’ (link) and ‘can you draw me a ramen bowl with no chopsticks’ (link). Both which I find hilarious. 

Given all this information, you, as a product manager or entrepreneur, must wonder how you can leverage GenAI in your products or your day-2-day workflows, while maximizing on the pros and avoiding the cons.

In this post I’ll address exactly that and hopefully it will help you make the right product decisions.

How to maximize your value from GenAI

When considering how GenAI can help you in your work, there are two dimensions that you need to consider:

1. How it can help you making your routine more efficient
2. How it can be used to enhance your product

Optimizing your routine using GenAI

I will divide this section to product managers on one side and product leaders & entrepreneurs on the other side, though sometimes the lines are blurred…

For product managers:

Product managers deal a lot with execution and the hands-on aspects of product management. That includes writing PRDs, conducting discovery processes, sprint planning and orchestrating the product delivery process.

Here are the ways I found GenAI to be useful in those aspects:

1. PRD templates – by leveraging the ‘memory’ aspect of ChatGPT (I believe its competition also has such capabilities, and if not, it’s just a matter of a few months) – you can teach it how you like to write PRDs, what’s the structure it should follow, and what is the essence of each section. Then, when you are about to work on a new feature, you can describe the feature to the AI and it will create the skeleton of the spec. Yes, you will definitely need to make adjustments, fix stuff and add your own, but it will save you the tedious boilerplate stuff. There are plenty of tutorials on how to do that. Here is one example (don’t get hung up on the specific template the author chooses, just on the technique, which you can apply with any template).
2. Customer interviews analysis – assuming you are recording your interviews, you can download the transcript of each interview (another output of GenAI, btw) and feed it all to ChatGPT or one of its competitors. However, just shoving it in and asking for insights probably won’t get you what you need (though I do recommend doing it, sometimes you will find gems). You need to consider your motivation for these interviews. What were you looking for? I like to predefine ‘buckets’ of possible results and ask the bot to map the interviews to one of these buckets. For example, if you are hunting for underserved pains, then come up with 6-8 possible pains that you already have on your list as suspects, and ask the bot to map the interview to one of these pains. If you are considering several value propositions – do the exact same thing, but this time the buckets represent several possible value propositions for your feature or product. If buckets are over simplified for you, then you can define a format for a spreadsheet with specific columns that you want the bot to fill (for example – ‘systems used by the customer’) and then each conversation will be translated to a row in your spreadsheet. Essentially, what you’ll be doing here is transferring unstructured data to structured data which is much easier to analyze.
3. Written communication optimization – we already discussed in previous posts (here for example) why communication is 90% of the product manager’s work. So, if you find yourself in a tough spot with written communication (bad Slack correspondence, sensitive emails, etc..) – then you can ask the bot to suggest some non-aggressive or sensitive replies that you can use. And needless to say that you can use it for grammar fixes if you are about to send an important message, or just want someone to proof your presentation.
4. Consolidating a backlog – now, if you follow me for quite some time, then you know my opinion about backlogs (if you don’t – read this). Bottom line, I’m not a fan. However, if your company forces you to store all historical feature requests and want to go over them for finding gems, then ChatGPT and the similars can help. You see, the main problem with feature requests is that different stakeholders can submit the exact same feature request, but using different terms and words. By exporting your Jira backlog (or whatever other system you are using) and feeding it to the bot – you can get some pointers for requests which are quite similar. It won’t be perfect, but it should save you some considerable time. 
5. User interface text strings – you are about to launch a new feature with new UI elements, but not sure about the phrasing of each of the labels and/or descriptions? Well, some companies employ specific people just for that. However, I’m personally not sure this is cost-effective anymore and especially for smaller companies. These days you can upload screenshots, explain the overall feature to the bot and ask it to fill a spreadsheet mapping all relevant controls on the screen to the relevant text for each.  

For product leaders & entrepreneurs

1. Competitive research – ChatGPT and its competitors can be quite useful in pinpointing potential competitors that are having the same feature set or are the leaders in a specific space.
2. Persona definition – given a product, you can leverage GenAI to brainstorm about the right personas for the product you plan to introduce. You can also share the transcript of some conversations around that. Once you narrowed it down to 2-3 personas – you can ask the bot to create a full persona analysis for each.
3. Thought leadership – want to build your presence on the web as someone who knows what he/she is talking about? Ask the bot for some ideas in your domain and ask it to suggest a skeleton for a post once you identified the topic. Take the skeleton and make it a full post and then ask the bot to go over and improve it. This can also be useful for shorter posts that you can post on social media.
4. Marketing campaigns – if you are about to launch a new product you’ll need to tackle the acquisition phase as one of the first steps. Very often this is addressed by targeted ad campaigns testing different messages. GenAI can be super useful on that front since it can generate dozens of messages around the same value proposition within a few seconds.

For me, those are the main use cases. However, there are plenty of more hacks that you use GenAI for. Just Google it if you’re curious.

Enhancing your product using GenAI

In my introduction post to GenAI (here) I suggested a general approach as for how to integrate GenAI into your product. 

Specifically, I gave these two tips:

1. Avoid using a third party GenAI engine if the feature or capability that you are working on is part of your core business. For example, if your company builds crime-detection software using video cameras, it’s risky to rely on a third-party AI engine for such a critical function. Either your team possesses the domain expertise to develop such a capability in-house, or you should solve a different problem. There is more than one reason for that, but the main one is that your whole business is totally dependent on a third party entity, and if that entity decides at some point that you can no longer use its technology – then your whole business crumbles.
2. Don’t apply GenAI solutions for features that don’t move the needle for your business. For example – let’s say you have a travel website and you decide to embed GenAI in it so it will generate a unique greeting for each visitor based on the geo they are visiting from. Hence, a random visitor from Japan will be greeted with a greeting that is generated in real time and it’s in Japanese. You did it because you thought it’s gonna be very cool and probably increase the traffic to your site. Alas, reality shows that this feature has no real impact on any of your business KPIs. At this point in time – the right thing to do would be to kill this feature, of course, because each visitor is just causing you to lose money.

These two principles definitely still apply today, so you should embrace them.

Aside from these two, here is an additional thing that I observed in the last couple of years:

The business risk is high

I quoted in the past that Martin Cagan (a known product guru) has identified four risks that you should consider when working on new products: the value risk, the usability risk, the feasibility (technological) risk and the business risk. I’m not going to discuss all these risks here. If you are interested, here is the original article.

I’d like to focus for a second on the business risk, because this is the most elusive one in my humble opinion, and many entrepreneurs & product managers often miss it or just overlook it. 

This risk covers scenarios in which you came up with a product that does provide value, it’s quite feasible from the technology point of view and you solved all of the usability issues with it. Yet, in these scenarios, when everything is put together, it may not work because the operational costs or other business realities simply don’t make sense when you try to scale the product beyond the first customers.

It could be that the margins are too low, it could be that it requires a tough market education, it could be that the sales cycles are too long, or a bunch of other reasons.

What I have observed is that with GenAI – the business risk becomes a true risk and must be addressed early on.

What do I mean by that?

GenAI is very sexy and can solve tons of problems. It’s relatively easy to come up with ideas of how to create products that deliver value with generative AI. For example – providing an image generation capabilities for the users (like ChatGPT does), providing a service that helps you summarize any article that you want on the web, providing a tool that generates business presentations for you at will or providing a plugin that generates SQL queries for you anytime you need.

Because it’s so easy to generate ideas for where GenAI can be useful, a flood of new startups emerged shortly after ChatGPT was introduced to the world. Many startups which tried to offer such services ended up closing their gates shortly after.

Why is that?

The main reason is that while these ideas provide value and the technological aspect is no longer an issue – they don’t scale well, mainly due to the costs involved. When you give  users a generative AI tool that they can use at will – there is a good chance they will abuse it because… why not? It’s fun, it’s cool and it provides value. I’m gladly paying you my 20$ a month – so what do you want from me?

The fact that many of these companies discovered is that a fixed cost and unlimited usages don’t work well together. Generative AI technology is expensive. Each time a user asks a tool to generate something for them – there is tons of processing that happens behind the scenes. This can be quite expensive, especially if you are delegating the heavy lifting to a third party engine.

The second reason why these startups didn’t strive is because there was no true barrier to entry. If your company doesn’t have any true secret sauce and you delegate the heavy work to a third party – what stops another company (or even just a couple of talented entrepreneurs) to offer the exact same service and steal your customers? The answer is – ‘nothing’.

Hence, the business risk is very real when it comes to GenAI, and because of that, my advice is as follows:

Consider the product or the feature that you intend to offer and understand in which bucket your product/feature falls when it comes to generative AI:

1. Bucket 1 – the GenAI technology is directly exposed to the user, and the user decides the frequency of invoking it.
2. Bucket 2- the GenAI technology is used in your backend, helping you to perform complicated tasks behind the scenes, and the end user just enjoys the output of this heavy lifting. They have no control over the frequency in which it’s going to be invoked.

If your product or feature falls within bucket 1 – then you are walking directly into a true business risk. You have only two options I’m aware of – 

1. Limit the amount of usages the user can do within a given period of time (e.g. 10 usages per day for example)
2. Charge the user a hefty amount of money in advance so statistically, when considering all your users/customers, you should have a nice margin left.

I strongly recommend the first option, because this option can easily scale with your business, keeping you on the safe side all the time (in terms of margins, at least). This is the approach ChatGPT takes, for example. If you try to generate too many images, even if you are a premium user – it will notify you that you’ve reached your daily quota.

Canva and other services take the same approach as well. You can utilize their GenAI services up to a certain amount of times per day.

The second option is a viable option as well, but it does require you to model your business very accurately and is very data driven. Most young companies don’t have enough data to price their products in that accuracy which may cause them to lose money, and changing the price later down the road can be tricky.

If your product or feature falls within the second bucket – then you are in a safer spot, because you have full control on the amount of usages of GenAI technology that you are going to do within a given period of time.

It allows you to accurately project the costs involved.

Still, it doesn’t mean that you are totally safe in terms of your business risk. It still could happen that the minimal amount of invocations of GenAI tech that you require to provide the value to your users or customers simply doesn’t make sense in terms of costs.

It also could be that even though your users can’t invoke these services themselves, your business will require to increase its usage with each new customer that comes on board. In that case you need to make sure that each customer is ‘worth it’ and in some scenarios you may need to draw an accurate profile of customers that you can sell to.

Defensibility

As I noted earlier, the attractiveness of GenAI can make us blind to our competition. The fact that it was easy for us to come up with our first version and start selling is not necessarily good!

If it was easy for us – is it because we have a strong expertise in this domain and we did something brilliant? Or is it because it’s just not rocket science?

If it’s the second – you are in real danger.

It’s great that you are already generating sales, but use that money to create a moat around your business.

Ideally, you should have thought about it in advance and you already have a plan, because true moats are not easy to design.

The best moat would be to have a strong network effect. You can read about it here. However, not all businesses can naturally create a network around them. In that case you need to consider upgrading your technology with in-house brilliancy to offer an added value your competition can’t have or it’s quite hard to build.

If you don’t have any of those – I’m sorry to inform you, but you are living on borrowed time. Enjoy it while you can, make the maximum amount of money that you are able to, and move on to the next thing, because it’s just a matter of time before bigger sharks will eat your business.

Summary

Two years after the GenAI revolution hit our shores is a good opportunity to reflect on what we’ve learned.

Since it was introduced there were tons of posts, Youtube tutorials and discussions teaching us how we can leverage GenAI both for our personal life & work for boosting productivity. 

However, while GenAI is a real game changer and a true enabler to initiatives that were blocked before it was introduced, it also possesses some business risks.

The technology is not cheap and doesn’t always scale well. If your GenAI features are user facing – make sure to design your business model and terms of service in a way that the usages of these features scale well with your business.

Aside from that you need to make sure that you have a true secret sauce and that secret sauce is developed in-house. Otherwise, the easiness of development can quickly turn against you by your competition.

That’s it for today. 

I hope you found this post useful.