Level 5 Software
Enterprise software has operated on a basic principle for the past decade: ROI correlates directly with user effort. Organizations accepted that getting value from their tools meant dedicated time and energy from their teams to use them correctly. Salesforce is the easiest example - it works great IF every single sales person diligently enters every detail into the database.?
The next decade of software will be defined by AI, but this will only live up to the hype if we fundamentally rethink how software should work. Instead of asking users to adapt their behavior, AI-powered tools should deliver value with zero user effort. The core principle should be: no one should have to do anything for your software to add value.
Level 1 Software
Let's take an example of something we are thinking about here at Windmill. Every executive and manager we talk to wants their team to spend less time in meetings. This was music to our ears and a perfect problem that we could help solve! We have all the data they are looking for.
And so we built a system that kept track of time spent in meetings. We could slice and dice by team and by department, and we used state-of-the-art AI to automatically classify meetings into different categories so you could get advanced analytics. It was awesome.
We showed this to a handful of design partners and the reaction was universal: "This is awesome - and yup, as I expected, we are in a ton of meetings..." followed by an inevitable "now what?" Our fancy graphs and analytics just confirmed what they already knew. Instead of solving the problem, we'd simply given managers more data to analyze and more decisions to make.
In hindsight, we built a level 1 system. Self-driving cars have 6 levels. Level 1 is driver assistance like automated lane keeping. Level 5 is full self-driving when the steering wheel gets removed. Previously, the best the manager could do was one-off manual inspections in Google Calendar - Level 0. Now with our tool, they had all that data aggregated in one place and could automatically keep track of this over time.
It wasn’t that our tool was useless - it just didn’t move the needle. The problem is not "I don't know how many meetings people are having," it's "we have too many meetings." Since that learning, we've rethought everything about our product. Specifically, people should not have to do anything in order for it to work.
领英推荐
Level 5 Software
Taking this fresh lens to meetings, we have a new level 5 approach we are currently building. Our system automatically detects meetings that are potentially a poor use of time (e.g., low attendance rate, poor agenda, constantly moved or cancelled) and then automatically reaches out to some attendees to ask them for feedback about the meeting. If enough signals are met and the feedback is consistent, the system automatically pings the meeting organizer telling them to adjust or cancel the meeting.?
The analytics aren’t there to solve the problem, they are there to show that it’s working! This is a system that solves the problem of removing ineffective meetings and is always running 24/7. It doesn’t take vacation, scales infinitely and no one has to do anything, and that's absolutely critical.
The Future of AI Software
I've seen countless examples of businesses that roll out ChatGPT or Glean or some other tool to their workforce expecting huge productivity gains. In practice, 5-10% of users use the tool since it takes work to get value. You have to know how to use it and when to use it and what types of questions to ask. In order to really move the needle, you need a system that does the work automatically.
This trend is starting to gain steam this year. ChatGPT recently rolled out tasks which you can use to set recurring reminders (essentially run this prompt at this time every day). LangChain recently wrote an article talking about Ambient Agents, which is a very similar concept.
In order for AI to have the impact people hope for, level 5 systems need to be designed. We won’t be building any more level 1 systems here at Windmill.