Startup idea: Zapier for consumer apps
I've been wondering why there haven't been any really useful 'digital personal assistants' developed over the last few months. There are several recurring use cases for which it could be incredibly useful (ordering Uber/Uber Eats, booking restaurants, booking tennis courts, etc). Wouldn't it be nice to just write, 'Book me a tennis court at @club today at 6 pm for 1 hour' or 'schedule UberX for today at 3 pm, pickup at my place, drop off at the airport'?
Can't one just get access to these consumer apps' APIs and develop this?
Cheers from Portugal
78 comments
[ 5.7 ms ] story [ 153 ms ] threadThe biggest problem is that APIs tend to be very restricted. So most assistants end up limited in what they can achieve.
Cheers from Portugal, too.
It's different for work, because the company pays for zapier, and people get paid to do work aka use zapier.
This however may just be a matter of time until the incentives align and all of this becomes possible. Eventually, it is my conviction that the web will be conversational first - and typing/clicking will be a secondary mode of action for more advanced queries.
Similar for social things. I need to coordinate movie tickets, dinner reservations, or tennis courts not just with the supplier but with my friends.
A generalized text interface for consumers would have to be not only as good, but noticeably better in order to create some monetizable advantage.
If this arises, I’d expect it to come from an existing body of users (Siri/Alexa) encouraging suppliers (Uber, OpenTable, DoorDash, movie theaters) to build towards a common API. Without that pump priming with tens of thousands of local users already on the platform with money to spend, it’ll be hard to gain traction I think.
In regards to the last sentence, 100% - I think this will eventually have to come from one of the big tech
If I could choose I’d much rather book by sending a calendar than a text message
Uber does offer an API, but I think part of the problem is that a lot of these providers either worry about becoming interchangeable, or believe - rightly or wrongly - that their brand is strong enough not to want to let intermediaries cut in.
So I think you're right in as much as that at least in some areas it'll take providers big enough to say "if you don't work with us all of our millions of customers will see your competitors first".
A smaller player will likely have alto start small and be content to not have as broad coverage as they'll like.
Maybe couple with papering over some with human intervention like more traditional concierge services until they can go to providers with proven demand.
Something that was very popular back in the day, and was a great filter for companies, and something that most of the current giants did to their competitors, before declaring via laws that nobody can do it onto them to secure their market position indefinitely.
To their credit, clicking on a restaurant and seeing available ordering platform options with fee info is nice.
So it's more stressful and time consuming than just doing it myself
I vaguely recall some discussion on HN about it, mostly focused on the 'why' question.
There's company in YC23 called Real Magic that I think are having another go.
Maybe try https://harpa.ai . It seems the most advanced / flexible browser extension in a survey I did. Not affiliated in any way.
These requests seem a lot less frequent and more complicated.
Context matters a lot. Which makes it very difficult.
What tennis court? What club? What if it is booked? What about payment? What about agreeing to waivers?
Who would build the integration, the club? The app?
What would be the workflow? 1. Join consumer zapier, 2. join tennis club, 3. oAuth them, 4. then type my command?
Having to write/speak the assembly language of each component of that every time in chat is not going to win in the market.
I don't see these statements as "assembly language", they are questions a human assistant would ask as well.
“There’s a court available 150 miles away; the Uber will be at your door at 3PM to get you to the court by 6” is a legal but non-sensical solution in your version of the interaction.
You have to have some kind of implicit rules and some kind of explicit configuration whether that’s “when I say my door, I mean X” or “when I play tennis, the default duration is an hour and I prefer club Y over town court T1 over town court T2.” And an implicit “people don’t want to spent more time in transit than at an activity, so first constrain search and issue a notice to the user if a solution chosen violates this”.
Looking at how we might approach your tennis court scenario I see a lot of mistakes that the AI could make. Maybe the court is members only, but it didn't realize the court is inside a gym or location where a human would be sure to check that. Or it finds one an hour away, so your Uber is expensive.
I assume if you do this often, you'd know about potential tennis courts in the area. So for me I'd rather do the research up front myself, but this could be personal preference. In the end, maybe we have to see how good these `AI + some code` apps get, but I wouldn't want to use the first versions.
I guess one reason is that you can literally do this yourself within the Uber app in 30 seconds without the need for an intermediary.
Arguably Android or iOS is already an app with buttons for tasks, and you press the button of the task you want like "get a car" or "bring me food" and it happens. The cool part is each button can bring up a UI designed to work best for that particular task.
This is why "everything apps" might not be the next big thing -- they're already here.
An everything app is just centralizing and bottlenecking what already exists: Android or iOS as the everything app with buttons for tasks.
Ok, but can I use my phone to message my tennis court reception to book a session?
As much as I hate Meta there's a ton of value in a universal messaging app. Seems like WeChat is one step further. There's a reason why Musk is attempting to take X in that direction.
1. https://ifttt.com
It's better to start with a workable business model and promise at the beginning rather than scam users with a bait-and-switch.
The real problem is that most people don’t want to pay for things.
Works better and doesn't require monthly payment. Just one time purchase of hardware.
It was unnecessarily complex to setup tho, because producers lock down their hardware to their closed cloud. Then they demand monthly payments because they have "ongoing cost", that shouldn't exist in the first place (IFTTT mentioned api fees of various producers as main reason for more aggressive monetization).
I'm happy to pay for a product. I'm not going to pay company that constructs closed ecosystem just because it allows to extract monthly fee from me.
I feel like someone must be making a easy tool to point llm at your own api but I haven't seen it yet.
https://gorilla.cs.berkeley.edu/
Most apps also want to make it easy for you to do repetitive tasks quickly. Doordash has a "recent orders" section where I can tap to reorder. That's pretty seamless.
This general idea holds across most large consumer apps: if something is done repetitively, the app has a way to make that process faster. As such, a personal assistant only becomes useful for edge cases like "ordering from a new restaurant where you already know what you want and know that it's on Doordash."
Even in your example of "book a tennis court at 6," you might run into issues. What if the court is booked at 6? I might have secondary times in mind, or I might be willing to settle for a 45 minute session if it's available. All of this is easier to deal with using a calendar interface. And what if different times cost different amounts of money? I wouldn't necessarily know what amount that hour at 6 would cost, and that may matter to me.
Poimt being that these apps work well for one thing: tasks that you do repetitively many times, that aren't sped up by the original app itself, and where things won't change. That ends up being a pretty limited usecase, even if you can keep the service cheap.
Our kids use a kid friendly debit card product where we can fund their accounts on demand. I also have a chore schedule system I built where they enter their chores each day. I'd love to be able to automate allowance based on chore completion but there's no way to integrate the two.
Every 2 weeks I enter all my accounts, credit cards, etc into a spreadsheet for graphing and tracking. I'd love if there was an api for these services to fetch the data instead of manually logging in.
And then their chore “paycheck” is every two weeks.
Like simulated work / prep for the real world.
take taxes out of their paycheck. Give one of their friends an allowance from those taxes, but don't make them do any work for it. don't pay them every two weeks due to "a computer glitch". pay the kids different rates and prohibit them from talking about it with each other. threaten to adopt an immigrant kid who will work faster and harder and eat less food, and can live in the closet under the stairs. make is clear that one of the two kids is replaceable with said immigrant child, so they fight with each other instead of unionizing. if one of them gets injured, don't feed them while they're not doing chores. make the other kids work twice as much but don't pay them anything extra, make it clear they should be glad they can do chores and haven't been replaced yet.
I only used it a few times and found it too clunky. But I loved the idea. Worth researching what happened to them.
What needs to happen is an AI-improved Alexa that doesn't talk like a damn robot, doesn't only respond to fixed script text patterns, and ties into all of the home automation, cloud services, and IoT things. Voice agents shouldn't just do things by guessing but should seek a bit of follow-up prompting to clarify ambiguous commands rather than picking the wrong one like Alexa does.
It is good in case, for an app that does everything already(a super-app), they can release a voice assistant which is confined to that app itself.
They are based in menlo park, exactly as you describe it they are doing. Not sure how good it is but they're trying.
If anyone cares about the name reply and I'll get it tomorrow when I go in.
Sorry for late reply
Rather than relying on an LLM every time, our tool uses LLMs to help you automate your task (with code) and then lets you call that “Magic Loop” via text, email, or webhook.
People are using it for surf notifications, triggering lights during touchdowns, notifying about low Tesla batteries, and more.
We have a few “first-class” APIs for scraping, weather, finance, etc. but you can also just add your own API keys/tokens (e.g. for a Slack webhook).
You can try it today: https://magicloops.dev
Building a single production integration connector is difficult and expensive (API licensing/T&C stuff will eat a LOT of your time, APIs across products vary wildly in terms of maturity and standards compliance, dealing with versions/deprecation, and don’t get me started on handling auth).
Building a library of them significantly more difficult and expensive. And this is why the products that do well tend to cost a fair amount of money, which consumers typically aren’t going to be interested in spending.
Open source tools like n8n are promising, but still clunky and unreliable. Too fiddly for consumers.
(Was a PM for an integration product, and have been intimately involved in this space).
In my experience of having tried it, it only worked right about 60% of the time, which wasn't worth the hassle of having to correct it all the time. It was easier just to pull up the app and see the immediate visual feedback every step of the way.
Currently in the early stages of building something like that (https://www.flexibly.ai)