Lessons Learned Building Enterprise Bots
Almost all of the early Slackbots have moved in the same direction — they have added strong web app components. I assume they’ve all done it for the same reason we did. When we started Talla, we tried to push everything through the bot interface in natural language. But the problem with B2B bots is, the same task/process/workflow at various companies will be quite different. Let me give you an example.
Lesson 1: Workflows vary across companies, which is hard to deal with in conversation.
Say you want to build a bot that manages the offer letter generation process for candidates at a 500 person company. What does the bot need to do? Well, the process could vary quite a bit. It could be as simple as this — you are the hiring manager, you generate the offer letter and extend the offer. Or maybe you generate the offer letter and get HR to approve it and then extend it. Or maybe you generate the letter and HR approves it then they extend it. Or maybe HR generates it and gets your approval and the approval of your boss and then they extend it. You get the picture.
If you want to build this process into a bot, you need to have a way to configure it for the variations in the process across companies. But now, try to do that configuration through a bot. Using natural language to tell the bot about the process is a nightmare. We got around it by building a web interface for configuration.
Lesson 2: Reporting and Analytics are important for enterprises, and hard to deal with in conversation.
Most enterprises love data. If they are going to deploy your bot, they want to see data about what it does, who uses it, and understand the impact it has on the company. Again, this is something difficult to do through a bot interface, and is better through a web interface.
You see the trend, right? More and more bot related tasks are going through a web interface. I assume the other bot companies that have moved this direction have seen this as well, which is why they’ve done it.
Lesson 3: You have to build most of your own stack.
The number of “business” bots that are built on other startup platforms continues to surprise me. If you want to build anything for an enterprise that is mission critical (and those are the most valuable apps) you have to offer a SLA. And you can’t offer an enterprise quality SLA if some other part of your tech stack doesn’t have one that is equivalent to, or better than, yours. Maybe some companies have the idea that this can wait as the bot ecosystem matures, but I don’t. So far, the early enterprise adopters we are talking to are asking about this. If you are raising money for, or funding an enterprise bot company, this means it will be more expensive than you think.
Lesson 4: Onboarding matters.
We built our own internal onboarding A/B testing tool. Why? Because conversational interfaces have a discoverability problem. Knowing what a bot can do is difficult, because there is no clear menu showing it all. We believe the way you get around this is by determining early intent and showing the user the piece of the bot they need to see. So we built our own A/B onboarding tool to drop new users in different experiences and measure the results.
Here is an interesting one. The way we measure success at the moment is whether or not a Talla user creates and executes some kind of workflow. We’ve used the onboarding tool to drop half of users into the web app, where we walk them through creating a workflow. We’ve dropped the other half into Slack, where we deliver them a workflow, so they can get the user experience, then take them to the web app to create one. Guess which one performs better?
When dropped into the web app, 68% of users create a workflow. When dropped into Slack, 32% do. How you onboard users to a bot will influence your success rate.
Lesson 5: Spreading through an organization is a challenge.
In Slack, sometimes one person in a 1000 employee company installs and uses your bot. How do you reach the other 999? We try to push to get to an administrator, or someone in I.T. who can let the entire company know about Talla. But it doesn’t always happen. And often, the installer is someone who doesn’t have the authority, or doesn’t feel comfortable, sending something to the entire company.
We built some infrastructure to do batches of direct messaging and record the results of various tests. A whole post about those tests is coming, but, for now, we learned that the engagement results are better than email. It’s a tricky space to navigate because you really want to ask permission to go spread through the organization, and if you do DM someone, you want to give them the ability to mute quickly so you don’t bother them again.
We had hoped to use a bottoms up approach to building Talla, but given the current state of the ecosystem, we’ve moved away from it. We think a sales driven process that focuses on understanding and driving full value to the organization is more useful and successful at this point in the market development.
Lesson 6: Events are great for leads.
I’ll end with a marketing insight. Startups often hate to go to events because they are expensive, and typically eat up a couple of days. But, in a space that is poorly defined like “bots,” events give you the chance to talk to a few hundred people in a short time span and iterate on your pitch. Figuring out the early messaging and positioning matters.
This space is impacted by so many interesting trends: the adoption of chat based messaging, A.I. technology, NLP and conversational interfaces — that is why it is incredibly exciting and fun to work in this space. But it is also poorly defined and confusing. We are learning a lot, and wanted to share with the rest of you who may also be trying to figure some of this out.
And by the way, Talla is hiring for almost all positions so, if you like working on these problems, reach out — email@example.com. Of course, if you need a bot that works like an HR service desk, with an intelligent workflow automation engine and a conversational knowledge base, we hope you will try Talla.