How to Train a Bot
All AI bots require some human review to function well and respond accurately. My team and I spoke with partners and internal staff to see how we might provide insight and control over a bot’s answering behavior.
Research and Design | Mainstay | Research: January – March 2021 | Design: March – October 2021
Deliverables
Research Plan & Analysis
Personas
Prototypes
Roadmap Recommendations
Skills and Experience
Exploratory User Research
Usability Research and Analysis
Product Roadmap Planning
Visual Design
Mainstay is a student engagement platform.
Their Chatbots use AI, Natural Language Processing, and scripted blurbs to connect students with universities for greater graduation success.
The bots are a mixture of proactive…
… and reactive messaging exchanges.
Terminology
Bot Manager: The university staff members who write content, monitor conversations, schedule outreach messages, maintain bot knowledge, and/or pick up conversations when the bot can’t sufficiently answer.
Contact: The students who receive these outreach messages, respond to interactive messages, and/or ask the bot questions, all through SMS or Webchat.
Bot: A Mainstay chatbot, which communicates via text using scheduled messages and pre-written blurbs of knowledge.
Knowledge Base: A content library of pre-written responses the bot can use
Knowledge Staff: Mainstay employees who spend time reviewing bot conversations and maintaining knowledge across many bots and universities.
Business Problem
We needed to scale.
The company needed to increase the number of accounts it could support while maintaining or improving bot accuracy in order to:
Expand the client base
Increase the number of positively impacted students
Maintain and improve bots with less human effort
Like most AI bots on the market, Mainstay bots will always require some human intervention to function well and respond accurately over time. Mainstay’s knowledge associates provided significant support for messaging review and knowledge management.
Over time, this became a non-scaling solution. Knowledge associates lacked the time to review every conversation, but our partners lacked a way to do it themselves.
Users
Partners
Many of our partners spent time regularly reviewing messages in their conversations manager, however not all partners are like this. We needed a way to surface problem areas in a way that would allow them to quickly address missed opportunities without feeling like the bot was failing at its job.
While some partners have a full team of employees to work on strategy and message review, many of our partners have a limited number of employees dedicated to managing a bot.
Knowledge Staff
Our Knowledge Staff browses messages that went unanswered by our bots and suggests a better match, which must be approved by the client.
While we have an amazing group of team members reviewing these messages, it was impossible to review the influx of messages coming through as we entered into the early months of the pandemic.
Discovery Research
We needed a way for partners to successfully manage their bots, and a way to give back time to our Knowledge Staff.
Learning Goals
What tasks and responsibilities take a significant chunk of time?
How do users know what they need to do next for their bot?
What tasks can be entirely automated?
Stakeholders
Mainstay Knowledge Staff
AI Team
Executive Team
Existing Platform Users
Research Methods
Open ended interviews
Observational sessions
Desirability studies
We were able to uncover several tasks that took a significant chunk of time, and some limitations in the platform’s current abilities.
We already knew that our partners needed a method for reviewing their incoming messages. Many of our best partners used our unbranded internal review system, which lacked workflow clarity. Through conversations with partners, we were also able to uncover a number of concerns they had to make it a more comfortable experience.
Overall, partners needed tools that would let them do more with less time, without giving up control of content and bot behavior.
Partners interviews typically consisted of open ended questions and usability testing for any prototypes in progress. Regularly opening with open ended questions and continuing discovery research allowed us to paint a clear picture of our partners over the course of the year.
Analysis
We narrowed down our findings to several important themes, based on the top challenges our partners face today.
Bot Metrics
We did not have a clear, reliable way to measure a bot's accuracy. The Automation Rate is frequently incorrect, and has become a pain point for Partner Success and the Support teams. AdmitHub frequently frames incorrect questions as bad or a flaw of our system, instead of highlighting them as an opportunity for growth or expansion.
Suggested Projects
Provide a metric or set of metrics that will indicate a bot’s behavior
Give actionable feedback for how one might improve those metrics
Content Management for Teams
Working in the Knowledge Base is cumbersome. Partners have mentioned various usability issues in the content management system that individually are unimportant, but collectively make the system unpleasant and time consuming.
Partners want efficiency, and they need to be able to see where they need to take action.
Suggested Projects
Create permissions for new users so Bot Managers can delegate work without giving up control of a powerful mass-messaging tool
Allow greater flexibility and control in content categories to work with partners’ existing organizational structures
Advanced filtering for easy and accurate search workflows
Nudges
Partners are overwhelmed by the amount of content that needs attention. They are unsure what to do next, and uncertain where their efforts are most valuable.
I recommended better content flagging, so that it would be obvious which messages and scripts were due for review. Knowledge Staff were also able to see important high level trends, and needed a way to send that information directly to Bot Managers.
Suggested Projects
Improve how we make Content Recommendations
Invest in the Content Recommendation Library
Improve message review for efficiency
A notification manager that can help Bot Managers set their own reminder
In-app messages and reminders that help focus partners on their next task
Knowledge Base Trend Automation
Human eyes are expensive and faulty. We do not use the data we already have (all those conversations) to improve Knowledge Bases and bots scientifically, and there's no way to go through thousands of conversation logs without some automation.
Suggested Projects
Collect feedback from students and share it back with Bot Managers for directional suggestions
Increase flexibility and customization in the platform’s content organization
Bulk editing tools for content
Conversation Automation
We will eventually need to reduce the number of hours Mainstay staff spends reviewing messages. Partners who lack the time and resources to review all student messages will need some support to direct students using AI as much as possible.
We can improve the content hierarchy so that a message which matches to many answers can indicate the bot should ask a follow up question for more precision.
Suggested Projects
Create pathways to information available when the incoming message is not clear
Develop conversation tools that allow the bot to ask for more information and synthesize multiple messages together
Collaborative Editing Tools
Partners with multiple bots or those that work in a university system have to manually review school-wide understandings to make sure they match. This is high effort, low quality, and leaves lots of room for errors.
Rigid content organization makes it difficult to share information with others, and a lack of permissions and team tooling makes it difficult to get input from content experts without opening the sensitive messaging capabilities to unauthorized people.
Suggested Projects
Reduce friction for collaborative content editing
Allow for restricted roles to increase collaboration and increase delegation of content editing
Allow connected universities to share knowledge easily and efficiently
Advanced Knowledge Maintenance
Once these methods have been implemented, there will still be a subset of partners who prefer to pay for our services and skip the review themselves. We must find a way to help Knowledge Staff prioritize certain messages and manage a bot efficiently.
Suggested Projects
Conversation prioritization
Message grouping