Skip to main content
Work

InSite 2.0: Rebuilding an AI Chat Agent Against the Clock

InSite 2.0: Rebuilding an AI Chat Agent Against the Clock
Act-On had been selling an AI chat agent called InSite. Customers put it on their websites, visitors asked it questions, and some of those visitors became leads. The product came from a vendor. Act-On white-labeled it, put its own name on the box, and resold it. Then the vendor told our VP of Product it would shut down at the end of the year. He brought the news to the team with a hard date: January 3rd. By then, Act-On needed its own chat agent in production, with existing customers moved over and still working. Subtract Thanksgiving and the winter holidays, and the runway got short in a hurry.
I started by auditing the vendor's tool. Every screen, every setting, every behavior we'd been reselling. Then the PM and I went into Gainsight to see which customers were using it, and how much. InSite had also been running on Act-On's own marketing site, which gave us real behavior to look at. 70% of visitors who opened the chat asked a question. 20% of engaged visitors submitted a form. Two behaviors, both real, both worth designing around. Then we talked to the seven customers who used it most. They cared about two things: feeding the agent their own content, and capturing visitor info through embedded forms. They also had a wish list. They wanted control over the visual styling, more knowledge sources, multiple agents for different purposes, and if-this-then-that logic so they could steer the agent when a visitor mentioned a specific word or phrase. That wish list became the roadmap. January 3rd decided the order. The directive for V2.0 was clear. Reach feature parity, ship by January 3rd, and plan fast follows to improve the experience. That made the cut line easy to draw and easy to defend. V2.0 had one agent per account, one knowledge source, light visual styling, embedded forms, and response rules. The rest of the customer wish list went into V2.1 and V2.2. Nobody had to fight about squeezing in one more thing, which is rare and lovely. InSite serves two people who will never talk to each other. The admin is a marketer. They're putting an AI agent on their company's live website, where it can say something wrong in front of a prospect. Every decision on the admin side is about making that risk feel manageable: knowing what's live, testing before anything ships, and recovering cleanly when they change their mind. The visitor lands on the customer's website with a question. They don't know Act-On exists, and they shouldn't have to. They want a quick, accurate answer. Maybe they fill out a form. The admin experience is what makes the visitor experience safe to ship.
Looply AI Chat panel with a blue header, a greeting asking what the visitor would like to know, and three suggested questions about mid-market services, pricing, and scheduling a demo
The chat after a visitor asks what services Looply offers: the agent's answer about implementation, support, and consulting, followed by three blue links to related pages
The agent replying to a request to contact the team with an embedded form: required first name, last name, and email fields, then company name, phone number, and an interest dropdown
The visitor side, prototyped in Figma Make. Greeting, suggestion chips, answers with links back to the source, and a form right in the conversation.
An agent answer describing a six to eight week implementation process, followed by three suggested replies about discovery, training, and go-liveA submitted contact form above a confirmation that the Looply team will reply within one business day, with chips to ask another question, browse resources, or view documentation
InSite visitor chat on the sample Looply websiteA clickable prototype with sample content. Pick a suggested question, then type a couple of follow-ups to reach the form.Open the prototype in a new tab
Admins build their agent in a composer with four tabs: Settings, Design, Greeting, and Responses. Design and Greeting show a live preview of the widget next to the controls, so a color change shows up the moment you make it. Responses is where the if-this-then-that request landed. Admins describe the voice and tone they want, then add response rules. Each rule has triggers written the way a visitor would say them ("A visitor asks about pricing") and a response, with an optional asset like a contact form or a case studies page. The agent matches the conversation to the closest rule. Marketers get control over the agent without learning prompt engineering.
The Responses tab of the InSite agent composer: a voice and tone field, then three response rules for pricing, case studies, and integrations, each listing its triggers, response, and linked asset
Response rules: triggers in the visitor's words, a response, and an optional asset. The hardest part of the composer was telling admins what's live. An agent has two separate states that are easy to mash into one. Visibility: can site visitors see it? Publication: are there edits sitting in draft that visitors haven't seen yet? Those combine four ways. Visible. Hidden. Visible with unpublished changes. Hidden with unpublished changes. Collapse that into one on/off switch, and nobody can tell what their website is actually showing. So we named every state and kept the names consistent everywhere: the header, the modals, the buttons, the help text. Visible and Hidden. Published version and Unpublished changes. Show and Hide. The header always tells you which of the four you're in.
Four versions of the composer header in a grid, each with a different status beside the agent name: Visible, Hidden, Visible (unpublished changes), and Hidden (unpublished changes)
The composer header in all four states. The little status next to the agent name is the most important text on the page. There was a debate about the main workflow. One model combined publishing changes and showing the agent into a single action, since that's what most admins want when they finish editing. The other kept them separate: queue up your changes, then decide when visitors see them. We landed on separation and kept the combined action as a shortcut for the common path. Try to show an agent that has unpublished changes, and the modal offers "Publish and show" right next to "Show agent."
A Show agent dialog explaining the agent will go live with its last published configuration, a yellow notice about unpublished changes, and three actions: Cancel, Publish and show, and Show agent
Two steps by default. One step when you need it. Marketers need to see their agent before visitors do. In the interactive preview, admins can chat with their agent before anything goes live. A toggle at the bottom switches between the published version and the unpublished changes, so you can compare what visitors see today with what you're about to ship. I wanted preview to run on the customer's actual website. That turned out to be technically hard, so we scoped it down to a generic browser window. Admins still get the feeling of seeing the agent on a real page, and engineering got to keep its December.
The interactive preview screen: a Hidden (unpublished changes) status, the Looply chat showing a contact form inside a blank browser window, a tooltip saying form submissions are disabled in preview, and a Published version and Unpublished changes toggle at the bottom
The interactive preview, with the published and unpublished toggle at the bottom. I built a working version of the visitor experience in Figma Make. It ended up doing four jobs. It got stakeholders on board, it went to engineering as a reference, it showed customers what was coming, and it let us test whether the conversation flow made sense before anyone wrote production code. V2.1 and V2.2 brought the rest of the wish list: multiple knowledge sources, content behind a login, authentication through HTTP headers, and multiple agents. All of it came out of discovery. Some of it arrived after the designs were already done. I stopped treating it as one moving target. I broke it into separate pieces and gave each one a focused pass against the architecture we already had. Claude helped here as a requirements partner. It kept the growing list straight and flagged gaps from a design perspective before they turned into surprises in development. The HTTP header authentication modal nearly broke my brain. It asks a marketer to configure something most marketers have never heard of. I distilled it as far as it would go. Indexing shaped the knowledge source design too. Scraping a website takes time, fails in creative ways, and has limits. So the knowledge source screens needed an updating status, error states, and scrape limits that made it clear what was happening. Content exclusion takes wildcards, so a marketer can drop every blog post from the agent's brain with one line.
The Knowledge sources page in Act-On: an active status banner, two source URLs, a content exclusion setting, and a searchable table of 420 collected pages with their type and last collected date
Knowledge sources: where the content comes from, what's excluded, and every page the agent has read. Customers wanted different agents for different jobs, like a sales agent for prospects and a support agent in a customer portal. That raised a new question. How do you decide which agent shows up on which page? The first approach asked admins to paste a variable into the page itself, which is a lot to ask of a marketer who just wants a sales agent on the pricing page. We replaced it with URL rules that live inside InSite. One agent is the default and appears everywhere. Secondary agents get assigned to specific URLs, with wildcard matching for whole sections of a site. An exclude list covers the pages that shouldn't have an agent at all. Three rules, no overlap, and a marketer can explain it to their boss in one breath. Agents and knowledge sources also got private and public settings, which separate content behind a login from content anyone can see.
The Agents page listing a default agent, a sales agent, and a customer portal agent with their knowledge sources and last updated times, a Create agent button, and a note about using different agents for different audiences
The agents list. One default, as many specialists as your plan allows. Some important InSite moments happen when the admin isn't even logged in. I designed the emails for those, including the one that goes out when an account hits its annual conversation limit. It explains that the agents are paused until the limit resets, confirms that conversation history is still in reporting, and points to the account manager for an upgrade. Forms got the same treatment. If someone edits an Act-On form in a way that breaks an InSite agent using it, the save is blocked, and the modal lists the unsupported fields and every linked agent. There was also a sales move baked into the product. Accounts without an InSite license can build agents, customize them, and test them in preview. Shipping one requires a license. Look at all the ways you can customize it. Try it out. Want to ship it? Let's talk. The Report tab answers two questions. How are visitors moving through the chat? And what are they asking about? A Sankey chart shows the journey, from opening the agent, to starting a conversation, to continuing it. Below that, the agent tags conversations with themes automatically, so admins can see what's bubbling up: billing questions, demo requests, product info. A conversation table at the bottom gets you into individual chats, with links clicked and forms submitted. It shipped as designed, which still feels like a small miracle.
The Report tab for an InSite agent: a Sankey chart flowing from opening the agent to starting and continuing a conversation, a themes table with bars for topics like product info and billing, and a conversation list with contacts, themes, links clicked, and forms submitted
Engagement as a journey, themes as a pulse check, and every conversation one click away. The visitor-facing widget can't look like Act-On. It has to look like the customer: their color, their font, their chat button, their corner of the screen. So I designed a set of InSite-specific components that are brand-agnostic on purpose and live outside Act-On's design system. Brand colors created an accessibility problem. Customers pick whatever their brand guide says, and the chat header still has to be readable. I worked out a formula with Claude that takes the chosen primary color and picks black or white text based on contrast, then handed it to our front-end developer. Customers get their brand. Visitors get text they can read. Nobody has to know what a contrast ratio is.
The Design tab of the composer: a live preview of the Looply chat widget and chat button on the left, and controls for primary color, widget image, font family, and widget position on the right
The Design tab. Pick a color, a font, a chat button, and a corner. The widget handles the rest. V2.0 shipped by January 3rd. We migrated existing customers to the new InSite and met with them one-on-one to get them set up. Some of those calls were the PM, customer success, and me together. Plenty were just me. Act-On isn't a big company, so everyone wears a lot of hats, and talking directly with customers is a comfortable part of the job for me. One piece of feedback stuck with me. During a discovery interview, a customer who was leaving Act-On told us the new InSite was a major improvement. They were leaving because a new parent company used a different platform. They had no reason to be nice about it. V2.0
  • Act-On's own AI chat agent, one per account
  • A knowledge source built from the customer's website, with content exclusion
  • Light visual styling: color, font, chat button image, and position
  • Embedded Act-On forms for lead capture
  • Response rules with triggers, responses, and optional assets
  • Draft and publish workflow with the four visibility states
  • Interactive preview with a published and unpublished toggle
  • Reporting with engagement, themes, and conversation history
V2.1 and V2.2
  • Multiple knowledge sources, including content behind a login and HTTP header authentication
  • Multiple agents with a default agent, URL rules, wildcards, and an exclude list
  • Private and public agents and knowledge sources
  • Conversation limit emails and form compatibility guardrails
  • Build and preview without a license, deploy with one
Brand-agnostic is its own design problem. Designing components that have to look like someone else's brand meant designing for colors and fonts I'd never see. The contrast formula did more for accessibility than any style guide could, because it worked on every color a customer might pick. Name every state. Hidden and Unpublished look like a wording choice right up until a marketer accidentally ships something. Naming decisions deserve the same rigor as layouts. Break the moving target into pieces. When scope grows mid-project, a focused pass on each new piece, checked against the existing architecture, keeps the whole thing coherent. Test the scary flows early. We made the publish and visibility calls on internal judgment. A few sessions with customers would have given everyone the same evidence to work from.

Related projects

AI Subject Line Optimizer
Act-On Software

AI Subject Line Optimizer

Designing Act-On's first customer-facing generative AI feature: a subject line analyzer that uses each account's own performance history to generate personalized alternatives.
Groups: Putting Data in Context
Sparta Science

Groups: Putting Data in Context

Designing a way for Sparta Science users to compare individual performance against meaningful peer groups rather than the entire dataset.
COVID-19 Guidance
TripIt

COVID-19 Guidance

Designing a centralized, digestible source of COVID-19 travel restrictions inside TripIt so travelers could stop hunting across government sites.
Safety Personalization
TripIt

Safety Personalization

Designing a personal risk threshold system inside TripIt so travelers could set their own safety preferences and get alerted when destinations exceeded them.
Settings Refresh
TripIt

Settings Refresh

Volunteering to lead a full rebuild of TripIt's outdated web settings, bringing it up to the accessibility and usability standards of the mobile app.