Logic Builder · Chatbot builder SaaS · AI features
Adding AI where it really helps
I designed 8 AI features for a tool that lets support teams build chatbots without code. Here are 4 of them. Each one started from a real problem people had, not from what AI can do.
Role
Senior Product Designer, design owner of Logic Builder
Team
Product owners, UX researchers, engineers and an LLM team. The ops team tested every feature.
Timeline
Jun 2025 – Jun 2026 · 12 months on the AI track
Logic Builder grew fast. People had more chatbots, bigger chatbots, sometimes hundreds in one account. Building, understanding, fixing and finding them all took too long.
Use what our LLM team could now build to make these everyday tasks faster. No AI for the sake of AI.
I owned the design of the AI work, from research to release. I ran the workshops where we chose what to build, designed 8 AI features end to end (4 are shown here), and tested each one with the ops team.
People built their first chatbot in half the time. They also understood, fixed and found chatbots faster.
01
Context
The product grew faster than its tools
Logic Builder is a tool for building support chatbots without code. You build a chatbot as a flow: a chain of steps on a canvas, like “greet the customer”, “ask for the order number”, “check it”, “answer”. Each step is called a node.
It started as an internal tool for one company’s own support ops team. I joined when the team grew to turn it into a SaaS product for other companies, like Verizon and Burger King. Their ops teams build chatbots in it every day, without engineers.
Two things happened at the same time. People made more and bigger flows, and some accounts had hundreds of them. And our LLM team could suddenly do much more.
So the question was simple: where would AI really help these people?
02
What we learned
Four problems kept coming back
Before we talked about AI, we mapped what we already knew from the ops team, user walkthroughs and customer feedback. We looked at four scenarios: new and experienced users, building a new flow or fixing an old one.

Four scenarios from our research. All four problems came up in every one.
03
Where AI helps
Every AI idea had to solve a real problem
I ran workshops with design, engineering and the LLM team. We started from the four problems, not from AI ideas. For every idea we asked: does it fix a real problem, can we build it, and how much will it cost to run?

Our impact vs effort board from the workshop. All the fully automatic ideas ended up in low impact, high effort.
We picked the four features that gave the most value for the effort. Those are the 4 in this case. In total I designed 8 AI features. The other four are spelling fixes, short node summaries, theme setup and smart tips.
04
Three rules
How AI should behave in our product
The workshops gave us more than a list of features. We also agreed on three rules that every AI feature had to follow. These rules are also why we said no to the fully automatic ideas.
People stay in control
AI can draft, explain and suggest. A person approves every change (human in the loop). Nothing goes live by itself.
Help matches the problem
A small problem gets a small fix: a typo takes one click. A big problem gets an explanation and a few options.
Cost is part of the design
Every AI request costs money, and big companies make a lot of them. Fewer questions, and AI search you can turn on and off, kept costs low.
05
What we built
Four features, one for each problem
Each feature below fixes one of the four problems. The videos show real product screens.
01
Discover
Research, workshops, rules
02
Quick wins
Explain the flow, Smart search
03
Big bets
Create a flow, Fix errors
04
Test and measure
Weekly tests with the ops team
JUN 2025
ABOUT 12 MONTHS
JUN 2026
1
Create a flow with AI
Starting from an empty canvas was the hardest part. New users didn’t know where to begin. Experienced users built almost the same flows again and again.
So the question was not “should AI help?” but “how much should AI do?”
THE QUESTION
Should AI ask questions step by step, or build the whole flow at once?
OPTION A
Step by step
AI asks you one question for each step and builds the flow slowly. It feels safe and teaches you the tool.
OPTION B
Chosen
All at once
You describe what you want in plain words. AI builds a full first version, and you edit it.
WHY WE CHOSE B
New users learned the tool in a few weeks. After that, answering question after question would feel slow. Every extra question also costs money at enterprise scale. And in testing, the ops team said they wanted to see the whole flow first, then edit it.
THE COST
New users get less guidance, and the result depends more on how well they describe the flow. We solved that with Explain the flow (below).

Step by step needs a question for every node. All at once needs one message.
1
Describe the flow you need in plain words
2
Copilot shows its plan and waits for a yes
3
The full flow appears on the canvas
Describe the flow in plain words. AI builds it on the canvas.
Before AI builds anything, Copilot (our AI assistant) shows its plan and waits for a yes. So you always know what AI is about to do.
2
Explain the flow
The ops team often worked on flows someone else built, or hadn’t opened in months. To understand one, they had to check dozens of nodes one by one.
Now Copilot explains the flow in a few short steps. Each step is a card. Click View, and the matching part of the flow lights up on the canvas. So you can check that AI got it right. You can also ask Copilot questions about any step.
1
Ask Copilot to explain the flow
2
Read a few key steps instead of every node
3
Click View to see each step on the canvas
Copilot explains the flow step by step and shows each step on the canvas.
We first made it to explain flows that AI had built. It turned out to help with any flow. It also gave new users the guidance they lost without step by step.
3
Fix errors with AI
The tool already found errors. The hard part was knowing what to fix and where to start. The error list showed problems but didn’t help to solve them.
1
Copilot lists the errors in the flow
2
It explains each one in plain words
3
Fix it in one click, or ask a follow-up
Copilot finds errors in the flow and fixes them with you, one click at a time.
Typos
Fix or skip in one click.
Settings problems
Copilot explains what’s wrong and offers a fix.
Structure problems
Copilot explains the problem and lets you pick a solution.
So the help matches the size of the problem. And nothing changes until you click.
4
Smart search
With 100+ flows, people remembered what a flow did, but not its name. So normal search by name didn’t help.
Smart search understands what you mean. Type “show me flows with sell / upsell” and you get flows about selling, even if those words aren’t in the name. You can also read a short AI summary of a flow right in the results.
1
Turn on AI search
2
Type what the flow does, not its name
3
Open a short AI summary from the results
Search by what a flow does. Read a short AI summary without leaving the list.
Not every search needs AI. For a simple search by name, normal search is faster and cheaper. So AI search is a switch you turn on, not the default.
06
Outcome
What changed
CREATE A FLOW
time to first flow
BEFORE
An empty canvas. Every node set up by hand.
AFTER
A full first draft, ready to edit.
EXPLAIN THE FLOW
time to read a flow
BEFORE
Checking dozens of nodes one by one.
AFTER
A few step cards, linked to the canvas.
FIX ERRORS
faster error fixing
BEFORE
A list of alerts, no help to fix them.
AFTER
Copilot explains the problem and offers the fix.
SMART SEARCH
time to find a flow
BEFORE
Scrolling and guessing exact names.
AFTER
Search by what a flow does.
Measured by our UX researchers and product owners in tests with the ops team, comparing the same tasks before and after each feature.
07
Reflections
The simplest feature won
The ops team used Smart search more than any other AI feature. And it was one of the easiest to build. The big features got more attention, but people used small, simple help the most.
That changed how I judge AI ideas. Now I ask three questions: does it fix a real problem, can people check what AI did, and is it worth the cost?
What I’d measure next: how much people change an AI draft before they publish it. That shows if the drafts are really good, not just fast.
The hardest part wasn’t designing AI features. It was deciding which ones were worth building.

