Every SaaS dashboard in 2024 had a chat with your data button glued to the top right corner. Most of them are gone now. Quietly removed in a redesign, hidden behind a settings flag, or still technically there but with usage graphs that look like a ski slope into the sea.
I have spent the last year watching three small SaaS products up close. Not naming them, because the lessons are more interesting than the logos. In each one, the team shipped two or three AI features in a hurry, then spent the next twelve months learning which ones their users actually liked. The pattern that came out of it surprised me, and once I saw it I started seeing it everywhere.
Here is the tour.
Product A, a project management tool
This one is a Trello-shaped thing for small teams. Boards, cards, threads of comments under each card. In early 2024 they shipped two AI features in the same release.
The first was a big purple button at the top of every board called chat with your projects. You could ask it things like what is blocked this week or which cards has Sarah moved recently. It opened a side panel. The panel had a text input, a send button, and a friendly greeting.
The second was much smaller. On any comment thread longer than five replies, a single line appeared at the top: one-line summary of this thread, followed by the summary itself. No button to click, no panel to open. Just a sentence at the top of a long conversation.
Twelve months later, the chat-with-your-projects feature was dead. Usage had collapsed to under one percent of weekly active users, and most of those were the founders dogfooding. The one-line thread summary, on the other hand, was being read by something like seventy percent of users who opened a long thread. It had quietly become one of the most-viewed pieces of text in the product.
The team killed the chat panel in a spring cleanup. Nobody complained. A few users wrote in to ask if the thread summaries could get even shorter.
Product B, a CRM for small sales teams
Different shape, same story. This product is a lightweight CRM. Contacts, deals, a pipeline view, a notes field on every contact.
They also shipped two AI features. The big one was an AI insights panel on the deal page. It listed things like this deal has been idle for fourteen days, consider following up and the last three deals at this stage closed within a week of the most recent call. It looked smart. It demoed great. Investors loved it.
The small one was inside the next-action field on every contact. As the user typed, the field would quietly suggest the rest of the sentence. Send the proposal by Friday. Schedule a follow-up call after the budget review. The suggestion was greyed out behind the cursor. Tab to accept, ignore to reject.
Twelve months later, the AI insights panel was being clicked by about four percent of users on any given week. The team had moved it lower on the page twice trying to get traction. The inline next-action suggestions were being accepted dozens of times per user per day. Sales reps did not even notice they were using an AI feature. They just noticed that typing the next action got faster.
One sales manager I talked to could not name a single AI feature in the product. When I pointed at the next-action field and said that is AI, she blinked and said oh, I thought it was just learning from my past entries. Which is, of course, exactly what it was doing. She had been using it constantly for six months.
Product C, an analytics tool
Last one. A small analytics product, the kind a marketing team buys to see what their website is doing.
The big swing was a natural-language query box. Ask anything about your data. You could type how many signups came from the newsletter last month, and it would try to build the chart. Sometimes it worked. Often it returned the wrong chart and you spent five minutes trying to figure out why. The product team kept tuning the prompts. Users kept giving up and going back to the chart builder.
The small one was an auto-generated weekly summary email. Every Monday morning, every account got an email with three short paragraphs. Traffic was up nine percent week-over-week, mostly from your most recent blog post. Newsletter signups dipped on Thursday after the email went out late. Your top page this week was the pricing page, viewed 4,200 times.
Twelve months later, the natural-language query box had a single-digit weekly usage rate. The weekly summary email had an eighty percent open rate. Customer success calls would routinely start with I read your Monday email and noticed that. The marketing team at one customer told me the email had replaced their internal Monday-morning analytics review.
The team has since doubled down on the email. They added a version for daily, a version for monthly, and a version that goes to Slack. The natural-language query box is still in the product, in a tab nobody clicks on.
The pattern
Three different products, three different problem domains, three pairs of AI features. In every single case, the big bold chat-shaped feature flopped, and the quiet inline assistant won. Once is a coincidence. Three times is a pattern.
Here is how I would phrase it after watching all three. AI features that interrupt the user with a chat UI fail. Features that quietly improve something the user was already doing succeed.
The chat-with-your-projects button asked the user to stop, change context, formulate a question, type it, and read an answer. The thread summary appeared at the top of a thread the user was about to read anyway. One added a step. The other removed one.
The AI insights panel asked the user to stop, scroll, read, decide if any of it was useful, and act. The inline next-action suggestion arrived inside a field the user was already typing into. One added a step. The other removed one.
The natural-language query box asked the user to stop using the chart builder they already knew, learn a new mental model, type a question, and hope the result was right. The weekly summary email arrived in an inbox the user was opening anyway. One added a step. The other removed one.
The rule of thumb
If I had to put it on a sticker, I would say this. AI should remove a step, not add one.
The features that win are the ones that look at what the user is already doing and quietly do part of it for them. The features that lose are the ones that show up with their own UI, their own mental model, and a polite request for the user's attention.
The best AI feature is the one users do not realise is AI. The sales manager who thought her next-action field was just learning from her past entries was using a feature that, by any technical definition, was AI. She did not care. She just noticed that her job got easier.
Friction kills more features than capability does. The chat panels in those three products were, on average, more capable than the inline assistants. They could do more, answer broader questions, handle more shapes of input. It did not matter. They cost the user a step, and the user did not have a step to spare.
If you are building an AI feature this year, my one piece of advice is to look at the workflow you are about to interrupt. If your feature lives inside that workflow, it has a chance. If it lives next to it, behind a button, in a panel, it probably does not. 🌱