All case studies

Software · Client work · 2026

An AI chat that collects ideas.The old form stayed as backup.

A software company collected ideas for using AI from its employees, through a long form. I built an AI assistant that asks the same questions in a chat, one or two at a time, and suggests an answer when someone is stuck. The form stayed as the backup: if the chat failed, the assistant sent the person the form. Every idea got the same scoring and a written report, whichever way it came in.

  • A backup built in

    the form, if the chat failed

    When the chat broke, the assistant sent the person the old form. Both fed the same list, and every idea was scored the same way.

  • Suggested answers

    instead of empty fields

    When someone said 'I don't know', the assistant offered an answer they could keep or change. On the form, that question stayed blank.

  • A report for every idea

    sent back by email

    Every idea came back to the person who sent it with a score and a reason for each part. The full report added a small test to run and a market check.

Client
Software company
Stack
  • n8n
  • Google Gemini

Context

The company ran a programme to get its people using AI at work. One part of it was simple: anyone could send in an idea for where AI could help.

Ideas came in through a form, and the form was long. It asked what the idea was, who it was for and what problem it solved. It asked what the idea was worth to the business, how long it would take to build and what could go wrong. Every answer landed in one list.

Problem

The hardest questions on the form were the ones about value and effort. What is this worth? How long would it take to build? Someone with a good idea often cannot answer those alone.

What a form cannot do

  • Help with a question the person cannot answer. 'I don't know' becomes an empty field.
  • Ask a follow-up question when an answer is unclear. A short answer goes in as it is.
  • Tell the person what happens next. The idea goes into the list and waits for somebody to read it.

So the job had two halves. Help people while they answer. Then give every idea an answer back.

What I built

An AI assistant in the company chat that interviewed the person about their idea, and a chain of scoring steps that read every saved idea and wrote a report back.

I built it to one rule: the chat changes how the questions are asked, not which questions. The assistant asked what the form asked and saved the answers in the same shape. An idea from the chat and an idea from the form looked the same afterwards.

  1. One or two questions at a time. Not the whole page at once. The assistant replied in the language the person wrote in, and it asked again when an answer was unclear.

  2. 'I don't know' got a suggested answer. The assistant offered an answer the person could keep or change. This is the one thing a chat does that a form cannot.

  3. Nothing was saved before yes. At the end the assistant showed a summary of the idea. The person corrected it or confirmed it. Only then was it saved, with their name and email added automatically.

  4. Every idea got the same scoring. A few fixed questions, like how well the idea fits the company and how hard it is to build. Each score came with a one-line reason and a recommendation.

  5. A test plan and a market check. The report said what to check first and how: a small test, one or two weeks long. Then it looked at whether something similar already existed. The report went back to the person by email.

The form as the backup

The obvious version of this replaces the form. The chat is newer and friendlier, so the form goes. I kept the form from the start, as the backup.

A chat with a model sometimes goes quiet. The reply comes back empty. When that happened, the assistant said so. It asked the person to repeat the last message, or it gave them the link to the form.

The form was not a dead end. Once a day, the form answers were copied into the same list as the chat ones. From there, every idea went through the same scoring and got the same kind of report, whichever way it came in.

That took the pressure off the chat. It had to be better than the form when it worked. When it did not work, it had to hand the person over to the form.

Result

Ideas arrived in one list, in one shape, whichever way they came in. Everyone who sent an idea got a summary and a written report back by email.

The score was advice, not a decision. The people running the programme read the reports and decided which ideas went ahead. The full report ended the same way: this is a first look, so talk to the team before you start building.

The hard part

The hard part was one word: yes.

Everything before it is an ordinary conversation. If the assistant gets an answer wrong, the summary shows it and the person fixes it. At yes, the person is finished. They close the chat and go back to work. If the idea was not saved at that moment, nobody would find out.

Yes is also the moment where a model can fail without anyone noticing. The model can reply "OK, I will save it" and save nothing. Or the reply can come back empty, right after the person said yes. People also confirm in many ways, and a short "ok" has to count as a yes.

So the assistant said “saved” only after the save had really worked. If the reply came back empty right after a yes, the person saw a clear message: the idea is not saved yet, type Save to try again, or use the form. And if saving failed completely, the person got an email with everything they had said, so they could send it in again without starting over. Nobody had to explain their idea twice.

What this actually says

Putting an AI chat in front of an old process is easy now. The usual next step is to switch the old process off, because the chat is newer and friendlier.

That is the step where things get lost. A chat that works almost every time makes a great demo. The person it fails loses the idea they just explained, and nobody finds out.

Here the chat did the part a form cannot: it helped people with the hard questions. The form did the part a chat cannot promise: it was always there.

An AI chat does not have to be perfect. It has to know where to send people when it is not.

See also.