Detailed
TL;DR
Lumen AI asked users what to extract. I gave them something to review first.
Lumen AI asked users what to extract. I gave them something to review first.
Lumen AI asked users what to extract. I gave them something to review first.
Overview
Lumen AI helps private equity teams research companies before investing and monitor their performance
afterward. I redesigned company setup, from document upload to reviewing AI-extracted information.
This story focuses on giving users an initial company profile they could check, edit, and extend without configuring every field upfront.
Before
After
Before
After
Problem statement
Setting up one company meant moving between separate flows.
Setting up one company meant moving between separate flows.
Lumen AI already had document extraction. But users had to upload files elsewhere and configure fields before seeing the results.
Lumen AI already had document extraction. But users had to upload files elsewhere and configure fields before seeing the results.
Lumen AI already had document extraction. Users could select files, choose the fields they needed, and provide prompts to guide the AI.
Lumen AI already had document extraction. Users could select files, choose the fields they needed, and provide prompts to guide the AI.
But the work around extraction was scattered. Documents had to be uploaded in the Data Room before they could be selected during setup.
But the work around extraction was scattered. Documents had to be uploaded in the Data Room before they could be selected during setup.
Document uploads required users to leave the setup flow.
Document uploads required users to leave the setup flow.
User don’t know what will appear against each selected fields.
User don’t know what will appear against each selected fields.
The extraction was already there. I looked next at how the steps around it fit together.
Diagnosis
Users needed to set up a company. The product separated upload, configuration, and review.
Users needed to set up a company. The product separated upload, configuration, and review.
I traced company setup back to the Data Room. Users uploaded documents there, then selected them again during setup.
I traced company setup back to the Data Room. Users uploaded documents there, then selected them again during setup.
I also noticed that field configuration came before results. At first, I treated that sequence as something to keep while connecting the surrounding steps.
I also noticed that field configuration came before results. At first, I treated that sequence as something to keep while connecting the surrounding steps.
Uploading didn’t carry through to setup. Users had already added their documents to the Data Room, but had to select them again to use them during company setup.
Uploading didn’t carry through to setup. Users had already added their documents to the Data Room, but had to select them again to use them during company setup.
The client wanted sources and confidence scores alongside extracted data and a way to switch between confidence scores. Which will triggered by something, not always present.
The client wanted sources and confidence scores alongside extracted data and a way to switch between confidence scores. Which will triggered by something, not always present.
My first move was to connect the steps users already had to take.
What I tried first
I brought the steps together. Configuration still came before review.
I brought the steps together. Configuration still came before review.
The client wanted setup to feel easier and less overwhelming. I assumed bringing the separate steps into one flow would help.
The client wanted setup to feel easier and less overwhelming. I assumed bringing the separate steps into one flow would help.
Attempt I made:
I organized setup into three steps.
I organized setup into three steps.
I organized setup into three steps: upload documents, choose the fields to extract, and review the results.
I organized setup into three steps: upload documents, choose the fields to extract, and review the results.
Users no longer had to select the same documents again or enter basic company details like company name, industry, business model, and website. AI finds this information from the uploaded documents and fills it in automatically.
Users no longer had to select the same documents again or enter basic company details like company name, industry, business model, and website. AI finds this information from the uploaded documents and fills it in automatically.
What I realized
Client feedback changed my focus. Users still had to choose fields without seeing what AI would return for them.
Client feedback changed my focus. Users still had to choose fields without seeing what AI would return for them.
The client tested this version and said they didn’t like this version. They wanted configuration options and extraction results on the same screen.
The client tested this version and said they didn’t like this version. They wanted configuration options and extraction results on the same screen.
That feedback changed my focus. I had noticed the order earlier, but left it intact. Users still had to choose fields without seeing what AI would return for them.
That feedback changed my focus. I had noticed the order earlier, but left it intact. Users still had to choose fields without seeing what AI would return for them.
The scattered flow wasn’t the main problem. The configuration was.
The decision
Give users extracted information before asking what else they need.
Give users extracted information before asking what else they need.
I needed a starting point that let users review information while keeping every customization option available.
I needed a starting point that let users review information while keeping every customization option available.
Connecting the steps hadn’t removed the upfront configuration. I needed a starting point that let users review information while keeping every customization option available.
Connecting the steps hadn’t removed the upfront configuration. I needed a starting point that let users review information while keeping every customization option available.
The reference
As I see the setup as building a company profile, with extraction working inside that process. So I decided explored how other products handled profile creation. A common pattern stood out:
As I see the setup as building a company profile, with extraction working inside that process. So I decided explored how other products handled profile creation. A common pattern stood out:
Information was grouped into sections. Each section showed the fields that belonged there.
Information was grouped into sections. Each section showed the fields that belonged there.
People could fill out what they needed. They could build the profile around their purpose.
People could fill out what they needed. They could build the profile around their purpose.
There was room to add more. The initial structure gave them a starting point they could extend.
There was room to add more. The initial structure gave them a starting point they could extend.
LinkedIn’s profile sections gave me a reference for Lumen AI.
LinkedIn’s profile sections gave me a reference for Lumen AI.
LinkedIn’s profile sections gave me a reference for Lumen AI. I assumed this familiar structure would help users understand what information they could include.
LinkedIn’s profile sections gave me a reference for Lumen AI. I assumed this familiar structure would help users understand what information they could include.
AI could populate the initial fields, while users decided what to correct or add.
AI could populate the initial fields, while users decided what to correct or add.
The change
Before: Upload documents, choose extraction fields, then review the results.
Before: Upload documents, choose extraction fields, then review the results.
After: Upload documents, review a default set of extracted information, then decide what to add.
After: Upload documents, review a default set of extracted information, then decide what to add.
I kept upload as the first step. Review & Confirm brought editable results and configuration together on one screen, as the client wanted.
I kept upload as the first step. Review & Confirm brought editable results and configuration together on one screen, as the client wanted.
Why this fit
The default information gave users a starting point. They could see what AI found before requesting more.
The default information gave users a starting point. They could see what AI found before requesting more.
Sections kept the setup flexible. Users could add fields or create sections for information specific to the company.
Sections kept the setup flexible. Users could add fields or create sections for information specific to the company.
Extraction could happen more than once. I separated “Run AI Extraction” from “Submit” so users could refine the profile before finishing.
Extraction could happen more than once. I separated “Run AI Extraction” from “Submit” so users could refine the profile before finishing.
The default information gave users a starting point.
The default information gave users a starting point.
Users could add fields or create sections.
Users could add fields or create sections.
I mapped how users could check sources, correct values, and request more information before submitting.
I mapped how users could check sources, correct values, and request more information before submitting.
The assumption
I assumed a familiar profile structure and default fields would provide a useful starting point. Both needed testing with investment professionals.
I assumed a familiar profile structure and default fields would provide a useful starting point. Both needed testing with investment professionals.
I assumed a familiar profile structure would make setup easier to understand. I also assumed the default fields would provide a useful starting point. Both needed testing with investment professionals.
I assumed a familiar profile structure would make setup easier to understand. I also assumed the default fields would provide a useful starting point. Both needed testing with investment professionals.
I would judge the approach by whether users could review a value, add missing information, and complete setup without confusing extraction with submission
I would judge the approach by whether users could review a value, add missing information, and complete setup without confusing extraction with submission
The decision gave a starting point. The screens had to help users make it their own.
The solution
Users can review what AI found before configuring more.
Users can review what AI found before configuring more.
After document upload, AI fills editable fields. Users can submit or add more fields and extract again.
After document upload, AI fills editable fields. Users can submit or add more fields and extract again.
After document upload, AI fills a default set of editable fields. Users can review and submit that information, or add more fields and extract again on the same screen.
After document upload, AI fills a default set of editable fields. Users can review and submit that information, or add more fields and extract again on the same screen.
I made document upload the first step, removing the need to upload elsewhere and select files again.
I made document upload the first step, removing the need to upload elsewhere and select files again.
I made document upload the first step so users no longer had to upload files elsewhere and select them again during setup. AI uses those documents to populate the initial company information.
I made document upload the first step so users no longer had to upload files elsewhere and select them again during setup. AI uses those documents to populate the initial company information.
I grouped extracted values into editable sections so users could correct information directly.
I grouped extracted values into editable sections so users could correct information directly.
I presented extracted values as editable fields, grouped into sections by category. Users can correct the information directly or submit it if the initial set covers their needs.
I presented extracted values as editable fields, grouped into sections by category. Users can correct the information directly or submit it if the initial set covers their needs.
I put sources and confidence scores in a drawer so users could compare evidence while keeping the form in view.
I put sources and confidence scores in a drawer so users could compare evidence while keeping the form in view.
The client wanted sources and confidence scores visible alongside extracted data. I put them in a drawer opened from each field’s info icon, where users can inspect the evidence and choose another source while keeping the form in view.
The client wanted sources and confidence scores visible alongside extracted data. I put them in a drawer opened from each field’s info icon, where users can inspect the evidence and choose another source while keeping the form in view.
Warnings identify missing data. Users can enter values or rerun extraction without losing manual edits.
Warnings identify missing data. Users can enter values or rerun extraction without losing manual edits.
I placed warnings directly under fields where no data was found or extraction failed. Users can enter a value manually, go back and upload more documents and run extraction again, without losing their other manual edits.
I placed warnings directly under fields where no data was found or extraction failed. Users can enter a value manually, go back and upload more documents and run extraction again, without losing their other manual edits.
Users can add fields with a name, type, and extraction prompt.
Users can add fields with a name, type, and extraction prompt.
I kept field customization within the relevant section so users can extend the initial information. Each added field has a name, a type, and an extraction prompt to guide the AI.
I kept field customization within the relevant section so users can extend the initial information. Each added field has a name, a type, and an extraction prompt to guide the AI.
Users can choose predefined sections or create a custom section.
Users can choose predefined sections or create a custom section.
I offered predefined sections alongside a custom-section option, drawing from LinkedIn’s profile structure. Users can choose the categories they need or create a section for information the defaults don’t cover.
I offered predefined sections alongside a custom-section option, drawing from LinkedIn’s profile structure. Users can choose the categories they need or create a section for information the defaults don’t cover.
Results appear on the same screen for review before submission.
Results appear on the same screen for review before submission.
Results appear on the same review screen, so users can check them before completing setup.
Results appear on the same review screen, so users can check them before completing setup.
The wider redesign covered portfolio monitoring, Metrics & P&L, and data management.
The wider redesign covered portfolio monitoring, Metrics & P&L, and data management.
Company setup was part of a wider redesign covering portfolio monitoring, Metrics & P&L, and data management.
Company setup was part of a wider redesign covering portfolio monitoring, Metrics & P&L, and data management.
The design brought review and configuration together. But what evidence did I have that it made setup easier?
Validation
The client preferred the revision. I still needed to test whether setup was easier.
The client preferred the revision. I still needed to test whether setup was easier.
The client tried the prototype and liked the approach. That supported the direction, but didn’t show whether investment teams could complete setup with less effort.
The client tried the prototype and liked the approach. That supported the direction, but didn’t show whether investment teams could complete setup with less effort.
The client tried the prototype and liked the revised approach. That supported the direction, but it didn’t show whether investment teams could complete setup with less effort. The next step would be to test the assumptions behind the design.
The client tried the prototype and liked the revised approach. That supported the direction, but it didn’t show whether investment teams could complete setup with less effort. The next step would be to test the assumptions behind the design.
What I'd measure
Setup completion. Can users submit a company profile with the information the task requires?
Setup completion. Can users submit a company profile with the information the task requires?
Time to a reviewed submission. Does the revised design reduce work once checking and corrections are included?
Time to a reviewed submission. Does the revised design reduce work once checking and corrections are included?
Repeated extraction. Can users rerun extraction and review the updated values without confusing it with submission?
Repeated extraction. Can users rerun extraction and review the updated values without confusing it with submission?
Errors left in the profile. Faster setup shouldn’t come at the cost of accepting incorrect data.
Errors left in the profile. Faster setup shouldn’t come at the cost of accepting incorrect data.
Setup completion.
Setup completion.
Time to a reviewed submission.
Time to a reviewed submission.
Rerunning extraction without confusing it with submission.
Rerunning extraction without confusing it with submission.
Errors left in the profile.
Errors left in the profile.
I’d test whether analysts could find “Add field,” then choose a field type and use a saved prompt without help.
I’d test whether analysts could find “Add field,” then choose a field type and use a saved prompt without help.
Moving configuration into a popup kept the controls off the main screen. I’d test whether analysts could find “Add field,” then choose a field type and use a saved prompt without help.
Moving configuration into a popup kept the controls off the main screen. I’d test whether analysts could find “Add field,” then choose a field type and use a saved prompt without help.
Could they submit when the default fields covered their needs?
Could they submit when the default fields covered their needs?
Could they choose a field type and use a saved prompt without help?
Could they choose a field type and use a saved prompt without help?
Could they discover “Add field” while reviewing the profile?
Could they discover “Add field” while reviewing the profile?
What would prove me wrong
I assumed the default fields would provide a useful starting point. If users had to add most of the information they needed, I’d reconsider those defaults.
I assumed the default fields would provide a useful starting point. If users had to add most of the information they needed, I’d reconsider those defaults.
I also assumed editable results would make review easier. If users submitted incorrect values without checking their sources, I’d revisit how the screen prompts review.
I also assumed editable results would make review easier. If users submitted incorrect values without checking their sources, I’d revisit how the screen prompts review.
If users had to add most of the information they needed, I’d reconsider those defaults.
If users had to add most of the information they needed, I’d reconsider those defaults.
If users submitted incorrect values without checking their sources, I’d revisit how the screen prompts review.
If users submitted incorrect values without checking their sources, I’d revisit how the screen prompts review.
How I'd test it
I’d compare the first proposal with the revised prototype using matched company documents. I’d vary which version participants tried first.
I’d compare the first proposal with the revised prototype using matched company documents. I’d vary which version participants tried first.
Each task would include a missing value, conflicting sources, and a field to add.
Each task would include a missing value, conflicting sources, and a field to add.
I’d observe whether investment professionals could resolve those issues and submit a complete profile without guidance.
I’d observe whether investment professionals could resolve those issues and submit a complete profile without guidance.
Whether the defaults and review controls worked for investment teams remained an open question.
Reflection
I’d keep the editable results. I’d test the starting fields sooner.
I’d keep the editable results. I’d test the starting fields sooner.
What I'd defend
Keeping extracted values editable beside the options for adding information, with manual edits protected when extraction runs again.
Keeping extracted values editable beside the options for adding information, with manual edits protected when extraction runs again.
Keeping extracted values editable beside the options for adding information. Users could correct a value, add a field, and extract again without moving to a separate configuration screen. I’d keep manual edits protected when extraction runs again.
Keeping extracted values editable beside the options for adding information. Users could correct a value, add a field, and extract again without moving to a separate configuration screen. I’d keep manual edits protected when extraction runs again.
What I'd do differently
I’d test what investment professionals would keep, change, or add before treating that starting set as settled.
I’d test what investment professionals would keep, change, or add before treating that starting set as settled.
I’d also check whether the information collected during setup supports the next task: reviewing the company’s performance.
I’d also check whether the information collected during setup supports the next task: reviewing the company’s performance.
I’d use this screen to ask analysts what information they need first, then revisit the setup defaults.
I’d use this screen to ask analysts what information they need first, then revisit the setup defaults.
The company overview brings performance figures, progress, and milestones together. I’d use this screen to ask analysts what information they need first, then revisit the setup defaults.
The company overview brings performance figures, progress, and milestones together. I’d use this screen to ask analysts what information they need first, then revisit the setup defaults.
What I'd watch
Whether a populated profile encouraged people to accept values without checking them. I hadn’t established whether users would open sources or how they would interpret confidence scores.
Whether a populated profile encouraged people to accept values without checking them. I hadn’t established whether users would open sources or how they would interpret confidence scores.
Whether a populated profile encouraged people to accept values without checking them. Sources and confidence scores were available, but I hadn’t established whether users would open them or how they would interpret the scores.
Whether a populated profile encouraged people to accept values without checking them. Sources and confidence scores were available, but I hadn’t established whether users would open them or how they would interpret the scores.
I gave users a way to review the AI’s work. I’d question what it asked users to accept first.
Showcase
The new
first five minutes
The new
first five minutes
Contact
Have a role in mind?
imran.uxi@gmail.com
Copied
© 2026 Imran. All Rights Reserved.
Contact
Have a role in mind?
imran.uxi@gmail.com
Copied
© 2026 Imran. All Rights Reserved.
Contact
Have a role in mind?
imran.uxi@gmail.com
Copied
© 2026 Imran. All Rights Reserved.
Contact
Have a role in mind?
imran.uxi@gmail.com
Copied
© 2026 Imran. All Rights Reserved.
Contact
Have a role in mind?
imran.uxi@gmail.com
Copied
© 2026 Imran. All Rights Reserved.