Redesigning Disclosure Assist, IBM's AI tool for ESG reporting.

Context

Hundreds of questions, one client's documents.

A CDP or ESRS disclosure runs to hundreds of questions on climate, water, forests and more. Consultants answer each one with evidence scattered across a client's reports. Disclosure Assist uses generative AI to write the first draft, in four steps.

1. Import

   Pick the questions

   Choose the framework release, then the themes and sectors in scope. Matching questions come from a master list.

   CDP 2025 · CC / F / W

2. Knowledge base

   Upload the evidence

   Add the client's reports and policies. They're stored in a vector database the AI searches.

   .pdf · .docx · .xlsx · .csv

3. Generate

   Draft every answer

   An LLM answers each question from the knowledge base: all at once, by module, or one at a time.

   Batch or single question

4. Score

   Check and improve

   The AI predicts a likely score, explains the gaps, and the consultant edits before exporting.

   Likely score · 8.0 / 10

CDP framework

A global disclosure system where companies report on climate change, forests and water security.

ESRS framework

European Sustainability Reporting Standards, required under the EU's CSRD.

Materiality scope

The themes and sectors that apply to a client, which decide which questions they answer.

Audit

Every step had a page, but nothing connected them.

I had no brief document to work from, so I started by reading the old portal screen by screen to understand what it did and where it got in the way. Tap a marker to see each issue on the project page.

1 2 3 4 5

1. 1Instructions took half the page

   A permanent "How it works" essay filled the left half of every project, on every visit.

2. 2Buttons that looked unfinished

   Dotted outlines made the main actions look like placeholders, not buttons.

3. 3No sense of progress

   Four equal tiles gave no order and no status. Nothing showed which steps were done.

4. 4An easy-to-miss framework switch

   CDP and ESRS sat as two small tabs in the corner, yet they changed everything on the page.

5. 5Typos in the main labels

   "Response Genaration" and "fromsections" undermined trust in a tool that writes for clients.

Heuristic evaluation

Scoring the old portal against Nielsen's 10 heuristics.

I rated every heuristic on Nielsen's 0–4 severity scale, from cosmetic to catastrophic. For a tool that runs long AI jobs, the biggest gap was not knowing what was happening. Select a row to see where each problem appears.

Old portal · Generate response

H01

#	Heuristic	What I found	Severity
H01	Visibility of system status	Batch generation is called "a long running process", yet there is no progress bar, no count of answered questions and no status on projects.	Catastrophic
H06	Recognition rather than recall	The four steps are separate pages behind equal tiles. Consultants must remember the order and which steps they've finished.	Major
H08	Aesthetic & minimalist design	Half of the home page is a product overview that returning users never need, and the project page repeats it as "How it works".	Major
H04	Consistency & standards	Four button styles on one form: dark grey "Add File", green "Add Tag", blue "Submit" and dotted outlines elsewhere.	Major
H05	Error prevention	Uploading a new question file overwrites the existing data. The only warning is small red text in a dotted box.	Minor
H02	Match with the real world	Themes and sectors appear only as codes (CC, F, W, GN, CN, TO-EPM), and the header offers "Llm Configuration".	Minor
H09	Help users recover from errors	Empty tables show only a header row, with no message about what to do next.	Minor
H03	User control & freedom	A generated answer sits in a read-only grey box. There is no way to move to the next or previous question from it.	Minor
H07	Flexibility & efficiency of use	The draft and its score live on separate tabs, so reviewing each question means switching back and forth.	Minor
H10	Help & documentation	Help is a block of static text with typos, with nothing in context at the step where it's needed.	Cosmetic

Principles

Make the AI's work easy to follow and check.

UX laws

Six laws of UX, applied to real screens.

Each decision in the redesign traces back to a well-known principle of how people perceive, decide and act.

01

Goal-gradient effect

People speed up as they see themselves getting closer to a goal.

In Disclosure AssistEvery project card and table row shows a progress bar, with green for complete and red for a project that needs attention.

02

Zeigarnik effect

Unfinished tasks stay on our minds more than finished ones.

In Disclosure AssistThe dashboard leads with what's left: "120 answered / 30 not answered" and "23 questions scored below 60%".

03

Hick's law

More choices, slower decisions.

In Disclosure AssistThe setup wizard asks for one decision per step, starting with a single choice: CDP or ESRS.

04

Jakob's law

People expect a product to work like the others they already use.

In Disclosure AssistTables, search, pagination and file uploads use standard Carbon components, the same patterns found across IBM's tools.

05

Law of proximity

Things placed close together are read as related.

In Disclosure AssistThe draft answer sits right next to its analysis and likely score, instead of on a separate Scoring tab.

06

Law of common region

Things inside a shared boundary read as a group.

In Disclosure AssistEach step is its own tile with an Open link, and the "How it works" help moved to a separate side panel.

Before & after

Same features, same AI, a clear order to follow.

The old project page and the new project dashboard, shown whole and side by side. Select either screen to see it full size.

Project page · old vs new

Side by side Stacked

OldLegacy portal

• Half the page is instructions
• Four equal tiles, no order
• No status or progress

NewCarbon redesign

• Help moved to a side panel
• A setup wizard for the four steps
• Completion, low scores and data gaps

Or drag the slider to see the same page change in place.

BeforeAfter

Tiny CDP / ESRS tabs in a corner
Framework tabs above the steps they control

Dotted placeholder buttons
Carbon tiles with a clear Open link

An empty "Dashboard" tile
Completion, low scores and a data gap table

Instructions in the main column
"How it works" in a side panel

Screen by screen

Five parts of the consultant journey, compared side by side.

ProjectsImport questionsKnowledge baseGenerateReview & score

Before

After

• One lone project tile became cards plus a searchable table
• Every project shows its progress and status
• The product overview moved to a side panel

My redesign

From a new project to a scored draft.

The redesigned consultant journey, built on Carbon. The setup wizard walks a first-time user through all four steps. Returning users go straight to the dashboard.

Setup wizard Project dashboard Projects Review & score Sign in

Four steps with a progress indicator. Each step explains why it matters, and Next unlocks once the step is done.

1. Step 1Choose framework
2. Step 2Import questions
3. Step 3Knowledge base
4. Step 4Generate & score

Visual system

IBM Carbon, used as intended.

Instead of the old portal's custom mix of colours, the redesign uses Carbon's tokens and components as they are.

Outcome

A tool that shows where you are and what's left.

This case study covers the consultant journey; I haven't included usage data. The next step would be to watch consultants run a full CDP questionnaire on the new screens and measure how long review takes per question.

Feedback

What the project manager said about the work.

From IBM's internal feedback tool, received on May 23, 2025.