Transcript Review Tool
Transcript Review Tool
Compliance | Workflow design
Compliance | Workflow design
2 months
2 months
AlphaSights
AlphaSights
context.
context.
AlphaSights produces transcripts of expert calls as a core part of its research offering, processing around 3.5–4k transcripts a day for 150k investment professionals. Before they can be published, an internal compliance team reviews each one to make sure nothing non-compliant gets through.
AlphaSights produces transcripts of expert calls as a core part of its research offering, processing around 3.5–4k transcripts a day for 150k investment professionals. Before they can be published, an internal compliance team reviews each one to make sure nothing non-compliant gets through.
before.
before.
after.
after.
problem.
problem.
The tool they used for this hadn't been touched in over five years. Reviewing a transcript meant clicking between tabs and windows, copying and pasting, and constantly losing context. A previous attempt to speed things up with AI flagging ended up adding to the noise rather than reducing it.
As the sole designer I needed to understand the existing workflow, figure out what was actually slowing people down, and build something fast enough to clear the backlog without asking the team to relearn how to do their job.
The tool they used for this hadn't been touched in over five years. Reviewing a transcript meant clicking between tabs and windows, copying and pasting, and constantly losing context. A previous attempt to speed things up with AI flagging ended up adding to the noise rather than reducing it.
As the sole designer I needed to understand the existing workflow, figure out what was actually slowing people down, and build something fast enough to clear the backlog without asking the team to relearn how to do their job.
workflow.
workflow.
To understand the workflow I conducted discovery shadowing sessions with 10+ users across seniority levels and validated findings with key stakeholders.
To understand the workflow I conducted discovery shadowing sessions with 10+ users across seniority levels and validated findings with key stakeholders.
design outcome.
design outcome.
The core decision was collapsing a multi-tab workflow into a single continuous view. Three panels, everything within glancing distance.
01. keep context throughout
01. keep context throughout
Keeping the overview panel on the left meant users can reference the details of the experts and companies mentioned wherever they are in the transcript.
Keeping the overview panel on the left meant users can reference the details of the experts and companies mentioned wherever they are in the transcript.
02. compare edits
02. compare edits
Side by side comparison puts the cognitive load on the user to spot the difference. Encoding changes through colour and underlining directly in the transcript means reviewers can track what's been redacted or added at a glance, without holding two states in their head at once.
Side by side comparison puts the cognitive load on the user to spot the difference. Encoding changes through colour and underlining directly in the transcript means reviewers can track what's been redacted or added at a glance, without holding two states in their head at once.
03. never lose your place
03. never lose your place
Putting flags, comments, and redactions in a dedicated panel means the transcript stays clean to read and edit. Reviewers don't have to context switch between following the content and actioning what the AI has flagged - both are visible at once, just not competing for the same space.
Putting flags, comments, and redactions in a dedicated panel means the transcript stays clean to read and edit. Reviewers don't have to context switch between following the content and actioning what the AI has flagged - both are visible at once, just not competing for the same space.
04. dark mode
04. dark mode
One of the biggest problems for our users was eye strain from reading transcripts all day. Dark mode was a non-negotiable for the update. The design challenge was getting contrast and legibility right, which meant re-evaluating every colour token in the existing design system rather than just inverting it.
One of the biggest problems for our users was eye strain from reading transcripts all day. Dark mode was a non-negotiable for the update. The design challenge was getting contrast and legibility right, which meant re-evaluating every colour token in the existing design system rather than just inverting it.
outcome.
-37%
Reduction in review error rate
+56%
Improvement in turnaround time
“This is such a game changer for our team, we finally take a bite out of our huge backlog.”
“This is such a game changer for our team, we finally take a bite out of our huge backlog.”
reflection.
reflection.
There were a lot of UX nuances in this workflow, so once I had high-fidelity mocks in Figma I used Cursor to build prototypes. This made iteration faster and simplified handoff - engineers could reference working interactions rather than trying to interpret static screens.
In future I'd push harder on the relationship between AI flags and reviewer attention. The current design surfaces flags clearly, but it treats all flags with roughly equal weight. In practice some flags need immediate action and others are low-confidence noise. A more opinionated hierarchy of flag severity would let reviewers triage faster rather than reading through every flag sequentially.
There were a lot of UX nuances in this workflow, so once I had high-fidelity mocks in Figma I used Cursor to build prototypes. This made iteration faster and simplified handoff - engineers could reference working interactions rather than trying to interpret static screens.
In future I'd push harder on the relationship between AI flags and reviewer attention. The current design surfaces flags clearly, but it treats all flags with roughly equal weight. In practice some flags need immediate action and others are low-confidence noise. A more opinionated hierarchy of flag severity would let reviewers triage faster rather than reading through every flag sequentially.