CS01: AI for Virtual Data Room Set Up & Diligence

TL;DR

Led design of Datasite AI across the full deal lifecycle — from the first 0-to-1 AI chat experience through Bulk Q&A, self-activation, and June UI refinements, with the Reviewer experience on deck. Beta cohort achieved 93/100 adoption score, 50% WAU/MAU stickiness, and 205K platform events in 10 weeks. Research grounded in 8 early adopter feedback sessions and a full M&A journey map spanning advisor selection through close.

Business Problem

M&A deal teams work across the full deal lifecycle — but AI was only positioned for one phase of it.

Datasite AI launched as a document interrogation tool for active diligence. But research showed that users had a limited mental model of when the tool was useful — overlooking high-value opportunities in pre-diligence, VDR setup, Q&A management, and closing. The design challenge was to evolve the product from a point solution into a trusted, end-to-end diligence accelerator.

Failed first sessions

Users faced blank prompts with no guidance and closed the panel without asking a single question

Narrow mental model

Users thought the tool was only for active diligence — missing pre-diligence preparation, risk scanning, and Q&A automation

Manual Q&A bottleneck

Bulk buyer request lists with hundreds of questions were processed one at a time — a major unsolved pain point across all accounts

Research Artifact

M&A Deal Lifecycle Journey Map

I built this journey map to ground the team in exactly where deal professionals — Daniel (banker), Desmond (client manager), and Deborah (buy-side reviewer) — need AI support across the full deal from advisor selection to close. Phases marked AI active are where Datasite AI adds direct value. Click any row to expand jobs to be done and pain points by persona.

Research

Primary Personas

Three distinct users with overlapping urgency but very different jobs to be done — built from early adopter research across investment banking, private equity, and M&A advisory contexts.

DA

Daniel

28 · Deal Coordinator · Investment Banking

Inv. BankingCorp Dev

"I run the deal"

Run multiple VDR-led deals with speed, accuracy, and visible control
Surface deal health, risks, and blockers immediately
Make decisions resolvable in one or two clicks

DB

Deborah

42 · Due Diligence Manager · Private Equity

Inv. BankingPrivate Equity

"I manage due diligence"

Build an accurate risk and value view without post-close surprises
Clear folder structure and fast intuitive search
Alert on Q&A and new docs relevant to her workstream only

DE

Desmond

38 · Client Manager · M&A Advisory

Inv. BankingM&A Advisory

"I manage the client relationship"

Help clients stay informed and confident — without surprises
Client-ready "single view" status without technical noise
Mobile-friendly summaries for on-the-go prep before calls

Customer Research

Top insights from 8 early adopter sessions

Conducted with 8 clients across seven organizations in Spring 2026.

Accuracy sentiment trended positively across sessions — users rated accuracy ~4/5 in the most recent four calls, up significantly from the first session.

Insight 01

Users have a narrow mental model of when AI is useful

Teams defaulted to using Datasite AI only during active diligence — missing pre-diligence risk scanning, document gap analysis, and VDR setup opportunities. Phase-based onboarding with concrete sample prompts is the clearest path to expanding adoption.

Insight 02

Document search is the core value proposition

Being "led directly to documents" was consistently cited as the most critical capability — not summaries or analysis, but precise navigation to the right file and section. Users want the AI to locate, identify, and cite. Not to replace the review.

Insight 03

Bulk request list processing is the #1 feature gap

Every session surfaced this: uploading an Excel-based buyer request list and having AI auto-populate answers with citations. Users described it as the difference between the tool being "nice to have" and becoming essential to the workflow.

Insight 04

Citation-first transparency is non-negotiable for trust

Users won't act on AI answers without inline citations linking to the exact page and passage. The ideal experience: hover a citation and see file name, upload date, page number, and a document snippet. Trust is built or broken at the citation level.

Insight 05

Role-based "lens" prompting is the most exciting concept

Querying through a financial, legal, tax, or IT security lens — tailoring the AI's perspective to a specialist discipline — was described as the most compelling emerging capability, especially for late-stage diligence when third-party advisors surface unexpected issues.

In their words

"If something would take me 20 minutes to look at, I would say maybe half that using the chat box. When you do this for 100 requests, the time... I definitely think it's helpful."

Investment Banker

"I would get asked to find stuff and expecting that to take like 30 to an hour, and this tool, it would take 20 seconds to find the data."

Investment Banker

Now

Core AI Chat & Self-Activation

The first release established the foundational experience — document interrogation with cited Q&A, meaningful empty states, and a PLG self-activation flow for users who hadn't yet purchased. The design bet was that getting the first session right would drive everything downstream.

1

Meaningful empty states over blank prompts

Empty states demonstrate value immediately rather than presenting a blank prompt. Action buttons surface high-value starting prompts: advisor-type reviews (financial, legal, IT security, tax), document gap analysis, sensitive data identification, and Q&A table generation. Each maps to a richer parameterized prompt sent on click.

We're building a rules engine to make these prompts more meaningful and contextual to each user — for example, if a user is in an empty project without any content, we surface prompts to help them get the necessary content into the project first. We're also introducing a prompt library so users can discover, save, and reuse prompts across their workflow.

View Prompt Library prototype →

2

Citations as the primary trust mechanism

Every AI response surfaces citations linking directly to source documents and pages. In compliance-sensitive deal contexts, users need to verify answers before acting on them. Citation design is load-bearing for adoption — not a nice-to-have.

View Citations prototype →

3

PLG self-activation embedded in the product

For users who haven't purchased Datasite AI, the interface surfaces a contextual upsell in place of the chat panel — communicating product value in the moment of need and enabling direct purchase without leaving the platform. Activation is fully sales-free.

View Prototype →

4

Document processing status indicator

A known UX gap: previously no signal to users that documents are being indexed to Blueflame after upload. Users would try to interrogate documents that weren't yet ready and receive degraded responses with no explanation. The indicator surfaces this status clearly.

Document indexing status indicator — showing indexing in progress and success states

Left: indexing in progress · Right: files indexed, everything's ready

Beta Outcomes — Feb 4 to Apr 17, 2026

What the first 10 weeks showed

60/100

Overall PES — healthy for a newly launched product

93/100

Adoption score — near-universal activation since launch

50%

WAU/MAU stickiness after 10 weeks

205K

Total platform events in the measurement window

50 clients activated — most active user reached 43 Datasite days

1,836 history button events — 95% return-to-history rate across 95 users

Open-to-submit gap identified — 2,927 prompt field interactions, 88 submits (~3%). Empty states and hero prompts are the primary design lever to close this gap.

Strong firm breadth — Jefferies, Moelis, Houlihan Lokey, Piper Sandler, William Blair, and corporate users

The most consistently requested feature across all 8 feedback sessions. Upload an Excel-based buyer request list, and Datasite AI scans the data room to populate answers with citations and confidence scores — one row per question, review and approve before sending.

"Garrett at Harris Williams immediately identified the possibility of uploading a buyer tracker and having AI scan the data room to point to relevant files. Alec at Jefferies said this made him 'very excited.'" — from the feedback sessions

We're moving forward with Option B — a design that brings the user into an embedded workflow with the ability to ask follow-up questions to a prompt. Option B scales well with future use cases and covers the prompt entry clearly so there's no confusion about where to interact.

Making the interface cleaner, more polished, and aligned to well-known AI interaction patterns. This update also aligns with a broader Datasite design system refresh — raising the quality bar across the full product.

UI refresh & design system alignment

A cleaner, more polished interface aligned to well-known AI interaction patterns. Refreshing Datasite AI's visual language alongside the broader design system update — raising the bar across the full product. View prototype →

Stopping an active prompt

Users told us they'd submit a prompt and immediately realize they forgot to include something. We're allowing them to stop mid-generation — and when they do, the previously submitted prompt returns to the dialog box so they can adjust it or delete it and start fresh. View prototype →

Toolbar animations

Starting with motion in the header to build awareness of the tool's state — thinking, generating, complete, error. Animation reduces cognitive load and creates a more trustworthy AI interaction. More animation will follow as the system matures. See the animation concepts in the embedded prototype below.

Expanding AI across the deal lifecycle

Today, Datasite AI is primarily used during active diligence. But as the journey map shows, there are high-value opportunities at every phase. Two stand out: Advisor Selection — where AI can help bankers rapidly research a target company, build competitive landscape summaries, and develop valuation context for pitch preparation; and Management Presentations — where AI can assist in drafting, structuring, and personalizing presentation decks based on known buyer profiles and deal data. These are phases where deal teams spend significant time on high-stakes work that AI is well-positioned to accelerate.

Toolbar Animation Concepts — Interactive Prototype

Later

What's on the horizon

Longer-horizon capabilities that expand the product's addressable market, unlock new revenue streams, and move Datasite AI closer to the vision of a trusted, end-to-end deal collaborator.

File upload

Enabling users to upload files directly through Datasite AI — documents that don't yet live in the VDR. Each uploaded and processed page would count toward the per-page billing rate, directly expanding the revenue surface of the product and making AI a more active participant in deal document workflows.

Everest / Help Agent (AI Assist)

The first Datasite core feature managed through Datasite AI. Answers questions about how to use the Datasite platform itself, pulling from the Salesforce knowledge base — reducing support load and helping new users get oriented without leaving the product.

Buy-side Reviewer experience

Every bidder has access to the same documents — the advantage goes to whoever can comprehend, synthesize, and act on the content the fastest. Designing for the Reviewer experience means accelerating buy-side diligence: navigating unfamiliar folder structures, managing Q&A, and building risk narratives across thousands of documents under time pressure. This opens a new revenue stream and a meaningfully larger addressable market for Datasite AI.

Automatic agent routing

As Datasite AI gains multiple specialized agents (document interrogation, help, Q&A drafting, VDR setup), intelligently routing each prompt to the best agent for the job — without requiring the user to choose — becomes critical. The system identifies intent from the prompt and applies the right agent automatically, reducing friction and making the full capability surface accessible to every user.

Reflection

What this work is teaching me

Designing AI products for high-stakes professional contexts is fundamentally different from designing for consumer engagement. In the M&A world, every answer has a counterparty — wrong information doesn't just frustrate a user, it can damage a deal relationship. That raises the stakes of every design decision around trust, transparency, and error states.

The most interesting shift in this work is watching the product move from reactive to proactive. Datasite AI started as document interrogation — a deal team asks a question, the AI answers with citations. Bulk Q&A takes it further: upload a list, get answers auto-populated. The Datasite MCP vision takes it further still: the AI sets up the data room, invites users, manages permissions. Getting that transition right — from answer machine to deal collaborator — is the real design frontier I'm working toward.