Companies keep years of board meeting papers, reports and minutes. To trace one decision, people used to open them one by one. Now they ask in plain language, and every answer shows which document and page it came from. It is the feature customers asked for.
Smart Assist: an AI chat assistant
with answers a board can trust
I designed Smart Assist, a chat-based search layer for Diligent Boards that lets board members and administrators ask questions of their materials in plain language and get answers grounded in citations they can verify, instead of a list of documents to open and read.
About the visuals: all screens and videos come from my prototype with synthetic data. It is not the final product, and no customer data appears.
In regulated companies, these people decide whether staff may use AI on sensitive board documents. They asked for answers they can check, an assistant that says when it does not know, no leaks, and clear privacy information. The design is built around each one.
Nobody reads a 800-page board pack word by word, yet board members are still expected to read what they are sent. Every answer links back to the source page, so the tool shortens the search, not the reading. Suggested first questions help less technical members get started.
Search returned documents.
Boards needed answers
Preparing for a meeting meant digging through several meeting packs. Smart Assist answers in natural language and pulls decisions, discussions, risks, actions and approvals out of years of board material, without searching book by book. I led the design from 0 to 1: UX and UI, and contributed to the guardrails and the AI's response behaviour.
Three roles, one shared need: answers they can check
Sales and customer service kept hearing requests for a chat-like way to search board documents, so I tested an early concept as a Figma Make prototype: 8 sessions, 15 participants, 10 organisations across the US, Canada, UK and Kenya. The sample is small and qualitative, so read the percentages as direction, not statistics.
Three roles use Smart Assist: admins who assemble the materials, directors who govern from them, and executives who answer to the board.
Board admins & staff
“The pack goes out Thursday and a director emails asking what we decided on that acquisition in 2021. I’m digging through PDFs. I’d love the help, but legal signs off before anything touches our board data.”
Board members
“I sit on three boards. The pack lands the night before, 400 pages. I don’t need it summarised, I need to know where to push, and what we said last time.”
Chief executives
“I’d rather hit the hard question in prep than in the room. And when a director asks where a number came from, I need to point to the page.”
One role also decides whether it launches: board admins gatekeep every AI rollout (about 60% of sessions), wary of directors pasting board packs into ChatGPT and of AI summaries replacing the duty to read.
Three key takeaways and decisions it has driven
Focused and assistive views
The business wanted an AI-first product. I kept Smart Assist assistive, not autonomous (yet): closer to grounded search than to an agent. I stayed honest about what it does today, laid patterns for where it is heading, and let intent choose between two modes:
- Focused full screen
A “sealed”, focused experience for cross-book research. The assistant takes over the screen and the answer becomes the focus.
Default: outside a book
- Assistive docked panel
An AI-assisted view beside the book, keeping the core app’s actions (navigation, comments, annotations) in reach.
Default: inside a book
Reactions split by role: admins leaned to full screen, directors to the panel.
"The board would maybe prefer the panel, but I'd prefer full screen, it's easier to navigate around."
Testing found no clear winner, so the PM and I set a context-aware default and left the final call to analytics: full screen outside a book, panel inside one. Users can switch, and the choice is remembered within the session.
Scope that follows where you are
Outside a book, Smart Assist searches all board materials. Inside one, it defaults to that book, so answers stay where the user is reading, and one switch widens the search to every book. A clearable context chip keeps the scope visible, a fix that came out of internal dogfooding, where wrong-book answers hurt trust most.
For a board, an answer you can't verify is a liability
Board materials are among the most sensitive documents a company holds, so adoption depends on whether the AI can be trusted: every answer has to trace back to a source and sensitive content can't leak.
Footnote-style inline citations got strongly positive reactions in all sessions (8/8). Users clicked them to verify, and the source-page preview was praised. But the book title wasn't visible without clicking through (5/8 sessions), and people need to know which document a citation points to at a glance.
Every claim carries a marker you can follow
Hovering a marker shows the book, document and page, so the book title is visible without clicking. If the source is in the book the user is reading, clicking scrolls to the passage. If it's in another book, it opens in a side drawer, so they can check without losing their place. Markers changed from 1a, 1b, 1c to plain 1, 2, 3, which tested better.
Next: highlight the exact passage
Highlighting the exact chunk on the source page a fact is drawn from would let users see where a claim comes from at a glance.
Highlight in the source chunkMake the first question easy to ask and safe to send
Starter prompts instead of a blank box
A blank input intimidates less tech-savvy directors, so the first run leads with starter prompts that show core use cases, with separate sets for admins and directors because their jobs differ. A “learn more about Smart Assist” prompt lets anyone check what the assistant can do before asking.
Smart Assist welcome state: full-screen first run with starter prompt cardsClicking a prompt fills the input; it doesn’t send
The user controls what goes out. I moved the prompts from exposed cards with one fixed prompt to chips, which are more versatile and cover more options. Analytics now track which prompts are used most; those results are not in yet.
Starter prompts and shareable prompt libraries (~80%) were highly valued for easing blank-canvas starts and lowering admin support load. Guided onboarding matters most for less tech-savvy, often older directors, who valued social proof, curated prompts, and a first-run walkthrough with a data-handling explainer.
“A lot of board members are older and not as tech savvy. Anything you can do to help them out saves them from coming to us with every little question.”
Data-policy answers stay one click away
Cautious customers won’t adopt an assistant they can’t trust with sensitive board data, and they want reassurance up front. “Learn more about data policy” anchors the welcome state and stays in the More menu in every conversation state, one click from the Help Center.
Learn more about data policy link opened from the welcome state, leading to the Help CenterUsers decide what the assistant keeps
Board conversations are sensitive, so users decide how long their chats are kept, and can rename or delete a single chat or clear their whole history. Retention is also set by organisation policy; my design covers the user-controlled side.
AI answering the right way, within limits
Trust is decided as much by how the assistant behaves at the edges as by what it answers, and that behaviour lives in written prompts. I pushed for us to own them, so its behaviour was designed, not inherited. I proposed the changes, the PM finalised them, and we tested together. The assistant has three grounding states: supported, partially supported and not found. In governance, partial is the most common case, so those answers lead with what was found and then name the gap, reading as a partial result, not a refusal. The system prompt sets its identity, tone and sourcing rules, and after the MVP round we revised it to tighten accuracy and guardrail scope. The tabs below show where it deliberately doesn’t just answer.
Smart Assist surfacing two conflicting annual-report filing dates from board minutes and a governance committee pack, and escalating the statutory dateWhen sources disagree, it surfaces both and escalates anything statutory rather than silently resolving.
Smart Assist answering a definitional EBITDA question and labelling it as not coming from the board's materials, with an offer to ground itDefinitional answers are labelled as not coming from board materials, with an offer to ground them.
Smart Assist declining a stock-picking request and redirecting to what the board's materials actually sayNon-governance and advice requests are declined and redirected to what the materials actually say.
Smart Assist saying it couldn't find information on the board's cryptocurrency policy instead of fabricating an answerBelow the relevance threshold, it says so plainly instead of fabricating.
It started from an answer-or-fail spec; dogfooding then surfaced misread negations and past-tense questions, wrong-book answers and shaky financial comparisons, and each round added explicit rules for them.
Customization that enhances, never filters
Because we don't train on conversations, personalisation has to carry relevance deliberately. In a governance setting that can never mean hiding information: it ranks, frames, and contextualises, never restricts. An audit member sees audit content first, a CFO outcomes first, but nothing is filtered out of the rest.
Set by an admin and visible to everyone. Five inputs (name, country, size, industry and a governance description) help the assistant read documents accurately: resolving "we" and "our board", recognising terminology, judging scale. It never uses them to generate regulatory or sector guidance from training data.
The two interfaces look different because one is the live product, mid design-system change, and the other is my prototype in the new theme.
Shapes how answers read, never what is retrieved.
- Role & focus order and frame results (audit first, actions first, outcomes first), and are set aside for legal or statutory questions.
- Custom instructions set format, length, and tone, overridable in any message, and never touch citations or grounding.
- Prompt preview shows exactly what the assistant sees, with an on/off toggle.
When inputs conflict, an explicit priority order resolves it: the in-message request wins first, then saved custom instructions, then role and focus, then organisation context, with governance-appropriate defaults underneath. The hard constraints (citations, grounding, permissions) override everything. Personalisation never touches retrieval, only how the answer reads after the right passages are found, and anything users type into settings is fenced as preferences, not instructions.
Faster, verifiable answers for every role
From tracing decisions to backing up numbers, each role gets an answer it can check against the source.
- Trace decisions across years. Cross-book search and decision tracing was the top use case for 70% of research participants. Smart Assist answers it from every pack at once, instead of one book at a time.
- Prepare faster. Directors find past discussions in an 800-page pack, with every answer linked to its source page. The tool shortens the search, not the reading.
- Back up every number. Executives point to the exact page behind a figure, instead of hunting for where it came from.
- AI that legal can approve. Verifiable answers, no leaks and the data policy one click away give the people who sign off on AI something they can say yes to.
What this project taught me about designing AI
In AI, the prompt is interaction design
Sharing the prompt with a central Platform team that serves every AI product in the org slowed us down. Once we could own the system prompt and most of the guardrails, we moved faster and could run thorough eval testing. For a conversational product, UX is as much about the prompt and the conversation as about UI, and design isn't automatically in the room for that work, so UX has to be brought in early.
Two right answers can conflict
Admins were right that directors must engage with the materials. Directors were right that no one reads 800 pages word by word. I stopped looking for a winner and designed for the tension: faster answers that always lead back to the source.






