An ops team of four supports about ninety people. The employee handbook is a 60-page PDF last opened by someone other than its author in March. There are eleven policy documents in a folder, an expenses guide, and three onboarding decks presented live every quarter. New starters read none of it. They ask their manager, who asks the ops team, who answers the same carry-over holiday question for the fourth time this month.
That is a retrieval problem, not a content problem. Here is the order to do the setup in.
Decide who can open it before you upload anything
Onboarding material is internal, so the access level you want is Owner group — "Group members only — people who have been granted access to your owner group." Signed-in members of your organization can open it; nobody else can, link or no link. Set it in the document editor on the Settings tab under Access Level, or pick it on the upload queue's Defaults for all strip so every file in a batch lands with the right level from the start.
Two neighbours are worth knowing. Owner admins only narrows a document to administrators of your organization — the right home for a policy still in draft. Registered means any signed-in Docutrain user, which is broader than your company and almost never what an HR team wants. See access levels explained.
Then bring the people in. The Users tab has an Invite User button asking for an Email Address; ordinary owner admins always invite registered members, and the Owner Group field is pre-filled and read-only with your own organization. Only the primary admin — the group's first administrator, who also handles billing — or a super admin can invite someone as an Owner Admin.
Invitations consume seats, and this catches people out: your plan's member allowance counts current members plus unexpired pending invitations. Hit the cap and the invite flow blocks with "User limit reached for your plan. Upgrade your plan to invite more users." The same check runs when you approve a join request, and single-seat plans exist — so a team intending to invite ninety people should sort its tier out first.
The invitation email names the inviter, the workspace and the role, and expires after 30 days. If the address already belongs to a Docutrain account, nobody signs up twice — they are added immediately and get a "You've been added to [organization]" email instead. Pending invitations sit in the Users list with a Pending Invitation chip and an Expires in N days warning; the row's Actions menu offers Resend invitation and Delete invitation. See users, roles and invitations.
Getting the material in
The handbook and policy PDFs go up through the PDF / DOC tab, three at a time if you queue them. They are prose with the occasional figure, so answer Document when Docutrain asks What type of PDF is this?
The onboarding decks need a different decision. Docutrain accepts .pptx directly, but the upload screen tells you plainly what you get: "To extract slide images and include them in AI responses, export your presentation as a PDF and upload that instead. Direct PowerPoint uploads only support text extraction." A PPTX upload trains on the slide text, which is enough to answer from, but the slides never appear in chat. If your benefits deck has a table easier to look at than to describe, export to PDF and choose Slide deck. Docutrain then captures every page as a full slide image and can show whole slides inside answers, trimming letterboxed margins automatically. Google Slides works the same way.
What a new starter actually meets
Two settings on the Messages tab shape the first thirty seconds.
Intro message is the formatted introduction shown in the chat sidebar, supporting paragraphs, bold, lists and links, with a live preview. The detail worth exploiting: if the introduction contains a list of sample questions, readers can click any of them to ask it straight away. Writing the four questions every new hire has turns the welcome text into a working menu. Regenerate abstract & keywords will draft a summary, sample questions and keywords from the document's own content to start from.
The Composer prompt sub-tab is an optional card shown above the chat input before the first message. Prompt above chat input holds the content, Card label gives it a header (the default reads "Getting started"), Allow user to dismiss decides whether readers can close it, Hide after chat begins removes it after the first message, and Dismiss persistence controls whether a dismissal lasts the session or about a year. A short card saying what the assistant does and does not cover — "policies and the handbook, not payroll queries" — sets expectations before anyone is disappointed.
If your team lives in Slack, the assistant can go there too. Docutrain's Slack integration pairs specific channels with specific documents; in a paired channel anyone mentions the bot with a question and the answer appears in the channel, so the whole team sees it. For documents restricted to your owner group, the bot checks the Slack user's email against Docutrain's records before answering and declines when they do not have access. There is no self-serve setup screen yet; installation and channel pairing are arranged with the Docutrain team.
Checking that any of it landed
Reading a policy and understanding it are different things. In the document editor, the Quiz tab has a Generate Quiz button. Docutrain sizes the bank from the document's length — minimum 10 questions, maximum 100 — and writes each one with an explanation and page citations.
Read the bank before showing it to anyone. Each question expands to show its options with the correct answer marked, the explanation and its Page chips; Edit lets you rewrite the question, adjust the options or fix the explanation.
Generating questions does not show them to readers — the Quiz toggle under UI Options → Engagement does that, and stays disabled until a bank exists. New starters then get a Quiz button in the chat tools, pick a length of 5, 10, 15 or 20 questions, and work through them with instant feedback and a review screen at the end. Attempts are unlimited, and the Quiz Statistics panel reports total attempts, average, highest and lowest scores. A low average on material your team knows well usually says more about the policy's wording than about the reader. Quizzes has the full picture.
When the policy changes
Open the document's Retrain tab and choose Begin Retraining. Replace all data suits a new edition of the handbook: the previous content keeps answering until the new content is fully processed, then the switchover happens in one step, so the assistant never goes dark. Add to existing data appends — right for a new appendix — and can be undone from the History tab, which Replace cannot.
Three follow-ups after a substantial retrain. The abstract is flagged stale and should be regenerated. Quiz questions are not regenerated automatically — they came from the old content, so review them. And leave a note on the training row in History: "v4 handbook, new parental leave section" reads better in eight months than a bare timestamp. Retraining and history goes deeper.
Finally, read the conversations. The Conversations tab is a live feed of every question asked, with a red dot on answers a reader marked unhelpful and the reason they chose — Factually inaccurate, Not helpful, Off topic and the rest. Those red dots are the cheapest content roadmap an ops team will get: a recurring question with a poor answer almost always means the handbook is missing a paragraph.