Company context

Harriet answers with the knowledge you give it. A well-organized knowledge base is the difference between a clear, citable answer and a vague one, so structure it around the questions your people actually ask.

Write for the way people ask

People get better answers when the knowledge base mirrors how they ask questions, not how compliance PDFs get filed. Name documents the way someone would type in chat: "How do I submit an expense claim?" rather than "Expenses policy v7 final_FINAL". Add a short summary to each document in the console so admins, and Harriet, can tell similar documents apart at a glance.

Split long policies into topic sections where possible. Replace a single "operations manual" upload with separate Expenses, Travel, and Procurement articles, each with a clear scope in its title. Someone asking about trains then finds Travel directly instead of wading through procurement thresholds.

Keep answers unambiguous

When several documents could match a question, answer quality depends as much on how you publish as on the AI. If two documents answer the same question differently, Harriet may surface both and people will see the confusion. Resolve overlaps in the source content, and keep one canonical document per topic per jurisdiction or audience.

• Split by audience and jurisdiction. Prefer separate files for materially different rules (countries, entities, business units) over one giant document that mixes them. Publish a UK handbook, a US handbook, and an India handbook rather than an 80-page global one, so a question from the UK hits the UK file first.
• Link between variants. Reference related documents in the body ("If you are in India, see …") instead of pasting every variant into one file.
• Assign one owner per document. Someone must update the document when policy, law, or process changes. Out-of-date content is the top cause of contradictory answers, and duplicated policy with no owner drifts.
• Align metadata with content. Use tags consistently, and make sure country and region metadata match the rules the file actually describes.

Sensitive material belongs behind user groups: restrict a document to specific groups and only members of those groups (and users with appropriate admin access) see it when Harriet searches company knowledge. Restrict the sensitive chapter rather than hiding everything in one large file, and confirm the group assignment before uploading, since mis-assignment is the most common cause of both "the wrong people saw this" and "nobody can find this".

Spot gaps from recurring questions

Harriet is only as good as the corpus you maintain. Gaps show up as repeated questions, low-confidence answers, or tickets opened right after a chat. Look in three places:

1. Review ticket themes. Tags or subjects that keep appearing (expenses, vendor setup, IT access) point at missing self-serve answers.
2. Check what your team answers by email. Anything answered repeatedly in the team's inbox should be a knowledge article instead.
3. Ask pilot champions. What do people still ask in Slack or Teams that Harriet should be handling?

When you find a gap, draft a short, factual article in plain language, assign group and geography metadata if the answer is not universal, and link it from related articles so Harriet can move between overview and detail. Five tickets a month asking "Do we pay for home internet?" become a two-paragraph policy snippet with an effective date, and similar tickets drop the next month. Date-stamp time-bound programs ("2025 travel policy FAQ") so you can retire them cleanly.

How the rest of Harriet draws on it

Company knowledge is not just for chat answers. Skills that include knowledge search ground their output in the same corpus, agent steps in workflows can search it when the Knowledge base skill is enabled, and every channel where people meet Harriet draws on the documents the asking person is allowed to see. Improving one article improves every surface at once.