Knowledge & data
Internal firm knowledge base
How does the deadline-extension process go again? Which text module do we use for overdue documents? Who covers for Ms Berg? In most firms the same person answers questions like these – between two sets of accounts.
AI knowledge system
01Intro
The firm knowledge base makes standards, responsibilities, text modules and processes findable – as a structured system the team looks things up in instead of asking around. And it can do more than a wiki: AI searches it, summarises, files new notes and step by step takes over defined routines – from meeting minutes to the morning deadline overview. The decisive cut here: it's about firm knowledge, not client files. Those stay in the specialist system, where they belong.
Up front: client data doesn't belong in this system – the knowledge base holds processes and standards, DATEV and specialist systems hold the files. I won't promise a "second brain in one go": the entry point is one bounded area of knowledge, otherwise it fails on upkeep. And the AI never writes unchecked – every change to a standard goes through sign-off.
02Demo
The knowledge base of the fictional “Reiter Tax Firm” – on the left the order in which the firm's knowledge lives (client data deliberately stays out), on the right three everyday moments with it. Click through and watch where it lights up in the tree.
The architecture behind it has two layers: at the bottom the vault – structured folders and notes that belong to the business. Above it the AI layer with documented rules: it reads only at first, then writes with approval, and sensitive areas stay locked. Calendar and tools dock on, routines run within clear limits.
The most honest proof of this model: this website came about exactly this way – I work in such a system myself every day, with AI as an employee inside it.
03How it works
- 01
Take stock: where does your firm knowledge live?
In heads, email threads, Word templates from 2019, notes on the pinboard. Plus: which questions interrupt whom most often, and what falls away when the longest-serving colleague leaves?
- 02
Build the structure, small and documented
A clear ordering system with simple, written-down rules: standards, processes, responsibilities, text modules. Boring, decisive – and the point where the line to client data is firmly drawn in.
- 03
Fill one area first
Not all at once: the area with the greatest pain comes first – often the deadline processes or the onboarding of new engagements. Existing templates move in curated, not copied.
- 04
Connect the AI: reading first, then writing
First the AI only answers questions from the knowledge – with a source reference. Once that holds, writing with sign-off is added: minutes, entries, change suggestions to standards. Rights and limits are set down as a documented work instruction.
- 05
Add routines – step by step
The morning overview with deadlines and follow-ups, the meeting minutes, the reminder list: the AI takes over defined tasks within clear limits – each routine only once the previous one runs.
04Stages
The entry point is Tier 1, the team knowledge base – for firms, it's exactly the right start.
Entry: the queryable knowledge base
One bounded area – say deadline processes and responsibilities – is structured and made queryable. New colleagues look things up instead of interrupting the team lead.
Extended: the AI works along
Minutes, summaries, change suggestions to standards – the AI writes along in the system, every change with sign-off.
Full build: routines run
A morning deadline card, follow-up reminders, reports – defined routines run on their own within clear limits. The knowledge base becomes the work system.
05Technical implementation
- The knowledge layer belongs to the firm: a Markdown vault (e.g. Obsidian), Notion or a comparable system – content without vendor lock-in.
- The AI layer is an assistant with access to the vault (e.g. Claude Code as the agentic layer) – with documented rules: which areas, which rights, which routines.
- Hard system boundary to client data: DATEV and specialist systems stay authoritative; the knowledge base references processes ("how do we do X"), never file contents.
- Tiered rights logic: reading first, writing only in defined areas with sign-off, sensitive areas excluded.
- Upkeep routines are set up alongside – owners and a review rhythm per knowledge area, otherwise the base ages like any wiki.
- Typical content: deadline logic, firm standards, text modules, responsibilities and cover arrangements, onboarding processes, meeting minutes.
for context · This shows how something like this usually gets built – not how it has to be built for you. What actually fits is worked out in the project: around your existing setup, your team, and what stays easy to maintain.