You're offline, showing the last version we saved.

Part of our AI in accounting coverage. See the full AI for accountants guide →

Ask a general AI tool about a client and it knows nothing: not the property deal they mentioned, not what you promised on the last call, not your rule for chasing late records. That is why so many AI answers come back generic. The fix is a second brain: a documented, organised store of your firm’s context that your AI and your whole team can read. I made the case for one when I opened day two of the AI in Practice Summit in November 2026. This guide turns it into something a small firm can build in a week, safely. Tool details are as at November 2026.

What a second brain is, and what it is not

Your first brain is what sits in people’s heads and private inboxes: a client mentioned a property deal, we promised something on that call but nobody remembers what, and it is stuck in one person’s inbox. In the second brain the same facts are written down, by client: they want to buy a property, we promised advice, the bank wants the numbers by Friday.

The idea is not mine; others have written about second brains for years. What changed is that AI can now read one: chat windows gave way to projects, which hold context, and then to MCP connectors, which reach into your other systems. A second brain does not replace your people. It gives everyone, and the AI, the same picture of what is going on.

You’re losing knowledge every time a team member leaves.

Trent McLaren, co-founder of The Firm, at the AI in Practice Summit, November 2026

For the business case, read our piece on the knowledge ledger. This guide is the build.

What goes in and what stays out

Split the content in two, the way I did on the day: client intelligence (each client’s goals, what they told you, what you promised) and team intelligence (how you work, who knows what, decisions made and lessons learned). Then be strict about what never goes in.

Goes inStays out
A one-page profile per client: entity structure, year end, goals, key contacts, communication preferencesTax file numbers, Social Security numbers, National Insurance numbers and other government identifiers
Summaries of advice given and decisions made, with datesBank account numbers, card numbers, passwords and portal logins
Meeting notes and action itemsFull source documents you have not checked for sensitive data
Your SOPs, checklists and review notesHealth, family or other sensitive personal information you do not need for the job
Pricing rules: how you scope, what triggers a fee review, what is in or out of scopeAnything a client asked you not to share or store
Email templates and your firm’s style rulesSuperseded versions of policies and advice

Pricing rules go in as a method, not a price list; the partner still sets each fee.

How to structure it

An AI reads a second brain the way a new starter would, so short, dated documents beat one giant export. A layout for a small firm:

  • Firm rulesone document with your tone of voice, sign-off rules, what the AI must never do, and who approves what.
  • Clientsone folder per client, holding the profile, a running log of advice and decisions, and meeting notes.
  • How we workSOPs and checklists for each recurring job, plus a lessons-learned log.
  • Who knows whatspecialisms and backups, so work does not stall during leave.

Put an owner and a “last reviewed” date on every document. The AI cannot tell a current policy from a 2023 one unless you say.

Once the documents exist, turn repeat jobs into skills. Gus McLennan of AI Dojo described a skill at the summit as a package of files with instructions telling the AI how to do a particular task. On the “Where to start with AI” panel, Andy Teece of Atomic Business Advisers said his firm built its house method into shared skills, rolled out across the team on a business plan.

Where it lives: the tools that can read it

You need two things: a store (where the documents sit) and an AI that can read the store. The store is often what you already run: SharePoint, Google Drive, SuiteFiles, Notion, or your practice management system such as Xero Practice Manager, FYI, AccountKit or Karbon. My team keeps its documentation in Notion and works on it through Claude. Pick one AI and commit to it, then ask your vendors whether they will let you connect it.

These are the features that let each AI read a firm’s context. Last verified 20 November 2026.

FeatureWhat it holdsSharing
Claude ProjectsUploaded files plus project instructions. Paid plans switch to retrieval as knowledge nears context limits, which Anthropic says expands capacity up to 10 times.On Team and Enterprise, shared with named members or the whole organisation, with view or edit permission.
Claude Google Drive connectorGoogle Docs added to a project stay synced to the latest version. Text only; embedded images are not read.Private projects only. An Owner must enable connectors for the organisation first.
ChatGPT ProjectsFiles, project instructions that override global custom instructions, and project memory.On ChatGPT Business, shared with workspace members, with chat or edit access.
Microsoft 365 Copilot NotebooksWord, Excel, PowerPoint, PDF and OneNote references, plus SharePoint folders or sites. For a shared location, Copilot picks up to 300 of the most relevant files you can access.Requires a Microsoft Copilot or Copilot Chat licence and a SharePoint or OneDrive licence to create notebooks.
MCP connectorsA live link from the AI to another system, such as Notion or a practice management tool, instead of uploaded copies.Set by each connector and your admin settings.

One date to diary if you built on custom GPTs: Chad Davis flagged in his AI year in review that OpenAI is retiring them, and OpenAI’s help centre FAQ lists 11 December 2026 for most workspaces. If MCP is new to you, start with MCP for accountants, explained without jargon, and if you have chosen Claude, our Claude Cowork setup guide covers the environment.

Capture: feeding it every conversation

Documents are the easy part. Most client context lives in email, calls, meetings and texts, and that is where firms bleed knowledge. AI is now good at organising messy context, so those channels should flow into the second brain.

Start with the two richest sources. Meetings: whatever notetaker you use, whether that is the one in Microsoft Teams or a meeting assistant such as Vinyl, file the summary and action items in the client’s folder after every call. Email: have the AI summarise any thread where advice was given, and add it to the client log.

Keeping client data safe

Most of the summit’s security questions came back to setup, not tools. Five rules from the panels:

  • Business plans onlyAnthropic says it does not train on data from its commercial products, which include Claude Team and Enterprise, unless the customer opts into its Development Partner Program (Anthropic privacy centre). OpenAI says it does not train on ChatGPT Business or Enterprise data by default (OpenAI business data page). Personal plans run on consumer terms, where you need to check the training setting yourself.
  • One approved platformGus McLennan told the “Teach your team” panel to standardise on one common platform for all firm data; staff drifting to their own tool is where settings go wrong.
  • Redact before it goes inOn the same panel James Rose of Content Snare described having AI build a separate redaction tool that uses no AI itself, so identifiers are removed before anything reaches the model.
  • Least accessBeau Gaudron of Growthwise suggested automation logins that see only what they need.
  • Tell clientsUpdate your privacy policy and engagement letters, and say how you use AI. Andy Teece called these, plus staff training, absolutely critical.

a paid account does not mean they do not train.

Beau Gaudron, Tech Ninja at Growthwise, at the AI in Practice Summit, November 2026

In Australia, the Office of the Australian Information Commissioner’s guidance on commercially available AI products recommends, as best practice, not entering personal information, particularly sensitive information, into publicly available generative AI tools (OAIC guidance). Check the OAIC or your own regulator for your obligations; this is journalism, not legal advice. For the deeper version, read is AI safe for client data?, and write the rules down in your firm’s AI policy.

Keeping it current

A second brain rots fast, because the AI confidently repeats whatever is in it. Once a month, ask the AI to list documents not reviewed in 90 days and anything contradicted by something newer, then archive what is stale. When the team tries something new, ask the AI to document what did and did not work and file it in the lessons log. Review AI drafts until you trust them, as you would a junior’s work.

Start this week

  • Mondaychoose one AI on a business plan and one store. Write them into your AI policy.
  • Tuesdaywrite the firm rules document and your three most-used SOPs.
  • Wednesdaybuild profiles for your ten most active clients, checked against the “stays out” column.
  • Thursdayset up the project, notebook or connector, share it, and test it with five real questions.
  • Fridaystart capture. File the week’s meeting notes, give each folder an owner, and book the first monthly review.

Frequently asked questions

Is a second brain the same as a document management system?

No. Your document management system can be the store, but a second brain is curated for reading: short, dated summaries and rules, pointing back to the source files.

Do I need Claude, or will Copilot do?

Copilot Notebooks can read SharePoint folders and Office files, so a Microsoft 365 firm can build one without another vendor. At the summit I said Copilot is the natural fit for many firms, though Claude and ChatGPT were further along at the time.

Should client meeting recordings go into it?

Keep summaries, decisions and action items rather than full recordings: they are easier to search and to check for sensitive details. Tell clients, ideally in the engagement letter.

How big does it need to be before it is useful?

Small. A firm rules document, three SOPs and ten client profiles already change the answers. Accuracy matters more than coverage, so grow it as work comes in.

What happens to the second brain when a staff member leaves?

That is the point of building it. If their notes are filed as they go, the knowledge stays. Remove their access on their last day.

Sponsored

Canyou: job-ready workers, zero chasing. Employee onboarding synced to Xero, MYOB and Cloud Payroll.

AI in AccountingKnowledge managementData securityClaudeChatGPTMicrosoft CopilotAI in Practice Summit

Was this useful?
Thanks — noted.
Free template

Get the editable AI policy template

A complete, editable AI policy for accounting and bookkeeping firms: approved tools register, client data rules, review, client disclosure, incidents and sign-off. Free with a firm profile: enter your email and we will send a code. No password. Read it in full first.

Have a password? Sign in

By continuing you agree to our terms and privacy policy.

Or build it with AI

Open Claude or ChatGPT with the prompt filled in. It writes the finished document for your firm straight away, then asks a few questions to tailor it. No client data needed.

Open in ClaudeOpen in ChatGPT

Works without client data. Don't paste client names, tax file numbers or ledgers into an AI tool your firm hasn't approved for them. AI data safety playbook.

Sponsored

Vinyl: client meetings, minuted, with actions in your workflow. Start for free.
Read next
Related
Listen next