Skip to main content
Learn about AI for accountingJoin live workshops

Buy What Writes, Build What Reports: A Rule for AI in Family Office Accounting

Nov 18, 20266 min readBy Truewind
Buy What Writes, Build What Reports: a rule for AI in family office accounting - Truewind illustration

Every family office now faces the same question from two directions. Vendors pitch AI for every workflow, and someone on your team has already built something useful in a general-purpose AI tool over a weekend. InfoGrate tracks more than 500 family office technology products, and 125 of them are AI native. Choosing among them has become a job of its own.

There is a sharper question underneath it: does this AI write to your books?

That question came out of a recent conversation between Dr. Tania Neild, Founder and CEO of InfoGrate, and Alex Lee, Co-founder and CEO of Truewind. Tania has spent her career building the systems family offices run on, from the National Security Agency to co-founding EnvestNet to building the aggregation, reconciliation and reporting platform at Private Client Resources. Her rule fits in one line.

Buy what writes. Build what reports.

If a tool writes to an external system, such as your general ledger, your ERP or Addepar, buy it. The cost of an error is too high to carry yourself, and a vendor carries the maintenance, the security and the updates.

If a tool summarizes, analyzes or reports, you can build it. The stakes are lower, and a mistake shows up in a draft instead of in the books.

Run your workflows through that rule and the answer gets clear quickly. Of ten common family office accounting workflows, seven write to the GL:

  • Custodial statements
  • Capital account statements
  • Bank reconciliation
  • Transaction coding
  • Investment activity
  • Property statements
  • Journal entries

Three report:

  • Variance review
  • Cash flow forecasting
  • Family reporting

Seven of ten are a buy. The three you can build are real wins, and they are a good place for your team to learn.

Why writing changes the math

A report with an error gets questioned in a meeting. An entry with an error flows into the ledger, the roll forward, next month's opening balance and the family's reporting. That difference is the whole argument, and it shows up in three places.

Repeatability. General-purpose AI is probabilistic. Ask it the same thing twice and you can get two different answers. For a summary, that's fine. For a reconciliation your controller has to sign every month, you need the same inputs to produce the same output. That takes fixed code for the arithmetic and AI only where judgment is needed, such as reading a statement in a format it hasn't seen.

Cost. Building something that writes safely costs more than most offices expect. By Tania's estimate, standing up a secure environment for one agent runs about $5,000, plus about $300 a month for the virtual desktop, firewall and usage. NIST-compliant code scanning tools add $20,000 to $40,000 a year. Each agent may need its own isolated setup, because agents shouldn't share access.

Security. An agent that writes can be steered. Tania has seen instructions hidden in a PDF as white text on a white page, code that came back embedded in search results and agents that changed their own browser settings. Her guardrail stack for any agent that writes:

  • An isolated virtual desktop, apart from anyone's own machine
  • A firewall that allows only the exact sites the agent needs
  • Browser restrictions as a second layer, never the only one
  • Credentials and API keys in a vault, never in a prompt or in code
  • Its own login, with limited permissions
  • A log of every action

If your office builds anyway, decide three things before anyone writes code: whether citizen developers or a deployment team owns the work, whether each step is deterministic or probabilistic, and whether it runs on third-party tools or your own cloud. Then write the policy: what people can build and deploy, where code is versioned, who reviews each change and who maintains it when a package goes out of date.

Where AI should not touch the books

Every office using AI in accounting needs a short list of things the AI never does on its own. Ours:

  • It never posts without a person. AI can prepare the entry and attach the workpaper. Someone on your team approves it before it reaches the GL.
  • It never does the math. Sums, tie-outs and roll forwards belong to fixed code that gives the same answer every month.
  • It never writes without guardrails. Anything that touches an external system needs every control above, or it should be a buy.

The approval step deserves design attention. A tool that asks for permission twenty times a run trains your reviewer to click yes. One clear approval, at the point where the entry is complete and the support is attached, keeps the review meaningful.

How to vet a vendor

Buying moves the risk to the vendor, so test it. Tania treats SOC 2 Type 2 as the floor. Of the roughly 125 AI vendors she has evaluated, more have failed her security review than passed, and she rechecks them every 6 to 12 months.

Ask every vendor:

  • What can the agent read, and what can it write?
  • Does it have its own login with limited permissions?
  • Is every action logged?
  • What happens if a statement contains instructions?
  • Does a person approve before anything posts?
  • Where does our data live, who can see it, and is it used to train models?

Then ask about depth. "Do you integrate with our ERP?" invites a yes. "Which modules can you write to?" tells you what the integration actually does.

For the record, here is how Truewind answers: Truewind is SOC 2 Type 2 certified, every entry is a draft with its workpaper attached, and nothing posts without human review.

What to do this quarter

Start with a list. Write down every accounting workflow your office runs and mark each one "writes" or "reports." Buy for the first column and let your team build in the second. Put the guardrail list in front of anyone building, and put the vendor questions in front of anyone buying.

Tania's advice for offices still weighing a build was to move now: get a vendor's guardrails working for you today, and revisit in two years if agents become cheaper to build safely.

If you want to see what buying looks like on your own statements, book a working session with Truewind and bring the statement that takes you the longest every month. To talk through your office's technology and operations, schedule a call with InfoGrate or connect with Tania Neild on LinkedIn.

Workpaper automation

Turn this into a close-ready workpaper

Start with sample files or upload your own statements to see how Truewind prepares review-ready workpapers and journal entries.