Your Accounting System Needs a Bouncer

There's a version of the AI-in-non-profits story being told right now that goes like this: AI will automate the tedious back-office work, free up your staff, and let you focus on your mission. That version is true. I believe it deeply — I'm building the tools to make it real.

But there's a part of the story nobody's telling, and the gap between those two versions is where a lot of organisations are going to get hurt.

The AI apocalypse you've been reading about — model hallucinations, data breaches, AI systems going rogue — those are real concerns. But they are not what quietly destroys a non-profit's financial integrity over the course of a year.

Here's the one that will: an agent that posts a $50,000 restricted grant as unrestricted revenue because nobody built the accounting rules into it.

I've been doing non-profit finance for 20 years. I've audited organisations. I've cleaned up messes that took years to create and quarters to unravel. And I can tell you with certainty: the most dangerous financial error is the quiet one.

The quiet errors that compound into audit findings and funder disputes:

• The wrong transaction type — a donation posted as a journal entry instead of a sales receipt
• The miscoded fund — restricted dollars sitting in an unrestricted bucket
• The misreported grant — supposed to be fully spent on programs, ends up allocated across the wrong lines

These are internal control problems. They become AI problems the moment you give an agent the ability to take actions without verification.

I want to be careful here. I have enormous respect for engineers and computer scientists. They build remarkable things. But there's a reason the financial systems that govern banks, public companies, and governments were not designed by engineers alone.

Internal controls exist because humans make errors. Systems make errors. And the only way to catch those errors before they compound is to build a verification layer between the action and the ledger.

In accounting, we call this:

• Segregation of duties — the person who initiates a transaction isn't the person who approves it
• Approval workflows — a threshold triggers human review
• Three-way matching — invoice, purchase order, and receipt must agree before payment
• Reconciliation — independent verification that two records of the same event match

These are the reason your auditors sign off. They're the reason your funders trust you with restricted dollars. They're the reason your board can sleep at night.

When I started building TERN, I made one foundational decision that everything else flows from: an agent should not be able to post a transaction, move a record, or trigger a financial workflow without first passing through a verification layer. No system should be trusted without controls.

What I'm building — and what I'm calling guardrails — is a compliance and verification layer that sits between your AI tools and your financial systems.

Before a donation posts to your accounting system, TERN checks:

Is this the correct transaction type? A donation from Stripe should almost always post as a sales receipt. This sounds obvious. But I've seen this exact error in almost every pre-built connector I've ever evaluated. It's common, it's hard to catch retroactively, and it misrepresents your revenue composition.

Does the fund coding match the restriction on this gift? A $10,000 grant restricted to program delivery carries obligations an unrestricted $10,000 donation does not, even at the same dollar amount. Posting them the same way is a governance failure.

Are there approval steps required before this posts? Some organisations want a human to review before anything over a certain threshold hits the ledger. That's good governance.

And at any point in a TERN workflow, I can inject a human approval checkpoint. Your team gets a notification in Microsoft Teams, Google Chat, or Slack — the tools you already use. One click to approve, one click to reject, with full context. The workflow pauses. It doesn't proceed until a human says so.

This is not how most automation tools work. Most tools are designed to move data from A to B as quickly as possible, with as little friction as possible. That's the right design for a lot of workflows. For the workflows that touch your financial records, it's the wrong one.

A strategic partner of ours processes over 50 donations per month from multiple fundraising platforms. Before we automated their gift processing, this took about 24 hours of staff time per month — manual reconciliation, cross-referencing three systems, hoping nothing was miscoded.

A low-code predecessor brought that down to about 4 hours. But it was brittle. It broke when platforms updated their formats. It had no verification logic. And it posted everything as journal entries.

TERN will bring that to near-zero, because we're verified. Every donation flows from the processor through the CRM into the accounting system as a correctly coded, fund-restricted receipt. If there's any ambiguity, the workflow pauses and asks a human. The audit trail is complete from donor to ledger.

That 83% reduction in processing time I mentioned? That was the unverified low-code version. TERN gets you to near-zero hours of confident, auditable automation.

One more thing worth knowing: you don't need to buy new software to use TERN.

If your organisation already has an AI subscription, you can licence TERN's connectors and run them through the AI you already pay for:

• Claude Pro
• Microsoft 365 Copilot
• OpenAI for Business
• Any AI platform that supports connecting external tools

Your AI, our guardrails.

If you don't have an AI plan, or you'd rather not think about it, we deploy a custom agent in your existing workspace. Your team interacts with TERN through conversations in the tools you already use:

• Microsoft Teams
• Google Chat
• Slack

No new software. No new logins. The guardrails are the same either way.

There are roughly 170,000 registered charities in Canada. Hundreds of thousands more non-profits in the US. Most of them are currently evaluating AI tools, being pitched AI tools, or feeling the pressure from their boards to do something with AI.

A meaningful percentage of them will implement automation without verification. The automations will work — in the sense that data will move from one system to another. And then, quietly, the errors will accumulate. The restricted funds will drift. The funder reports will misrepresent program spend. The auditors will find something.

And then the sector will do what sectors do when trust is broken: it will pull back. The EDs who got burned will warn their peers. The funders who saw misreporting will tighten their requirements. The technology that could have transformed the sector's administrative capacity will be associated with the organisations that moved too fast.

I think that would be a genuine loss — not just for the sector, but for everyone who benefits from the work these organisations do.

I'm building guardrails because this specific domain requires that controls come first. A computer scientist optimising for speed and capability is doing their job correctly — and in most workflow contexts, that's exactly the right priority. But in the workflows that touch restricted funds, fund coding, and funder accountability, speed without verification is how errors compound quietly over years. That's what twenty years of non-profit finance work taught me, and it's baked into every workflow TERN runs.

TERN is in active development with a small set of Strategic Partners who are co-creating the platform. If your organisation has real operational complexity and you want to be part of building this rather than just adopting it, reach out. Space is limited and intentional.

If you're not ready for that level of engagement, the Mission Multiplier Program is the place to build the skills and frameworks that will make you ready.

And if you're thinking through what responsible AI adoption looks like for your organisation — that's exactly what I'm here for. Start there.