Why I built our AI systems in-house instead of buying them

Twelve months ago, I thought AI was a toy. I asked ChatGPT a few questions in late 2024, felt clever when it hallucinated on something simple, and moved on. Then a mentor pushed me to lean into it over the Christmas break. Today, my Brisbane firm runs dozens of AI systems in production. Somewhere in the middle of that curve, I had a large proposal in front of me from a vendor offering to build all of this for us. I said no, and I want to explain why. I ran through the full twelve months at The Firm's AI in Practice Summit, and this is the version of the argument I want practice owners to sit with.

The proposal looked good on paper. Someone else does the building, we get the AI systems, we get on with running the firm. What stopped me was the IP question. At the end of the deal, I would not have owned any of it. If I cancelled six months in, I walk away with nothing. All the context, all the tuning, all the work sits in someone else's platform. I would be stuck.

I decided to learn it and build it myself instead. Twelve months later, we own everything. The context we built into our systems is portable. If Claude falls behind and Gemini pulls ahead, I can move platforms and take our context with me. That independence matters more than the twelve months of learning cost.

Context is the whole game

The single biggest lesson from the last twelve months is that context is everything.

The AI is only as useful as what it knows about your firm. I ran a live comparison at the summit. Two Claude windows side by side. On the left, a blank Claude that barely knows I am an accountant. On the right, our master prompt project, which knows 338,000 words about Inspire, our clients, our processes and our people. The outputs look similar on the surface. The magic is in the small things. The right-hand side knows ninety-seven percent of our clients run on Xero. It knows we do tax planning in the lead-up to 30 June. It knows our review process, our team, our house style.

Building that context took work. The more you already have documented, the easier it is. And once the context exists, it compounds. Every new agent we build inherits what the last one learned. Every process we teach the AI becomes something the next process can reference.

That is the asset. Not the AI itself. The context we have built into the AI.

Turn a meeting recording into an SOP in under an hour

The first working system I want practice owners to see is our SOP generator. We use Vinyl to record team meetings, then feed the transcript into Claude with a prompt asking it to turn the meeting into a standard operating procedure. Because Claude already knows how our QMS is written, how we do document control, and the structure our SOPs follow, the output lands close to publishable on the first pass.

A well-written SOP in a mid-sized firm can take five, ten or fifteen hours. Our version comes out in under an hour, ready for revision. Most SOPs I generate go through two or three revision passes, which is normal. The mistake people make is expecting the AI to nail it first time and giving up when it does not.

Use an AI review agent to catch errors before the reviewer

The second working system is our year-end review agent. It takes financial statements and tax returns for a family group and runs a full review. It suggests journals to fix errors it finds in the Xero file. It flags issues like missed Division 7A interest and gives the specific journal to raise. It works out where a preparer's numbers do not tie back to the workpapers.

The point of this is not to replace the reviewer. It is to catch the errors before the reviewer sees them. A team member submitting a job to review can run it through the agent first and lift the quality before it ever reaches a manager. On the reviewer's side, the back-and-forth on review points collapses. What used to be a twenty-minute exchange, where the preparer has to get their head back into the job, becomes a direct question the reviewer can ask the agent while the file is open.

When we ran the agent across hundreds of year-end jobs and looked only at the review-point patterns, we could see exactly where our team was strong and weak. Systemic errors surfaced. That pointed us straight to focused training. Division 7A came up repeatedly, so we ran training on Division 7A.

Build one afternoon of AML compliance instead of paying a vendor

The third working system was the AML compliance build. Software vendors were quoting us thousands of dollars a year for AML tracking, ahead of the 1 July rules. In one Monday afternoon with Claude, I had it research the requirements, produce our first run of customised policies, and set up a workflow that takes an onboarding transcript from Vinyl and generates the CDD documentation we need.

I double-checked everything Claude produced, obviously. What Claude did was compress the work of a compliance consultant into an afternoon and give me policies tailored to our firm rather than generic templates. We have not needed to pay a vendor for it since.

Give the team a framework, not just access

Getting AI working for you is the first job. Rolling it out to the team is a different job, and it is harder.

The rule I now enforce is that nobody brings me a question without doing the thinking first. Take the question to the AI. Come back to me with three proposed answers and the one you think we should go with. I add the context you did not have, but you have already done ninety percent of the work.

That single rule cut the volume of questions coming into me every day by a significant margin. Team members do their own thinking, learn faster, and only escalate when they genuinely need the twenty years of context they cannot get from an AI.

The starting rules

If you are looking at your own twelve months ahead, five rules will save you time.

One, pick a problem worth solving. Spending five hours to save five minutes a month is not a project. Pick something you already know is costing your firm real time or real errors.

Two, run a five to ten day sprint on it. Expect dozens of revisions. Our tax-planning agent took forty-plus. Push through.

Three, keep a human in the loop. AI does not touch client-facing work without a person signing off. Ever.

Four, fix the process before you automate it. Automating a broken process just gives you a broken process running faster.

Five, own the IP. Wherever you build, make sure you can walk out with your context intact. That is the asset. Do not build it inside someone else's platform if you can help it.

You can watch the full session here.