There's a request on Xero's product ideas forum asking for the ability to import and export bank rules. It has 540 votes. Practitioners have been asking for it since 2017. In November 2025 Xero said it wasn't on the roadmap, preferring to improve automatic matching so fewer rules would be needed, and in January this year it moved to "Accepted" while the product team looks at how it might fit. It still hasn't shipped.
QuickBooks Online has done this for years. Open Rules, choose Export rules, save the file, then open the other company and import it.
Same job, two answers. And bank rules are instructions you wrote: your judgment about how this client's transactions should be treated, written down so a machine repeats it. On one platform that judgment travels. On the other it stays where you typed it.
That was a nuisance when a rule took thirty seconds to rebuild. It stops being a nuisance now, because the instructions your firm is writing this year carry a great deal more of your firm inside them.
Part 1 of this series argued that the build is where your fee belongs. Part 2 argued that the rescue should cost more than the original bookkeeping did. Both told you to charge full value for encoding a client's system. This one is about where you put that encoding, because that choice decides whether it compounds for your firm or disappears.
Two ways your firm loses the same AI asset
Back in June, when Ramp launched Stack and called it an AI operating system built specifically for accounting firms, I told you to ask one question before you teach any platform your methodology: can you export your skills if you leave? Six weeks on, not one vendor has published an answer. So here's what to do with the answer when you finally get it.
Start with the fact that you can lose the same asset through two different doors.
The first door is the client's. XeroForce, announced in May and now in early access with general availability promised this year, lets you build agents in plain language inside Xero. Agents are AI that carries out a sequence of work on its own rather than just answering a question. Xero's own framing is that it becomes the orchestration hub and core financial operating system, and the agents can be shaped for your firm, one client, or an industry. Build them in the client's organisation and the client leaves with them, along with every rule you wrote.
The second door is the vendor's. Bookkeeping automation now arrives pre-built: Pilot's Meridian, Artifact, Wurthy, Ramp Stack. File quality review has Double, which most of you knew as Keeper until it rebranded last October. Teach any of these how your firm works, then decide to switch, and your encoded judgment stays behind in a product you no longer pay for.
Same asset. Opposite exits. One test, in both directions: can you take the instructions with you?
The instruction is the asset, not the output
This is worth being precise about, because the valuable thing is easy to mistake.
The output is the categorized transaction, the reconciled account, the drafted workpaper. It's disposable. Next month there's another one.
The instruction is the rule that decided it. The bank rule. The review points you always hit on a particular return type. The five questions you ask before you'll sign a year-end file. The exception list you run before a client's quarterly meeting.
That's the part which took you fifteen years to develop, and it's the part that now goes into a system so a machine can repeat it at scale.
Which means the asset your firm is building this year is a body of written instructions. Treat it like one. You wouldn't accept a working paper system you couldn't export.
Three answers a vendor can give, and only one gives you AI portability
Ask the question and you'll get one of three answers.
You can't see it. The tool does the work and produces a result, and the routine behind it isn't visible to you. Double falls here as far as I can tell, and it publishes nothing on whether the scripts driving its checks can be read or extracted. This is not a criticism of the output, which is good. It's a statement about what you're accumulating: usage, not an asset.
You can see it, it's contractually yours, and the only way out is by hand. Ramp Stack is the clearest case, and give it credit, because it does more here than most. It captures your processes as skills, which are editable documents describing how your firm does the work, and Ramp contractually treats those encoded processes as your firm's intellectual property. But there's no external API and no open standard, so getting those skills out means copying and pasting them one at a time. Nothing is technically blocked. The exit is just priced in effort, and that price scales with your client list.
You can see it, own it, and move it. This is the open-rails posture. Karbon, Firm360, and Kick opened MCP servers in June, and Filed, Symmetry, and TaxAct added them this month. MCP is a shared standard that lets the AI you choose plug into a vendor's data, like a universal socket instead of a custom cable for every pairing. The instructions sit with you and reach into their system, rather than living inside it. Monday's roundup had the simplest version of this: six minutes of recorded screen work produced an editable procedure file, and a file on your own drive is the one thing in this whole argument that nobody can take away.
One distinction to hold on to, because vendors will blur it. Wurthy promises that its agents explain their reasoning before acting and that you set the automation boundaries. That's real and it's more than most offer. It's also visibility, not portability. Watching an agent think is not the same as lifting the boundary you set and installing it somewhere else.
Recreatable is not the same as portable
Here's the honest counterweight, and it cuts against the panic.
Almost none of this is genuinely proprietary. Every routine those tools run could be recreated by a competent practitioner with access to the right data. Nothing is locked in a vault. So the switching cost isn't capability.
The switching cost is capability multiplied by your client count. You can rebuild one client's rule set in an afternoon. You cannot rebuild it two hundred times, which is exactly the number nobody puts in the business case when the trial starts. That multiplication is the real cost of an answer like "you can copy and paste it," and it's calculable today, before you commit.
Run the vendor extraction test before you teach it, not after
None of this is an argument against buying tools. The pre-built agents are the reason these products are worth money, and a firm that insists on building everything itself will spend the next two years on plumbing while its competitors close books. Buy them.
Just price the exit at the moment of entry. Four questions, on the demo call, before you teach the thing anything:
Can I read the instruction the tool is following, or only see its result? Can I export it as a file I keep, and does that include the routines you built as well as the ones I typed in? If the only way out is copying it by hand, how long does one client take? And what is that number multiplied by the clients I'd be moving?
A vendor who answers all four cleanly has earned your methodology. A vendor who deflects the second one has told you what you needed to know. And if you're building inside a client's own platform rather than a tool you buy, ask the same four questions about your engagement letter instead, because that's the door the client walks out of.
Run-time keeps getting cheaper. The design and the accountability underneath it don't. But a design you can't carry out of the building isn't an asset your firm owns, it's a rental you happen to have furnished. So pick the client whose system you've built out the most, and ask yourself the only question that matters: if you had to move them next month, what would come with you?
If the answer is "not much," the gap between your platforms is where your own portable instructions should live. Download the free AI in the Gaps Toolkit at theaiaccountant.ai/toolkit for 100 scored workflows that sit outside anyone's walls.