The most common complaint I hear from landlords who have tried AI bookkeeping tools: "It makes the same mistake every month."
They correct it. Next month, the same document comes in. Same wrong category. Same wrong property. Same wrong transaction type. And they correct it again.
That is not AI getting smarter. That is a form with extra steps.
When I built DoorVault, I knew that was not good enough. If Knox was going to handle the entire back office of a rental portfolio, it had to actually learn. Not just process documents. Learn your preferences. Remember your corrections. Get better over time so that the manual work shrinks week by week, not just shifts from one place to another.
That is what the Knox Learning Loop does.
What the Knox Learning Loop Is
Every time Knox proposes a change and you correct it, Knox does not just apply your fix. It stores the correction pattern for your account. The next time the same type of document comes in from the same source, Knox applies your correction automatically.
Your corrections are scoped entirely to your account. They never affect anyone else's setup.
A simple example: your property manager sends a monthly statement that includes a line item called "Owner Draw." Knox initially categorizes it as an Expense. You correct it to PM Disbursement. That correction is logged in the Knox Activity Log with a before and after snapshot at the field level. The next month, when the same PM statement arrives with the same line item, Knox already knows. Owner Draw maps to PM Disbursement. No correction needed.
One correction. Done.
Why This Matters More Than It Sounds
Most landlords have at least a few recurring quirks in their data. A PM that uses unusual shorthand for maintenance charges. A bank transaction description that comes through abbreviated. A recurring vendor that spans two expense categories depending on the job type.
In a traditional tool, you correct these every single month. That is the hidden tax of automation that does not actually adapt.
In DoorVault, you correct them once.
The effect compounds over time. In week 1, Knox gets 8 out of 10 transactions right. You correct 2. By week 4, those 2 patterns are already handled. Knox gets 10 out of 10. In week 8, a new vendor appears. You correct it once. By week 12, you have not touched a categorization in three weeks.
The manual overhead does not stay flat. It shrinks.
That is the practical difference between AI that processes and AI that learns.
How to Trigger a Correction
There are two places where corrections feed the Knox Learning Loop.
The Knox Activity Log. The Activity Log shows every change Knox has made, with a full before and after snapshot at the field level. If a transaction landed in the wrong category, you click the field, update it, and save. Knox logs your correction and associates it with the pattern that triggered the original decision. Single field revert is available for every entry, so if you change your mind, a rollback takes one click.
The Batch Review Page. When you upload multiple documents at once with Trust Knox set to OFF, every proposed change lands on the Batch Review page instead of being applied automatically. You can review each proposed transaction, approve what Knox got right, and correct what it did not. Corrections made on the Batch Review page feed the Learning Loop exactly the same way as corrections made in the Activity Log.
The Batch Review page is persistent across sessions. Upload 30 PM statements on a Sunday night, close the tab, come back Monday morning, and all your pending review work is still there. Approve per document, per field, or apply everything at once when you are ready.
Both paths build the same pattern library. One just happens during live review. The other happens after the fact.
The Difference Between General AI and Purpose-Built Bookkeeping AI
A lot of AI tools in real estate right now are general purpose language models bolted onto a real estate interface. They process documents in isolation, with no memory of your portfolio, no context about your preferences, and no ability to learn from your corrections over time.
Knox is different because it was built specifically to run a rental portfolio.
Every document Knox processes is connected to a property, an entity, a property manager, and a full transaction history. Knox does not evaluate a line item in a vacuum. It evaluates it in the context of your entire account.
That is why the Learning Loop works. When you correct Owner Draw to PM Disbursement, Knox is not just filing away a generic label substitution. It is recording that for this PM, in this LLC, for this type of line item, this is how you want it handled. The correction is context-aware.
That level of specificity is what makes the correction durable. It does not just apply to the next instance of Owner Draw everywhere. It applies to the next instance of Owner Draw in the right context, for the right property, under the right entity.
What This Looks Like Across 90 Days
Here is a realistic picture of the Learning Loop in action across a 10-property portfolio.
Month 1: Knox processes your PM statements, insurance renewals, tax bills, and bank transactions. Initial accuracy is high on standard document types. You correct a handful of context-specific categorizations: your PM's shorthand, your vendor naming conventions, your entity structure quirks.
Month 2: The corrections from Month 1 are already applied. Knox handles the recurring patterns automatically. New document types may surface a correction or two. Your review time has dropped noticeably.
Month 3: Knox knows your account. Review is down to genuine outliers and truly new items. Monthly admin is 15 minutes instead of 6 hours.
This is not theoretical. It is the architecture Knox Intelligence was built around. The whole point is AI that proposes, learns, and never touches your data without permission. The learning piece is what makes the other two pieces worth having over the long run.
This Is Still Your Data
A concern I hear from landlords: "If the AI is learning and applying corrections automatically, how do I stay in control?"
The answer is the Trust Knox toggle.
With Trust Knox set to OFF, Knox proposes every change before applying anything. You see the proposed transaction, the proposed category, the proposed property assignment. You approve or reject each one. Corrections you make during review feed the Learning Loop. Nothing changes in your books without your explicit sign-off.
With Trust Knox set to ON, Knox applies changes automatically. The Activity Log still captures every action with before and after snapshots, and single field revert is always available. You choose the speed versus control tradeoff that fits your comfort level, and you can change it at any time.
The Learning Loop works the same way in both modes. In OFF mode, corrections happen at review. In ON mode, corrections happen in the Activity Log after the fact. Either way, Knox gets better.
The control is not an add-on. It is built into how Knox processes everything.
Getting Started
Knox starts learning from your first correction. There is no training period, no setup configuration, no schema mapping required. Just process your first batch of documents, correct what Knox gets wrong, and the pattern library starts building automatically.
The Starter plan is free for up to 2 properties. Knox Intelligence, the Learning Loop, the Activity Log, and the Batch Review page are all included. You can see exactly how the system learns before you commit to anything.
Start free at https://doorvault.app