Case study · Proof of concept

CAM reconciliation: AI reads, code does the math

We built a proof of concept for CAM reconciliation around one rule: the AI reads and explains, plain code does the math, and a person approves. It runs on a fictional shopping center with four tenants and 105 ledger rows, and the agent’s answers are checked against the leases and the code before they’re shown. It has not yet been tested on real leases.

Last updated

Built Nov 6, 2026Fictional shopping center

Demo. This isn’t accounting, legal or tax advice, and it comes with no warranty. Results depend entirely on the lease terms and figures entered, so check every statement against the leases and your records before you send it.

Every year, a commercial landlord compares what each tenant paid in estimated common area charges with what the lease says they owe. The arithmetic is simple. The hard part is that every lease has its own rules, and amendments quietly change them. A controller told us the CAM setup in their property software had never been done, because daily work never left time for it, so the reconciliation happens in Excel. We built this to see whether AI could carry the reading without being trusted with the numbers.

Open the workbench

4 tenantsin one fictional center, each with different lease rules
105 ledger rowsfrom two export layouts, in one set of cost pools
9 review items4 lease rules and 5 ledger rows that need a person
0 figures from AIevery amount comes from tested code
The design

AI reads, code calculates, a person approves

A language model is good at finding the clause that matters and bad at being the only record of a number. So the work is split. The model turns leases into rules and points at ledger rows that look wrong. Ordinary code applies those rules and produces each figure, so the same inputs always give the same answer. The controller reviews only what was flagged, and each amount on a statement opens its ledger rows and its lease clause.

The sample

Six traps we planted

An amendment that replaces a clause
The gift shop’s cap went from 5% over last year’s bill to a compounding cap from a 2022 base. Reading the original lease alone gives the wrong bill.
An amendment that isn’t in force yet
The restaurant’s fee drops to 10%, in an amendment signed in 2025 that starts in 2026. So 2025 stays at 15%.
A change partway through the year
The restaurant expanded on July 1, so 2025 uses 4,000 square feet for six months and 5,000 for six.
A lease that says nothing
The gym’s lease never mentions gross-up, though the others allow it. Silence isn’t permission, so a person decides.
A capital item coded as repairs
A roof replacement sat in repairs. The leases exclude capital replacements.
The same invoice in two systems
An invoice paid in June appears again in July, right after the center moved systems.
Guardrails

What stops the AI being wrong quietly

A quote must exist
When the agent cites the lease, our code looks for those words in the lease. If they aren’t there, the page says so.
A figure must come from code
If an answer holds an amount no tool returned, the answer is stopped and not shown.
What-if runs in code
To ask what a change would do, the agent calls a tool that recalculates. It can’t work out a difference itself.
The tools only read
The agent can look things up and recalculate a copy. It has no tool that changes a decision or approves a statement.
Only a person approves
The approve button stays off until every flagged item has been decided.

The checks are tested with a scripted model that deliberately misbehaves: invented quotes, made-up amounts, unknown tools, endless loops. The expected figures in the math tests were recomputed separately with exact fractions.

First live run

What the model did, Nov 6, 2026

We ran the sample once through DeepSeek’s deepseek-flash model. Its reading matched the rule sheet on all 16 lease rules, including the 2026 fee change. It quoted the lease correctly on 15, and on the silent gym lease it quoted nothing and asked for a person. In the ledger it found all 5 planted rows.

It wasn’t perfect, and that’s why a person approves. It flagged both copies of the duplicate invoice for exclusion, which would have removed a real cost, and it would have excluded the storm cleanup that the sample keeps. One run on invented leases shows the approach can work. It isn’t a measure of accuracy.

Limits

What this doesn’t prove yet

  • The leases, the ledger and the column layouts are made up for this example. Real leases are longer, messier and sometimes scanned.
  • The rule sheet covers the terms in these four leases. A real one would need more kinds of rule.
  • This example sends its fictional leases and ledger to an AI provider through our server. Real rents and leases would need their own setup, with an AI provider and data terms the client agrees to first.

Have leases and a ledger of your own to try this on? Tell us about them.