.png){kind=logo}
Tellus App, a real estate and financial technology company, uses AI to draft code and review proposed changes. Its AI governance policy puts the division of labor in six words: "Humans provide taste. AI provides scale." People answer for the finished work: a qualified reviewer must understand and approve a change before it reaches customers.
"People define the problem, review the work and decide when it's ready for customers; AI does more of the building," says Jeromee (JJ) Johnson, President of Tellus App, Inc. and former head of BATS Options. "That lets a small team move quickly, and it makes ownership more important, not less."
Tellus's policy identifies a risk of AI-assisted development: "comprehension debt," when a team merges code it doesn't fully understand.
"The AI wrote it and the tests pass" is not enough to approve a change. The reviewer must be able to explain what the code does and why. If the developer who submitted it can't explain it, it isn't ready for review.
The policy also calls for engineering leaders to spot-check recently merged AI-assisted work to confirm that people understand what they approved.
Before AI begins building a change, a person defines the problem, the design, the limits and how success will be judged. AI-assisted code changes then move through three stages:
The build. AI drafts the change within those limits, and the developer submits it for review.
An AI review. A separate AI review checks the change for correctness, security patterns and test coverage. Its verdict never stands in for a person's.
A human approval. A code owner for that part of the system reviews the change, the AI's findings and the test evidence, then approves it or sends it back. For code headed to production, GitHub's branch protection requires that approval before a change can merge, and no one can approve their own change.
The person who directs the AI remains accountable for what it produces. Anyone in the chain can stop or override AI work at any time. The policy is blunt about it: "The AI did it" is not a defense.
Tellus uses synthetic test data in place of real customer records. In July 2026, it added another safeguard: a check for personal data and secrets in the screenshots, logs and videos its AI build system produces during testing. The check runs before that evidence is attached to a change for internal review.
A quality engineer reviewing the first version found gaps in what it could detect, and the team added six detectors within days. The check began in advisory mode, flagging matches while the team evaluated its results. Under the team's rules, quality engineering decides when the evidence supports turning enforcement on. The check now blocks flagged test evidence from being attached for internal review.
Human review improved the automated check before it began enforcing the rule.
For a company whose customers save and collect rent through its app, money and customer information get the strictest rules. Before AI-assisted code or calculations can affect a customer's balance, interest, rewards or a lending decision, a qualified person must independently verify the underlying math and logic. AI may help draft or check that logic, but its own assessment does not satisfy the verification requirement.
The policy prohibits sending the most sensitive customer information, such as financial records, to outside AI providers under any contract. It permits AI processing of that information only with models and hardware Tellus controls, with access restricted and use logged. Using a personal or free AI account for Tellus work is banned outright. These AI rules sit within Tellus's broader information security policies for data classification, access and logging.
AI-drafted customer messages get a person's review too. A support reply, notification or in-app text drafted by AI is approved by a person before a customer sees it, while automated messages such as confirmations use templates people approved in advance. Tellus's security officer can suspend any AI tool immediately if it poses a risk.
Johnson's approach draws on his work in financial markets. He built the BATS Options Exchange and led it through the May 2010 Flash Crash. As head of BATS Options, he later led an industry effort to harmonize the exchanges' rules for handling erroneous options trades. The SEC approved BATS's rule changes in March 2015.
"Automation is powerful, and its failures are fast," he says. "The higher the stakes, the stronger the check."
That is also why Tellus keeps human approval in the workflow as AI speeds up the work that comes before it. As Johnson wrote in his Forbes Technology Council article, "Protective friction is the control layer between model output and real-world action."
Tellus's approach offers six habits other teams building with AI can apply:
Define what requires human approval. At Tellus, that includes AI-generated financial logic or calculations, production code changes, compliance decisions and AI-drafted customer communications.
Set the design and boundaries before AI starts building. People define the problem, the scope and how success will be judged.
Keep AI review and human approval distinct. An automated assessment informs the decision; a qualified person makes it.
Approve only what you can explain. Passing tests does not replace understanding the work.
Set data rules by sensitivity. Tellus keeps its most sensitive customer information away from outside AI providers.
Check the evidence as well as the code. Screenshots, logs and videos need safeguards for personal data and secrets, even when tests use synthetic records.
AI helps Tellus produce and review more work. People remain responsible for understanding what it does and deciding when it is ready for customers.
Tellus is not a bank. Tellus is not FDIC-insured.