Every development firm you talk to in 2026 will tell you they use AI responsibly. The claim costs nothing to make and is nearly impossible to disprove from a slide.
There is one request that works: show me a pull request where a reviewer rejected generated code, and walk me through why.
Teams with a real pipeline find one in about ninety seconds, usually with mild enthusiasm, because rejecting generated code is a normal Tuesday for them. Teams performing the practice search for a while, then explain that their process is more nuanced than that.
Here is what you are actually auditing, and what a healthy answer contains.
The two failure modes at opposite ends
"We don't use AI — all our code is written by hand." In 2026 this is not a craft position. Manual boilerplate is not a quality signal; it is a billing model. What it usually means is that the firm's pricing depends on hours that no longer need to be spent, and nobody internally wants to be the person who says so.
"AI writes about 80% of our code." Stated with pride and no review model attached, this is a description of a liability. The problem is not the percentage. Generated code is frequently good. The problem is that no one has drawn the line where human accountability begins.
Both answers are about the same missing thing: an explicit boundary between what is drafted and what is owned.
What a healthy pipeline looks like
The firms handling this well describe something like the following, and can produce evidence for each step.
Generation is scoped, not universal. Models draft implementation against an architecture a human designed. They do not choose the data model, the auth strategy, the caching layer or the sync algorithm. Those decisions have consequences that outlive the sprint, and generation is optimized for plausibility rather than consequence.
Review is mandatory and human, with no exemption for generated code. In fact good teams review it harder, because the failure mode is different: generated code tends to be syntactically clean and semantically confident, which is precisely the combination that slips past a tired reviewer. A hand-written bug usually looks like a bug. A generated one looks like a decision.
Automated gates run before the human. Static analysis, dependency scanning, secret detection, test coverage thresholds, and a build that fails on new warnings. The point is to spend human review attention on logic and design rather than on things a linter can catch — and generated code produces a specific flavour of noise these gates catch cheaply.
Security-sensitive paths have named owners. Authentication, session handling, payment flows, cryptography, and anything touching personal data are reviewed by a specific person, not by whoever picked up the queue. Ask who that person is on your project. If there is no name, there is no accountability.
Tests are not generated by the same pass that generated the implementation. This is the check most teams miss. If the model writes the code and then writes the tests, the tests encode the implementation's assumptions — including its mistakes — and pass in perfect agreement with a bug. Tests should come from the specification, ideally written first, by a different process or person.
The questions that reveal the truth quickly
Beyond the pull request request, four follow-ups do most of the work:
-
What data leaves our environment when you use your tooling? Source code, customer data, internal documentation? Which providers? What retention terms? For a client under Quebec's Law 25 or in a regulated sector, this is not curiosity — it is a cross-border disclosure question with real obligations attached.
-
How do you handle dependency suggestions? Models suggest packages with confidence, including some that are unmaintained, unnecessary or occasionally nonexistent. A team with no dependency review policy has an attack surface growing on autopilot.
-
What's your policy on generated code you don't understand? The only defensible answer is that it does not merge. If a reviewer cannot explain what a block does and why it is there, someone will be trying to debug it at 2am with less context than exists right now.
-
Has velocity changed, and where did the time go? A truthful answer is specific: more iteration cycles inside the same budget, more time in discovery and instrumentation, more thorough review. An answer of "we're just faster" with no reallocation described usually means the saved hours became margin, not quality.
What this is really measuring
You are not measuring their AI tooling. Tooling converges quickly and everyone will have the same models within a quarter.
You are measuring whether the firm has an opinion about where machine output stops and engineering judgment starts — and whether that opinion is written down, enforced in CI, and visible in their commit history. That is a proxy for how they will behave on every other judgment call in your project, most of which you will never see.
The teams that thought hard about this are the same teams that will tell you a feature in your brief is a bad idea. That correlation is not accidental.
Broader context — Canadian cost bands, delivery models, privacy architecture and the full vetting checklist — is here: Mobile App Development Company Canada: The 2026 Buyer's Guide. How we structure delivery is on our mobile app development page.
Frequently Asked Questions
Is it a red flag if an agency uses AI to write code?
No — refusing to is the bigger flag in 2026. What matters is scope and review: models drafting implementation against a human-designed architecture, mandatory human review with no exemption for generated code, and named owners for security-sensitive paths.
Why ask for a pull request where generated code was rejected?
Because it is the one artifact that cannot be produced retroactively. A team with a real review culture has many; a team performing the practice has none, and the search itself is the answer.
Should tests be generated alongside the implementation?
Not by the same pass. Tests generated from the implementation encode its assumptions, including its bugs, and then pass in agreement with them. Derive tests from the specification, ideally written before the implementation.
What should I ask about data leaving my environment?
Which providers receive source code, customer data or internal documentation; what the retention terms are; and whether processing region can be controlled. Under Quebec's Law 25 or in regulated sectors this is a cross-border disclosure question with concrete obligations.
How does AI change how much review time is needed?
It increases it per line of accepted code. Generated code is syntactically clean and semantically confident, which is exactly what slips past a tired reviewer. Hand-written bugs usually look like bugs; generated ones look like decisions.
If implementation is faster, should my project be cheaper?
Either cheaper for implementation-heavy work or materially more iteration for the same budget. What you should not accept is unchanged pricing, unchanged cadence and unchanged review depth while output per hour has multiplied.