CxO + VP Engineering briefing 01 / 14

Slide 01

Stop Reviewing Code. Start Proving It Works.

CxO + VP Engineering + Board
Core argument

Code review was supposed to be about rigor. It became a rubber stamp. AI review will not fix it. What fixes it is building systems that prove correctness.

Your teams spend $343K/year on a ceremony where 85% of the output is about style, naming, and social norms. Your production incidents come from the gaps the ceremony does not cover. The answer is not better reviewers. It is verification gates that prove the product works.

The shift From humans reading diffs to systems proving correctness. Continuous delivery, not ceremony.

Slide 02

"Code Review" Is a Phrase. Not a Practice. It Means Something Different on Every Team.

The SAD MF framing

Team A: Knowledge sharing

Senior walks junior through the change. Explains trade-offs. Everyone learns. Valuable. But that is mentoring, not quality assurance.

Team B: Risk reduction

Second pair of eyes before production. Also valuable. But only if the reviewer reads every line, traces logic paths, and checks edge cases. Almost nobody does.

Team C: Vanity metric

"We reviewed 100% of PRs this quarter." On a slide deck nobody in the room believes. Measuring comment volume without quality filters. A SQL injection flag and a semicolon suggestion count the same.

Ask ten engineering leaders what code review is and you will get ten answers. That is not a practice. That is a phrase everyone uses to describe something different.

SAD MF (Scaled Agile DevOps Maturity Framework) — sadmf.com

Slide 03

LGTM. Four Letters That Replaced Quality Assurance With Social Obligation.

What review actually is
What LGTM actually means

I am busy. I trust you. This PR has been open for three days and I feel guilty.

It means I do not understand this area of the code well enough to say anything useful, but I am not going to admit that. So I will approve it and hope the tests catch whatever I missed.

That is not rigor. That is a rain dance.

The numbers
Average review time 11 min

SmartBear data. Eleven minutes for a change that took eight hours to build.

Reviewer focus Style

Microsoft research: reviewers spend most time on style and formatting. Things a linter handles in milliseconds.

Slide 04

$343,000 a Year on Ceremony. Your Incidents Come From the Gaps It Does Not Cover.

The CFO slide
Daily reviews 80 PRs

Conservative estimate for a 200-person engineering org.

Daily cost $1,320

14.7 hours of senior engineer time per day at $90/hr loaded cost.

Annual cost $343K

Spent on a process where 85% of the output is about style, naming, and social norms. Work a linter does for free.

The incidents that reach production are the ones that pass review and pass tests. Semantic bugs. Integration failures. Edge cases nobody thought to check. The things eleven-minute reviews were never going to find.

You are spending $343K on ceremony and your incidents come from the gaps

Slide 05

Only 15% of Review Comments Are About Defects. You Are Calling This Quality Assurance.

The data
Style, naming, documentation, social norms85%
Actual defects15%
Formal inspection ~60%

Capers Jones: structured, multi-participant, documented code inspection catches ~60% of defects. That is good. That is not what your teams do.

PR review 5-15%

Industry estimates for pull request review done in eleven minutes between meetings. Your CI pipeline catches more. Your linter catches more.

What you track Whether

Most teams track whether reviews happen. Not whether they produce value. That is like tracking whether your team wore helmets, not whether they scored.

Slide 06

Army Sensors. Neonatal Dosing. When Getting It Wrong Means Someone Dies.

Why this matters personally

Chemical weapon detection

Sensor networks for the US Army. Chemical weapon leak detection on a base where the team sat. If the software missed a reading or threw a false negative, the people breathing that air were us. We did not LGTM that code. We verified it. We validated it. We ran it.

Neonatal medication dosing

Babies in the NICU whose bodies could not tolerate a rounding error. The difference between a therapeutic dose and a lethal dose for a two-pound infant is measured in micrograms. We did not skim that code in eleven minutes between meetings. We proved it was correct.

Slide 07

AI Review Is Better Than Your Team. It Is Still Reviewing Against Training Data.

The AI pitch and its limits
What AI does well

A frontier model will do the 45-minute review in 30 seconds. Consistently. On every PR. Without Friday fatigue.

It reads every line, traces logic, checks edge cases, flags security issues. It does what your best reviewer does, but on every single PR without exception.

Real value More thorough than your eleven-minute human reviewer. Every time. That is not hype.
What AI reviews against

Patterns from training data. Ten million examples of Express middleware. Not your billing race condition.

It does not know your state machine has an undocumented transition three customers depend on. It does not know the function was written to work around a vendor API bug from 2019.

The gap A very fast, very thorough generic review. Still misses what causes your production incidents.

Slide 08

The Billing Race Condition. The Undocumented State Transition. The Vendor Workaround From 2019.

What AI misses
Miss 1

System-specific failures

Your billing service has a race condition when two invoices close in the same millisecond. No training data covers that. No generic review catches it. Only someone who knows your system or a test that exercises that exact path.

Miss 2

Undocumented dependencies

Your state machine has a transition that three customers depend on. It is not in the docs. It is not in the tests. It is in one engineer's head. A reviewer, human or AI, cannot catch what is not documented.

Miss 3

Historical workarounds

The function was written to work around a vendor API bug from 2019. A reviewer sees "ugly code" and refactors it. The workaround disappears. The bug returns. You build the thing, then you check the thing. The defect already exists.

Pattern Your production incidents come from system-specific context that no reviewer, human or AI, can reliably catch by reading diffs. The answer is not better reviews. It is better verification.

Slide 09

Farley, Humble, Finster. The Answer Has Been Published Since 2010. Most Teams Still Have Not Implemented Half of It.

Continuous delivery
The principle (2010)

The goal is not to find defects after they are introduced. The goal is to shorten the feedback loop until defects are caught within minutes. Automatically. Every time.

Not by a reviewer. By the system. Deming said it about manufacturing fifty years before Farley and Humble translated it into software.

DORA data Teams deploying multiple times per day have lower change failure rates than teams deploying monthly. Not because they review more carefully. Because their systems verify more continuously.

Slide 10

Every Stage Is a Gate. Every Gate Has a Binary Answer. Pass or Fail. No "Looks Good to Me."

The pipeline model
01

Compile + Build

Every commit triggers a build. The artifact is identical to what runs in production. No manual packaging. No "it works on my machine."

02

Full test suite

Unit, integration, contract, end-to-end. Not "run the fast ones." All of them. Every time. The pipeline does not get tired on Fridays.

03

Security policy

Automated security scanning. Dependency checks. Compliance rules encoded as code. Not a checklist someone fills out quarterly.

04

Performance benchmarks

Critical-path latency. Memory usage. Throughput. Regressions caught before merge, not after customers complain.

05

Rollback verification

Confirm the deployment can roll back cleanly. If you cannot undo it, you should not ship it.

06

Production artifact

The artifact that passes all gates is the artifact that deploys. No rebuild. No re-package. What was tested is what ships.

Slide 11

The Agent Does Not Review Your Code. The Agent Builds the Gates. Then Runs Inside Them.

Domain-aware verification
Gate

Domain model validation

An AI gate that understands your domain model validates that a pricing change does not create negative-margin scenarios across your product catalog. Not pattern matching. System-aware verification.

Gate

Contract verification

An AI gate that has ingested your API contracts verifies that a schema change does not break downstream consumers in ways a static type checker cannot see.

Gate

Compliance verification

An AI gate that knows your compliance requirements flags a data retention change that would put you out of HIPAA compliance before it ever reaches a human screen.

One is informed by training data. The other is informed by your system's actual requirements. That is the difference between AI reviewing code and AI verifying the product is correct.

AI as pipeline participant, not as diff reader

Slide 12

You Do Not Stop Doing Code Review on Monday. You Start Building the Gates That Replace It.

The transition plan
Q1

Highest-churn files. Most common incidents.

Have your agents write the tests that cover those paths. Measure change failure rate before and after. You should see movement within 90 days.

Q2

Expand coverage. Add contract and performance gates.

Contract tests between services. Performance benchmarks for critical paths. Start tracking what your pipeline catches that your reviewers did not.

Q3

Shift the conversation. Review becomes mentoring.

When the pipeline catches more defects than reviewers, treat code review as knowledge sharing. Not quality gate. Reviews become about design, mentoring, shared understanding. The pipeline finds the bugs.

Regulated industries An automated gate that runs the same checks on every commit produces a traceable, reproducible audit trail. Stronger compliance posture than "Steve reviewed it on Tuesday and clicked approve."

Slide 13

What to Do Monday. Three Actions That Change the Trajectory.

Immediate actions
01

Change what you measure

Stop measuring whether code reviews happen. Start measuring whether your pipeline catches defects before production. If your metric is "PR approval rate," you are measuring ceremony.

02

Audit your last 20 incidents

For each one, ask two questions. Would our review process have caught this? Would our pipeline have caught this? If neither, that is a gap in your validation gates. Not a reason to add more reviewers.

03

Build three tests, not three reviewers

If your payment processing has three edge cases that caused incidents, the fix is not a more careful reviewer. The fix is three tests that make those edge cases impossible to ship.

Read Farley and Humble's Continuous Delivery. Read it again if you read it in 2010. I promise you did not implement half of it.

The data has been clear for years — deployment frequency correlates with lower failure rates

Slide 14

Get Rid of the Phrase "Code Review" Entirely. Describe What Actually Happens Instead.

Decision close
Stop saying

Code review keeps quality high. We need to review more carefully. We reviewed 100% of PRs this quarter.

Code review has value as knowledge sharing, as mentoring, as shared understanding. That value is real. But it was never quality assurance. It was a ritual performed because it felt rigorous.