AI is a power tool, not a teammate. Power tools are fantastic when you respect their edges and dangerous when you pretend they do not have any. For platform work, the line is surprisingly clear.
Where it genuinely helps
Toil: log triage, YAML boilerplate, one-off scripts, regex nobody wants to write
Scaffolding: a first-draft Terraform module or Helm chart to react to
Documentation: turning a working config into a readable runbook
Incident support: summarizing a noisy thread into a timeline
Glue code: adapters, migrations, and format conversions
Use it to generate a starting point you then own and refine, for example, scaffolding a module, then reviewing every line:
// AI-drafted, human-reviewed: an S3 bucket with sane defaults.
resource "aws_s3_bucket" "artifacts" {
bucket = "acme-build-artifacts"
}
resource "aws_s3_bucket_public_access_block" "artifacts" {
bucket = aws_s3_bucket.artifacts.id
block_public_acls = true
block_public_policy = true
ignore_public_acls = true
restrict_public_buckets = true
}Where it hurts
Confident wrong answers on APIs that changed after training
Security posture: it will happily suggest an over-permissive IAM policy
Novel design: it averages the internet, and your hard problems are not average
Anything where a plausible-but-wrong answer costs you a production incident
The failure mode is not gibberish, it is fluent, reasonable output that is subtly wrong. That is exactly the kind of mistake humans wave through.
A workflow that keeps you honest
Draft with AI, but treat the output as a proposal, not a decision
Diff and test everything before it reaches a cluster
Never let generated IaC apply without a human-read plan
Keep a paved road so the easy path is also the reviewed path
Automate the toil, own the judgment. AI is leverage on the parts you understand, not a substitute for understanding them.
The platform engineers who win with AI are not the ones who trust it most, they are the ones who verify fastest. Speed plus scrutiny is the whole game.
