Infrastructure reviews are where the biggest mistakes get caught and the slowest work happens. Spacelift has been shipping an AI copilot that does a first pass on your Terraform and OpenTofu changes.
Spacelift started as a platform for running infrastructure as code pipelines. It handles the parts that other CI tools ignore, like policy checks, policy as code, and approval flows. The copilot layers on top of that and reads a plan before a human reviewer has to.
A second set of eyes on every plan
Every Terraform plan is a promise about what the tooling intends to do. The planner shows resources it wants to create, update, or destroy. The risk lives in the difference between what you asked for and what the plan says will happen.
The copilot reads that plan and compares it to the change you submitted. It flags the moments where the two disagree, or where an action looks risky. A broad security group change, a resource rename that will cause a replacement, or a deletion that looks like it can take something important with it. Those are the cases it checks first.
It also generates an explanation in plain English. Where a reviewer used to open a multi-hundred line plan and parse it line by line, the copilot summarizes what the plan does and where the attention should go.
Where the review flow gets faster
The real friction in infrastructure change is the reviewer loop. Developers push a pull request, then wait for a platform engineer to read a plan and sign off. That wait is where velocity dies, especially when the change is routine.
The copilot changes the shape of that loop. Automated checks filter out the trivial approvals, and the reviewer spends attention only on the changes that genuinely need a human. Teams I have seen use this to cut the backlog of waiting plans and keep the code moving.
It also raises the floor for less experienced engineers. Someone new to Terraform gets a plain language note about why their change would destroy a resource, which teaches them faster than a reject with no explanation.
The reality check
This is a copilot, not an autopilot. It can review a plan and flag a concern, but it cannot fully judge intent. A plan might look wrong to the model and still be exactly what the team wants. The final call belongs to a person who knows the business reason behind the change.
It is also a reviewer aid, not a replacement for policy as code. If you need a hard rule, like no changes that open the cluster to the internet, you enforce that with a policy engine and not with a suggestion. The copilot surfaces things. The policy stops them.
Naming accuracy matters here in an industry that trips over it. Spacelift runs plans for Terraform and OpenTofu, and the copilot reasons over both. If you use one and not the other, the behavior should match what you have configured rather than assuming.
Consistency is another quiet win. A rotation of reviewers will have different tolerances for risk. A copilot applies the same lens to every plan, which makes review outcomes less dependent on whoever happens to be on call.
What this points to is a shift in where the work happens. The most valuable application of AI to infrastructure is not generating config from a prompt. It is catching the mistake before a human has to find it. Spacelift is one of a handful of tools pursuing that angle, and the approach is more defensible than a chat window in your build. It answers a question you would have asked anyway, and it gets a head start on it.