What AI-Augmented DevOps Actually Means

Every consultancy has added AI to its website in the last two years. Almost none of them will tell you which parts of the work it changed, because for most of them the answer is "the marketing copy".

This is a more specific account: where AI agents do speed up infrastructure work, where they are a liability, and how you can tell which you are being sold.

The work splits cleanly in two

Infrastructure work divides into decisions and mechanics, and the split is sharper than people expect.

Decisions are things like: should this be a queue or a stream, what is the blast radius when this account is compromised, is the cost of multi-region justified by the actual availability requirement, is this a platform problem or an organisational one. These need context that mostly is not written down. Commercial pressure, team capability, what was tried in 2023 and why it failed.

Mechanics are the rest, and there is far more of it than anyone admits: writing the twentieth Terraform module that resembles the previous nineteen, porting pipelines from one CI platform to another, generating tests for existing behaviour, sweeping a repository for every place a deprecated pattern appears, writing the documentation that should have been written at the time.

Agents are very good at mechanics and unreliable at decisions. Almost everything useful follows from taking that seriously.

Where the gain is largest

The biggest wins are not where people expect. Writing new code faster was never the bottleneck, so that is not where the gain shows up. It shows up in work that is valuable but was previously uneconomic.

Breadth. Auditing every module in a large estate instead of sampling three and extrapolating. A human reviewer checks a representative sample because checking all 200 would take a fortnight. An agent checks all 200. Problems that were previously found by incident get found by audit instead.

Migrations. Mechanical translation is exactly the shape of work agents do well: same semantics, different syntax, repeated hundreds of times with small variations that a regular expression cannot handle but a model can.

The work that always gets cut. Tests for legacy behaviour, documentation, consistent naming. This is skipped under deadline, and skipping it is why estates decay. When it costs an hour instead of a week, it stops being the first thing sacrificed.

Where it goes wrong

Three failure modes, all of which we have hit.

Confident wrongness in unfamiliar territory. An agent working in a codebase it has not been given context for will produce something plausible and subtly wrong. The fix is unglamorous: written conventions, architectural context, and explicit statements of what must never be touched.

Review debt. If agents generate more change than your team can meaningfully review, you have not gained capacity. You have moved the bottleneck, and made it harder to see. A queue of plausible diffs nobody has read closely is worse than a slower team.

Production surprises. Anything touching live data, security boundaries or state needs a person who understands the consequences. Not as ceremony. The failure modes are irreversible and the agent has no stake in them.

How to tell whether a consultancy means it

Ask for the artefact.

A team that really works this way has something to show: a repository, a pipeline, an estate they built, with history you can read. Ask what proportion of it was agent-generated and how it was reviewed. Ask which parts they deliberately kept manual, because a good answer names specific things. Migrations, secrets handling, anything with an irreversible failure mode.

An outfit that has added AI to its pricing page but cannot show you a diff is selling you the same engagement with a different adjective.

What it means commercially

Done properly, this changes what is worth doing, not just how fast it happens. Work that was previously too expensive to justify becomes affordable. Comprehensive test coverage on legacy systems, full-estate audits, documentation that stays current.

That is the honest pitch. Not "the same work at half the price", which invites you to wonder what was removed, but a wider range of work worth commissioning at all.