■ HIGH RISK ■ Technology
Substantially, for the standard service-and-database work that makes up most backend jobs. What holds out is the systems end, where correctness under concurrency, failure and load is decided by judgement rather than by pattern matching.
“APIs and CRUD are easy pickings for AI. Distributed systems at scale? Not yet.”
Our AI replacement risk score — how we score jobs
Most backend work is plumbing with rules attached: designing endpoints, writing handlers, validating input, modelling tables and writing migrations, wiring authentication and authorisation, queuing background jobs, integrating third-party APIs, adding caching, writing integration tests, and keeping the whole thing observable. Then there is the operational half, which is on-call rotations, slow queries appearing at month-end, a dependency deprecating an endpoint, and the periodic migration from whatever the last team chose to whatever this team prefers.
Generation handles a striking amount of that competently. Scaffolding an entire REST or GraphQL service with validation, tests, migrations and documentation is close to a solved prompt. Agents write ORM queries, generate client SDKs, produce infrastructure-as-code, upgrade dependencies and translate services between languages. Log and trace analysis tools summarise incidents automatically and often name the culprit deployment. Given that a large fraction of backend roles are, honestly, CRUD over a relational database with business rules attached, our score of 68 is not an insult so much as an inventory.
The residue is where correctness stops being local. Data modelling decisions constrain a company for years, and a model optimising for the current ticket will happily denormalise you into a corner. Concurrency, idempotency, exactly-once delusions, distributed transactions, cache invalidation and partial failure are all areas where confident generated code is frequently subtly wrong, and the failure surfaces months later as data corruption. Capacity planning, cost control on a cloud bill, security architecture around authorisation boundaries, and multi-region consistency choices demand reasoning about a specific system's traffic and history. And the person paged when replication lag stalls a payment queue is still a human with root access and something to lose.
Automatability: our editorial assessment of current and near-term AI capability
Pressure is immediate and compounding through 2030. Service scaffolding is already routinely generated, and teams are visibly building more services per engineer. Expect backend headcount to concentrate at the platform and distributed-systems end, with application-level API work increasingly produced by agents under review. The counterweight is that AI-heavy products need serious backend infrastructure for retrieval, queuing, cost management and evaluation, which absorbs some of the displaced capacity.
It can write a working one quickly, and for simple applications that is often sufficient. The problems appear at scale and over time: schema decisions that become expensive to reverse, subtle concurrency bugs, missing authorisation checks, and caching that silently serves stale data. Generated backends need review by someone who understands the failure modes they will hit in a year.
Marginally, and only at the deeper end. Routine API and database work is just as generatable as UI components, and there is more of it. The difference is that backend has a genuine high-complexity tier around distributed systems, data consistency and production operations where mistakes are expensive and human judgement still clearly outperforms generation.
Data modelling under real business change, concurrency and consistency reasoning, observability and incident response, security and authorisation design, and cost-aware architecture on cloud infrastructure. Add the infrastructure that AI applications require, such as vector and hybrid retrieval, queuing, rate limiting and evaluation pipelines, because that demand is growing while routine CRUD demand shrinks.
Most likely yes, per unit of product delivered. The visible pattern is fewer engineers shipping more services, with generation covering the repetitive layers. Total demand does not collapse, because more software exists and AI features need substantial backend support, but the composition shifts sharply toward senior engineers who can own systems rather than implement endpoints.