Which engineering work can AI actually do?

Semantic analysis of 69,156 real tickets from the Spring Framework project reveals three tiers of AI readiness — and shows how the balance shifts as a project matures.

Data source: Spring Framework Jira archive, 2003–2021  ·  Embeddings: OpenAI text-embedding-3-large  ·  Clustering: UMAP + HDBSCAN

69k
tickets analysed across 20 years of development
22%
safe to automate — no human review needed
30%
AI-assisted — draft + human sign-off
48%
escalate to senior engineer — too risky for AI
AI readiness shifts as a project matures
Percentage of tickets per year by readiness tier. As Spring's easy work gets resolved, complex architecture problems dominate.
Overall split — all 69k tickets
Across the full project history.
What the trend means

As Spring matured, the automatable work got done and what was left became genuinely hard. By 2020 nearly 70% of new tickets were architecture-level problems — security enhancements, JPA edge cases, distributed messaging, framework internals. The routine work didn't disappear — it was resolved. What remained required real expertise.

This is not a story about AI making engineers redundant. It is the opposite. As AI absorbs the repetitive tier, engineers are left with the work that is hardest, most impactful, and most valuable to their careers. The triage tool that routes work correctly becomes more critical as adoption grows, not less.

Product insight

AI automation doesn't shrink the escalate pile over time — it accelerates through the easy work and exposes the hard problems faster. The triage tool becomes more valuable as adoption grows, not less.

Embedding density validates the classification
Mean cosine similarity to the 20 nearest neighbours within each tier. High density = tight, repetitive patterns.
Automate 0.537  — tight, repetitive patterns
Assist 0.496  — edge cases, each one slightly different
Escalate 0.526  — tight, but the pattern is "this is hard"

The Assist tier having the lowest density is the key signal — these are the genuinely ambiguous tickets that sit at the boundary between known and unknown territory. That is exactly where a human in the loop adds the most value: not rubber-stamping AI output, but catching the cases where the AI is confidently wrong.

Readiness by cluster
AutomateDocumentation corrections & edits2,913
AutomateSpring dependency upgrades2,663
AutomateSpring Data deprecation updates2,039
AutomateRelease version updates1,794
AutomateCodebase consistency & refactoring1,457
AssistRequestMapping enhancements3,036
AssistContent negotiation issues2,504
AssistTransaction management issues1,628
AssistMongoDB mapping issues1,805
EscalateJDBC compatibility & performance2,911
EscalateJMS message listener issues1,982
EscalateAuthentication framework enhancements1,279
EscalateJPA entity mapping issues1,744
What the data tells us
Automatable work is repetitive by nature. Docs fixes, version bumps, deprecation replacements, and codebase formatting follow tight patterns — embedding density 0.54, the highest of any tier. AI can handle these confidently from historical precedent alone.
The assist tier is the most ambiguous. Lowest embedding density (0.50) — each ticket is slightly different from its neighbours. These are exactly the cases where AI drafts a solution but a human must verify correctness before merge.
AI adoption accelerates the hard problem. From 2010 to 2020, automatable work fell from 19% to 5% while escalate-level work rose from 36% to 68%. As AI clears the easy backlog, what remains is genuinely hard — making triage and oversight more critical, not less.
Your backlog has the same patterns. Every engineering team accumulates automatable work alongside genuinely hard problems. Connect your Jira or GitHub project and task2vec maps your history the same way — so you know which tickets to hand to an AI agent and which ones need a senior engineer, before anyone wastes time on the wrong one.
Cross-dataset validation

The same pattern holds across 1 million Apache tickets

To validate that this is not a Spring-specific quirk, we ran the same analysis on the full Apache Software Foundation Jira archive — 25+ projects spanning Spark, Kafka, Hadoop, Cassandra, Flink, and more.

1.0M
Apache tickets across 25+ open-source projects
31%
safe to automate — consistent with Spring finding
52%
AI-assisted — larger share in younger projects
17%
escalate to senior engineer
Apache year trend — automatable work stays stable
Unlike Spring's sharp decline, Apache's automatable work holds at ~31% from 2005 to 2021. The project mix includes newer projects at earlier maturity stages, offsetting the drift.
Spring vs Apache — overall split
Spring's higher escalate % reflects its single-framework depth. Apache's breadth includes more routine infra work. Both datasets show the same core pattern.
AI readiness varies dramatically by project domain

Same classification across 25 Apache projects reveals that domain complexity — not just team size — determines the tier mix.

Project Auto Assist Escalate Tickets
OFBIZ59%32%9%12,470
YARN49%39%12%10,769
HDFS49%39%12%15,588
SPARK44%42%14%37,443
HBASE43%40%17%26,421
KAFKA32%49%20%12,312
FLINK39%40%21%25,492
IMPALA29%39%33%11,031
AMBARI9%59%32%25,384
CASSANDRA9%83%8%17,115

OFBIZ (business app) automates 59% of its work. CASSANDRA (distributed database) has 83% in the assist tier — every change is novel but not necessarily critical. AMBARI and IMPALA have the highest escalate rates, driven by infrastructure complexity. No two projects share the same profile — exactly what task2vec maps.

Tickets as a mirror

What your Jira reveals about your organisation

Without reading a single roadmap document or running a single survey, ticket data alone reconstructs what a team actually builds, where its expertise concentrates, and which direction its work has been drifting. Here is what the two archives tell us.

Spring Framework — the Java enterprise middleware stack

69,156 tickets across 80+ sub-projects. The distribution reveals a deliberate platform strategy: own the abstraction layer between application code and everything else — every database, every messaging system, every security protocol.

The 12 Spring Data sub-projects (JPA, MongoDB, Redis, Cassandra, Elasticsearch…) each have their own project key. That is a deliberate architectural bet visible purely from ticket structure — no roadmap document needed.

Apache Software Foundation — the data infrastructure layer of the internet

1,014,926 tickets across 350+ projects. The domain breakdown reveals a single thesis: build the distributed systems that move, store, and process data at scales single machines cannot handle.

Big data alone accounts for 15% of all tickets — Spark, Hive, Hadoop, HDFS, Impala. Add streaming (Kafka, Flink, Beam) and distributed databases (Cassandra, HBase, Ignite) and over a quarter of the Foundation's work is the data pipeline stack.

Real strategy detection

Every one of these insights — what the team builds, where expertise concentrates, which direction work is drifting — came from ticket metadata alone. No interviews. No surveys. No process changes. Connect your Jira or GitHub and task2vec produces the same map for your team in minutes.

Validation finding

Across 1,084,082 tickets from two completely independent codebases — Spring Framework and 25 Apache projects — the same three-tier pattern emerges. Roughly 25–31% of work is safely automatable, 30–52% benefits from AI assistance with human review, and 17–48% requires senior engineering judgment. The exact ratio depends on project domain and maturity. What does not change is that all three tiers are always present — and without a tool to tell them apart, teams route work by guesswork.

Try it live

Paste any ticket and see which tier it falls into — scored against 40 k real Spring outcomes. No rules. No LLM. Pure historical signal.

Score a ticket →