A new software-modernisation case study deserves attention, but not for the usual reason.
Asana says it used AI coding agents to remove Enzyme, an ageing frontend testing library, from its codebase in about two calendar weeks. The company had previously estimated that finishing the migration through its existing approach could take another five years.
The eye-catching numbers are Asana's own estimates: approximately 1.5 weeks of engineering effort, about US$12,000 in model and infrastructure costs, and roughly US$6 million in fully loaded engineering effort for the broader manual scope. Those figures should be treated as a case study, not a universal return-on-investment calculator. The more transferable lesson is how the work was structured.
Asana did not ask an agent to modernise an application and hope for the best. It selected a repetitive migration with a clear finish line, divided the codebase into bounded work areas, ran up to four agents in separate copies, used fast automated checks, and required engineers to review every proposed change.
That combination is the real opportunity for businesses carrying years of deferred software maintenance.
What Asana Actually Did
According to Asana's engineering account, the company had been moving frontend tests from Enzyme to React Testing Library since 2022. Enzyme was no longer actively supported, created friction with newer React versions and encouraged tests tied closely to implementation details.
The team gave each coding agent a short goal: migrate one directory, follow existing codebase conventions, run the relevant tests and start with easier files. Up to four agents worked in parallel, each isolated in a separate copy of the repository. Engineers checked progress morning and evening, reviewed changes and opened pull requests.
The company's report highlights four conditions that made the task unusually suitable:
- A precise definition of done: remove all remaining Enzyme use.
- Many existing examples: the repository already contained the preferred React Testing Library patterns and helpers.
- Fast feedback: type checks, linting, tests and CI could reject incorrect changes.
- Human review: agents prepared changes; engineers decided what entered the codebase.
OpenAI's 18 August case study confirms the same project details, including the use of four parallel agents, twice-daily engineering checks and review of proposed changes.
The Headline Is Speed; the Lesson Is Readiness
It is tempting to turn this into a simple claim that AI can compress any five-year project into a fortnight. It cannot. Asana explicitly cautions that not every long-running project will collapse into weeks.
This migration worked because the desired transformation was already understood. The team was not asking AI to invent a new product, resolve ambiguous stakeholder needs or make untestable architectural decisions. It was applying a known target pattern repeatedly across a large codebase.
That distinction matters. AI agents are strongest when they can:
- recognise a pattern in good existing examples;
- apply it to many bounded locations;
- receive rapid, objective feedback;
- produce small enough changes for meaningful review;
- stop or escalate when the expected pattern does not fit.
For a business, the question is therefore not merely, "Which coding agent should we buy?" It is, "Which parts of our maintenance backlog have this shape, and is our delivery environment ready to verify the work?"
Good Candidates for AI-Assisted Modernisation
The best first projects are usually repetitive, measurable and reversible. Examples include:
- migrating tests from a deprecated framework to an established replacement;
- updating deprecated API calls where the target pattern is documented;
- converting components to an approved design-system pattern;
- adding missing type annotations or resolving a defined class of static-analysis findings;
- upgrading framework conventions across modules with strong regression coverage;
- replacing a legacy logging or configuration pattern;
- removing a dependency after its usages have been migrated and verified.
Poor first candidates include loosely defined rewrites, changes with no reliable tests, transformations that depend on undocumented business rules, and work where a subtle error could affect payments, privacy, safety or regulatory obligations without an effective review boundary.
This is consistent with our earlier guide, Where AI Actually Pays Off in Software Development: AI performs best on bounded, verifiable tasks. The Asana example adds useful evidence about what that principle looks like at migration scale.
Build the Harness Before Scaling the Agents
Asana describes the surrounding environment as the "harness": the documentation, conventions, examples and feedback loops that steer the agent. For most organisations, improving that harness will create more value than writing a longer prompt.
A practical modernisation harness should include:
- A baseline: record the current build, test and quality state before automation starts. Existing failures must not be mistaken for agent regressions.
- A machine-checkable finish line: for example, zero imports from a retired library, all targeted tests converted, and no reduction in coverage.
- Representative examples: keep several approved implementations in the repository so the agent can follow the project's real patterns.
- Fast local checks: agents need quick lint, type and targeted-test feedback before the slower full pipeline runs.
- CI parity: local and pipeline checks should agree. A workflow that passes locally and fails unpredictably in CI wastes both compute and engineering attention.
- Small change boundaries: divide work by directory, component or migration class. Do not let parallel agents modify overlapping areas without a deliberate integration plan.
- Protected approval: the agent may create a branch or pull request, but it should not approve its own work or bypass deployment controls.
- Traceability: capture the task, tool and model version, changed files, checks performed, failures, retries and human decision.
These are ordinary software-engineering controls, not AI theatre. They also support the governance approach described in Governing Agentic AI in Government Software Delivery: scoped identity, least privilege, deterministic verification, human accountability and a safe way to stop.
Stale Documentation Becomes an Operational Defect
One of Asana's most useful observations was that some internal guidance still recommended Enzyme. Human developers might notice that a document is old, ask a colleague or rely on newer experience. An agent can repeatedly apply the stale instruction at machine speed.
Before an AI-assisted migration, review the sources the agent is likely to read:
- repository instructions and contributor guides;
- architecture decision records;
- framework examples and starter templates;
- test helpers and fixtures;
- generated documentation;
- tickets, wikis and copied code samples.
Mark superseded guidance clearly, remove instructions that are actively wrong, and state which source has authority when documents conflict. Documentation hygiene is now part of runtime reliability for agentic development.
How to Run a Responsible Two-Week Pilot
A business does not need to commit its entire modernisation backlog to AI. A short pilot can test whether the economics are real in its own environment.
- Days 1–2: select and measure. Choose one repetitive migration, count the affected locations, sample complexity and record the current manual completion rate.
- Days 3–4: repair the harness. Update stale guidance, add representative examples, stabilise targeted checks and define success and stop conditions.
- Days 5–7: run a small batch. Give an agent one bounded area with branch-only permissions. Review every change and classify failures.
- Days 8–10: expand carefully. Parallelise only independent work areas. Keep changes small and prevent overlapping edits.
- Days 11–12: integrate and test. Run the complete regression, security and quality pipeline and inspect unexpected behavioural changes.
- Days 13–14: decide with evidence. Compare accepted output, review effort, elapsed time, defects and total model, infrastructure and engineering costs against the baseline.
The pilot succeeds if it produces a reliable decision. That decision may be to scale the workflow, improve tests before continuing, narrow the task, or stop. A dramatic demonstration without trustworthy verification is not a modernisation strategy.
What This Means for GovCMS and Drupal Teams
GovCMS and Drupal estates often contain exactly the kind of long-running maintenance work that deserves assessment: deprecated APIs, custom modules, theme changes, configuration patterns, test gaps and upgrade preparation.
Suitable agent-assisted work may include preparing patches for deprecated Drupal APIs, generating tests around established module behaviour, checking custom code against project standards, or converting repetitive theme and component patterns. However, the same controls apply: use representative examples, run the full project-specific test and security pipeline, keep production credentials out of the agent environment, and require qualified review.
Government teams also need to account for records, privacy, hosting, procurement and security requirements. An agent's speed does not relax those obligations. It makes traceability and permission design more important because more changes can be produced in less time.
Working with NAITEC Digital
NAITEC Digital helps businesses and government teams identify legacy-modernisation work that is genuinely suitable for AI automation, then build the delivery harness required to do it safely.
We can assess a backlog, establish a measurable pilot, strengthen automated testing and CI/CD, encode project conventions, integrate branch-based coding agents, and design review and approval controls. Our goal is not to maximise agent activity. It is to reduce maintenance cost without creating a faster path to unreviewed defects.
NAITEC Digital is a Newcastle, NSW software consultancy, a BuyICT registered supplier, and a GovCMS/Drupal specialist on the Drupal Services Panel. Our capabilities span AI integration, automation and custom software delivery and government digital services, GovCMS and Drupal.
If a migration has been sitting in your backlog for years, talk to NAITEC Digital. We can determine whether it has the right shape for a controlled AI-assisted pilot.
Frequently Asked Questions
Did AI really complete five years of Asana engineering work in two weeks?
Asana reports that it removed Enzyme in about 1.5 weeks of engineering effort across two calendar weeks, after previously estimating roughly five more years under its existing migration approach. It also reports about US$12,000 in model and infrastructure costs versus an estimated US$6 million for the broader manual scope. These are company-reported estimates from one unusually suitable project, not a general promise for every migration.
What kinds of legacy migration are best suited to AI coding agents?
Repetitive transformations with an established target pattern, many good examples, a machine-checkable finish line, fast automated tests and small reviewable change boundaries are the strongest candidates.
Why are tests so important for AI modernisation?
Agents can produce changes much faster than humans can inspect every detail manually. Automated tests, type checks, linting, security scanning and CI provide rapid objective feedback and let reviewers concentrate on behaviour, architecture and risk.
Should an AI agent be allowed to merge or deploy its own migration changes?
Usually no. Let the agent prepare changes in an isolated branch or pull request, then use protected checks and an authorised human reviewer before merge or deployment. Separate proposing a change from approving it.
What should a business measure in an AI modernisation pilot?
Measure accepted changes, review time, failure and rework rates, regression results, elapsed time, model and infrastructure spend, engineering effort and the remaining backlog. Compare those results with a manual baseline rather than relying on generated-code volume.
Can NAITEC Digital help assess an AI-assisted migration?
Yes. NAITEC Digital can identify suitable work, strengthen the test and delivery harness, implement a bounded pilot and establish the governance and review evidence needed to decide whether the workflow should scale. Contact us to discuss your backlog.