How it works
Code counts. The model drafts. A person decides.
Language models are good at refactoring patterns and bad at counting, so each step is done by the right kind of tool. Every step is written to an audit trail.
- 01
Collect
Parsers read code, OSGi configs, dispatcher, indexes and your BPA report.
DeterministicDone by code alone: the same input always gives the same result. - 02
Classify
Every finding maps to a Cloud Service pattern with file and line.
DeterministicDone by code alone: the same input always gives the same result. - 03
Plan
Our private model ranks risk and drafts the change for each pattern.
AI on our serverDone by a model on our own server, then checked by code before a person sees it. - 04
Refactor
Small pull requests, one pattern at a time, with notes and tests.
AI on our serverDone by a model on our own server, then checked by code before a person sees it. - 05
Review
A senior AEM engineer approves every pull request. Your team merges.
HumanA named person approves; nothing merges without it. - 06
Verify
Cloud Manager build, AEM Analyser and parity checks; all steps logged.
DeterministicDone by code alone: the same input always gives the same result.
Step by step
What happens at each stage
The fact sheet comes first and is produced by code alone. The model only works on what the fact sheet contains, and nothing reaches your main branch without a person approving it.
Collect
Parsers read your repository, OSGi configurations, dispatcher files, package filters, Oak indexes, your BPA report and content statistics. The output is a fact sheet: every finding with file, line and rule. No model is involved, so the inventory is complete and repeatable.
Classify
Each finding is matched to a Cloud Service pattern from the open rule set (33 code rules and 31 BPA codes), with the documented fix, how automatable it is and the planning hours.
Plan
For each pattern, the model on our own server proposes the target (Cloud Service or an Edge Delivery block), the risk, and whether it can be fixed mechanically. It reasons only about items in the fact sheet, so it cannot invent findings.
Refactor
5 codemods apply the mechanical rewrites without any AI. Patterns that need judgement are rewritten by the model and must pass deterministic checks before a person sees them. Adobe's own migration tools are used where they cover a pattern.
Review
A named senior AEM engineer reviews and approves every pull request before it reaches you, and your team merges. Rejections and edits are fed back so the next change is better.
Verify
Cloud Manager build and code quality gates, the AEM Analyser, generated regression tests, visual comparisons of key pages and an SEO parity check of URLs, redirects and metadata.
Where the AI runs
Your code stays with models we operate
Client code is processed by AI models running on servers we operate. It is not sent to third-party AI services.
- Private by default. The routing rule is enforced in code: tasks that contain client code cannot go to an external AI service.
- External models only with consent. If a task would benefit from an external model such as Anthropic's Claude, we ask first and use it only with your written consent, recorded in the audit log.
- Checked before a person sees it. Every AI rewrite must pass deterministic checks, for example that no imports are invented.
- Every step recorded. Model calls, check results and approvals are appended to a hash-chained log. Editing or deleting an entry breaks the chain, so you can prove who approved what and which model produced each change.
Limits
What we do not automate
Some work needs people who know your business. These stay with engineers and with your team.
Security and integration design
Authentication, security and integrations with other systems are designed by engineers, not generated.
Product decisions
Replacing AEM Communities or moving a site to Edge Delivery Services are decisions your team makes in the portal.
Merging to your main branch
Every pull request is approved by a named engineer, and your team merges it.
Start with the fact sheet
The readiness scan gives you a first fact sheet in your browser: every finding with file, line, fix and effort. No sign-up, and nothing is uploaded.