Understanding AI-Generated Documentation Quality
aprity uses AI to write readable documentation from your Salesforce metadata. Knowing what is computed, what is written by AI and how it is checked helps you decide how far to trust each page.
Facts first, AI second
aprity follows one principle: AI explains facts; it does not decide them.
Computed from your metadata, without AI
- The list of objects, fields, rules, flows, classes and other components.
- Dependencies: which component references which.
- Execution order: when triggers, flows and validation rules run on save.
- Which operations (create, update, delete) a rule applies to, when this can be established.
- Access: profiles and permission sets and what they grant.
Written by AI
- Business descriptions of objects, rules and processes.
- The narrative of functional processes.
- Summaries and classifications (for example grouping rules by topic).
Intent is only stated when there is a source
aprity states why something exists only when it can quote a source in your org: a code comment, a description, help text, an error message, a label. Without a source, the documentation says what the component does and leaves the why out, rather than inventing a reason. This is why business context sometimes looks missing: your team can add it with feedback.
The fidelity certificate
Each analysis produces a fidelity certificate, in Coverage and fidelity > Verification. During the analysis, before the metadata is purged, aprity takes a sample of claims from the documentation and checks them against your org's actual metadata.
- Two verification modes. Some items are checked without AI, against the org model. Narrative items are reviewed by a judge model. The certificate shows a score for each mode, and the overall score weights them by the number of items checked.
- Verdicts. Each claim is Verified, Divergence, Partial or Not verifiable. Divergences are shown in full, with the source excerpt.
- Scope. The score covers only what was verified. The certificate says what it does not reach: content types no verifier samples, and objects whose documentation was not produced.
Read the score with its scope: a high score on a small sample, or on a version with objects not documented, says less than it seems. See Read the fidelity certificate.
Signs of good output
- Specific references to your fields, objects and conditions.
- The same terms used for the same concepts across pages.
- Rules described with their real conditions and exceptions.
Signs that feedback is needed
- Generic text ("this rule validates data") without the conditions.
- A business purpose that is wrong for your company.
- Missing context that only your team knows (a contract, a regulation, an external process).
- Your company's vocabulary replaced by generic words.
What feedback can and cannot fix
Give feedback for factual errors, missing context, vocabulary and clarity. Feedback on rules, object overviews, processes and components is taken into account by the next analysis. See Building a Feedback Loop.
Feedback does not fix:
- Confusing metadata. If a rule or a field has a misleading name in Salesforce, fix it there. aprity documents what exists.
- The structure of the documentation. The organization of pages and sections is computed, not written by AI.
Known limits
- External systems. aprity can see that your org calls an external system, and what it sends. It cannot describe what the external system does.
- Very complex logic. Deeply nested Apex logic may be summarized. The main paths are described; not every edge case.
- One org at a time. Links between two of your Salesforce orgs are not documented.