AI transparency: how CVPitch uses AI on candidate CVs
CVPitch AI transparency: where Anthropic's Claude Haiku is used, what it never does, no model training, human approval, job match and the EU AI Act.
This AI transparency note explains, in plain terms, where CVPitch uses artificial intelligence on candidate CVs, which model it uses, what the AI is never allowed to do, and where a person always decides. Recruiters, candidates and clients should be able to see how a CV was produced, so we publish this alongside our security and privacy pages.
AI transparency in short
- One provider, one model family. AI features use Anthropic's Claude, Haiku model, through Anthropic's commercial API.
- AI is optional per CV. "Format only — no AI" formats a CV by rules and, for PDFs with a text layer and Word files, sends nothing to an AI provider. Scans and photos still need AI text recognition.
- AI extracts and proposes; people decide. Every fact shows its source passage, AI wording is a proposal you accept or reject, and a recruiter approves every CV before it is exported or shared.
- No decisions about candidates. CVPitch does not score, rank, filter or reject candidates. Job match is rule-based and shows evidence; it does not use AI judgement.
- No training on your data. CVPitch does not train AI models on candidate data and does not sell it.
Where AI is used
| Feature | Uses AI? | What the AI receives | What comes back |
|---|---|---|---|
| CV extraction (default mode) | Yes, Claude Haiku | The CV's text, split into numbered source blocks | Structured facts, each with a quote that must match the source exactly |
| Format only — no AI | No, except text recognition for scans and photos | Nothing for PDFs with a text layer and Word files; page images of scans and photos for text recognition | Sections and lines laid out by rules, each a literal quote of the source |
| Text recognition (OCR) for scans and photos | Yes, Claude Haiku | The page images | A transcription, flagged for a recruiter to check against the image |
| AI writing: rewrite, summary, proofreading, chat edits, translation | Yes, when a recruiter asks | Only the fields being edited, with the name and contact details replaced by placeholders where detected | Proposed changes shown word by word |
| Employer and institution descriptors | Yes, when a recruiter asks | Employer and institution names | Suggested descriptors such as "Global investment bank" |
| Schema extraction API | Yes, when an integration calls it | The submission's source blocks and the caller's JSON Schema | Values matching the schema, with any string not found word for word in the source listed for review |
| Job match | No | Nothing is sent | Requirement statuses from rules, each with its CV passage |
| Contact removal, rendering, hidden-text removal, brand autofill, email routing | No | Nothing is sent | Deterministic results |
What the AI is never allowed to do
- Act on instructions inside a CV. Candidate documents are treated as data. Text hidden in a CV, such as white-on-white or tiny text, is detected and stripped at upload, reported in review with counts, and never written to logs.
- Use tools. The model cannot browse, call integrations, send email or take any action. It returns structured data that is validated before use.
- Introduce unsupported facts silently. An extracted value must appear word for word in a quote that matches the source. If not, the extraction is rejected rather than shown as fact. Facts without support are marked unverified.
- Change protected facts in AI writing. Rule-based checks drop suggestions that change names, employers, job titles, dates, qualifications or locations, or that add numbers, links, contact details or names not found in the CV. These checks can miss a change, which is why every suggestion is reviewed.
- Approve anything. Neither the AI, the API, email intake, ATS connectors nor AI assistants connected through MCP can approve a submission.
Human review and approval
A recruiter reviews each CV beside its source, resolves review flags, and approves the exact version to be sent. The approval covers the facts, the client template version and the contact-removal settings; any later change, including an accepted AI suggestion, needs approving again. Every fact is labelled with its origin: extracted with evidence, recruiter-corrected, recruiter-added or AI-assisted. If extraction missed something, the recruiter can draw a box on the original page and assign the words to a field, without another AI call.
Job match and human oversight
Job match lists the requirements in a job description and marks each one evidenced, partly evidenced or not shown, with the CV passage beside it, then drafts an editable pitch email and a list of gaps to confirm with the candidate. It is deterministic: the statuses come from rules applied to the CV text, not from a model's opinion. It does not produce a score, a ranking or a recommendation to reject anyone, and nothing is sent automatically. The recruiter decides whether and how to put a candidate forward.
The EU AI Act
The EU AI Act lists AI systems intended for recruitment or selection, for example to filter applications or evaluate candidates, among its high-risk uses (Annex III, point 4). We have designed CVPitch so that its AI formats and extracts what a CV already says, while decisions about candidates stay with people: there is no automated scoring, ranking or filtering, job match is rule-based, and every output is approved by a recruiter.
We are not making a legal determination about how the Act applies to CVPitch or to your use of it. How a tool is classified can depend on how it is used, so agencies using CVPitch as part of hiring decisions should assess their own obligations, and can contact us for information about the system that supports that assessment.
Data, providers and limits
- Provider: Anthropic (Claude API). Our privacy page and the subprocessor table set out where data goes.
- No training: CVPitch does not use candidate data to train models. See Anthropic's own commercial terms for how it handles API data.
- Bounded cost and retries: every AI request has a hard cost cap, retries are limited, failed requests never use your credits, and a monthly processing budget applies across the service.
- Accuracy: AI extraction and text recognition can be wrong. We do not publish accuracy percentages; the fields and languages page sets out what we have evaluated.
- Labelling: local demonstrations without a live AI provider use a visibly labelled fake AI.
Frequently asked questions
Can I use CVPitch without AI?
Yes. Choose "Format only — no AI" when you upload. For PDFs with a text layer and Word files, no text is sent to an AI provider, the candidate's wording is kept exactly, and no AI credits are used. Scans and photos still need AI text recognition (OCR) to read the page, which sends the page images to the AI provider and uses 1 AI credit per page. Review, approval and contact removal work as usual.
Does CVPitch's AI decide which candidates to send?
No. CVPitch does not score, rank or filter candidates. Job match shows the evidence for each requirement using rules, and a recruiter decides.
Is candidate data used to train AI?
No. CVPitch does not train AI models on candidate data or sell it. AI processing uses Anthropic's commercial API.
Which AI model does CVPitch use?
Anthropic's Claude, Haiku model, for extraction, text recognition and the optional AI writing tools. See features for what each tool does.
Turn your next CV into a client-ready submission
Upload a candidate CV, check every fact against the source, remove contact details and export your agency’s branded DOCX and PDF. Your first 10 CVs are free.