AI Extraction for Candidate Data

Why It Matters

Reviewing dozens of applications means manually reading each resume, finding the salary expectation buried in a cover letter, and typing it into a field. That adds up fast. AI extraction reads every piece of candidate material submitted — resume, form responses, questionnaires — and fills in your custom data fields automatically, so your team opens each application with the key facts already surfaced.

How It Works

AI extraction runs in three steps:

1. Mark a field as AI-extractable — on any custom data field, enable the AI extraction toggle
2. Write an extraction prompt — tell the AI exactly what to look for in that field
3. AI processes on application — when a candidate applies, extraction runs automatically and populates the field

The extracted value sits alongside a confidence score so your team knows how certain the AI was. High and medium-confidence values surface on kanban cards and in Slack. Low-confidence values are stored but flagged for review.

Enabling AI Extraction

Open a job posting, go to Data Fields, and edit any field. Toggle AI Extractable on. A prompt field will appear — this is required.

Write a prompt that tells the AI exactly what to find. The prompt becomes the extraction instruction sent to the model.

Examples of good prompts:

Field	Good Prompt
Salary Expectation	Extract the candidate’s expected salary, compensation range, or rate. Return null if not mentioned.
Availability	Extract when the candidate can start or their notice period. Return null if not mentioned.
Location Preference	Extract the candidate’s location, city, country, or remote work preference.
Years of Experience	Extract total years of professional experience. Return a number. Return null if not clear.
Open to Relocation	Return true if the candidate mentions willingness to relocate. Return false if they say they cannot. Return null if not mentioned.
Technical Skills	Extract all programming languages, frameworks, and tools mentioned. Return as a list.

Prompts that say “Return null if not mentioned” reduce false positives — the AI won’t guess when the information isn’t there.

What Gets Analyzed

The AI receives everything the candidate has submitted, organized into sections:

Source	What’s Included
Candidate profile	Name, email, phone, GitHub username, LinkedIn URL
Resume	Full resume text, parsed work history, education, skills list, experience summary
Form responses	All answers submitted on the application form
Questionnaire answers	Written responses from any questionnaire stages the candidate completed

The AI only receives material the candidate submitted — nothing from your internal notes or team evaluations.

Writing Effective Prompts

The prompt is the most important part of extraction quality. Match your prompt to the field type:

Field Type	Example Field	Prompt That Works
Text	Salary Expectation	“Extract the candidate’s expected salary or compensation range. Include currency if mentioned. Return null if not mentioned.”
Number	Years of Experience	“Count total years of professional experience from work history. Return a number only. Return null if not determinable.”
Boolean	Open to Management	“Return true if the candidate has managed a team or expresses interest in management. Return false if they prefer individual contributor work. Return null if not mentioned.”
Tags	Technical Skills	“Extract all programming languages, frameworks, databases, and tools the candidate mentions. Return as a list of strings.”
Select	Seniority Level	“Classify the candidate as Junior, Mid, or Senior based on their experience. Return null if not enough information.”
Date	Available From	“Extract the earliest date or timeframe the candidate can start. Convert to a date if possible. Return null if not mentioned.”

Keep prompts focused on a single piece of information. Compound prompts (“extract salary and notice period”) produce inconsistent results.

Confidence Scores

Every AI-extracted value comes with a confidence score from 0.0 to 1.0, grouped into three tiers:

Tier	Score Range	What It Means	Shown on Cards/Slack
High	0.75 – 1.0	AI found a clear, explicit answer	Yes
Medium	0.50 – 0.74	AI found a likely answer with some ambiguity	Yes
Low	Below 0.50	AI guessed or inferred — verify before acting	No

High and medium-confidence values appear on kanban cards and in Slack candidate channels. Low-confidence values are stored in the application record but kept out of compact views to avoid surfacing unreliable data.

When reviewing an application, confidence is shown next to each AI-filled field so your team knows at a glance whether to trust the value or read the source material themselves.

Manual Override

Any AI-extracted value can be edited inline. Click the field value on the application page to edit it. Once you save, the field is marked as manually set — the AI will not overwrite it on future extractions.

The original AI-extracted value is preserved in the record even after you edit it, so you can always see what the AI found before your team made a change.

To force re-extraction on a manually-edited field, clear the value first, then run re-extraction.

Re-extraction

Automatic: Extraction runs once when a candidate applies. New applications are processed automatically — no action needed.

Manual re-extraction: Open the application, scroll to the Data Fields section, and click Re-extract. This re-runs extraction against all current candidate material using the current field prompts.

Re-extraction is useful when:

• You add a new AI-extractable field after the application was received
• You updated a field’s prompt and want fresh results
• The initial extraction failed

Re-extraction respects manual overrides — fields your team has edited will not be overwritten.

Extraction Status

Each application has an extraction status visible in the Data Fields section:

Status	Meaning
Pending	Extraction is queued but hasn’t started yet
Processing	Extraction is running now
Completed	Extraction finished — results are populated
Failed	Extraction encountered an error — see the error message for details

Common failure reasons include no source material (the candidate submitted nothing the AI can read) or an AI credit limit being reached. If extraction fails, fix the underlying issue and use the manual re-extract button to retry.

Quick Checklist

• Add custom data fields to the job posting
• Enable AI Extractable on fields you want auto-filled
• Write a focused extraction prompt for each field
• Use “Return null if not mentioned” to avoid false positives
• Review medium-confidence values before making decisions
• Edit any incorrect AI values inline — your edit locks the field from re-extraction
• Use Re-extract after adding new fields to existing applications