Data Annotator Job Description Templates for Small AI and Data Teams
6 free templates covering general, image and video, text and NLP, audio, clinical, and QA lead annotation roles, with classification and pay guidance. Download as DOCX.
The first data annotator I ever hired produced beautiful work for six weeks. Then we hired a second one, and within a week the dataset was unusable. Nothing had gone wrong with either person. The problem was that every ambiguous case had been resolved inside one head, and no rule had ever been written down.
That is the whole difficulty of annotation hiring in one story. The job description is not really about the person, it is about whether you have decided what a correct label is. Most annotator postings skip straight to attention to detail and a list of tools, and they end up hiring for speed on work that only pays off in consistency.
At FirstHR we write hiring templates for teams doing this without an HR department: a small AI company, a data services vendor, a clinic building a documentation model, an operations team labeling its own backlog. The six templates below cover general annotation, image and video, text, audio, clinical, and the QA reviewer who owns the quality bar.
TL;DR
A data annotator labels raw data so a model can learn from it, working from a written guideline rather than personal judgment. The role is non-exempt hourly work in essentially every case, and contractor agreements are a common misclassification trap. There is no dedicated federal wage code, so benchmark to nearby occupations. Six templates below, downloadable as DOCX.
What a Data Annotator Actually Does
A data annotator turns raw data into training data by applying labels a machine learning model can learn from. The labels come from a written guideline that defines every class and every edge case, and the annotator applies that guideline consistently rather than deciding case by case.
The work is more structured than the phrase data labeling suggests. A mature project has a class taxonomy, a set of tie-break rules, a gold standard set of pre-adjudicated items, an accuracy target, and an agreement metric between annotators. The annotator works inside all of it, and the best ones improve the guideline instead of quietly absorbing its gaps.
Two things surprise employers new to this. First, on any project running longer than a few months, most of the day becomes reviewing and correcting labels a model already proposed, which is a different skill from labeling from scratch. Second, the federal AI Risk Management Framework published by the National Institute of Standards and Technology treats the provenance and quality of training data as a governance question, not a back-office one, which is a useful argument when someone proposes doing annotation as cheaply as possible.
One Title, Four Different Jobs
Data annotator is a single title covering at least four jobs that screen for different traits and pay differently. Posting a generic annotation role instead of a modality-specific one is the fastest way to fill a pipeline with people who cannot do the work you actually have.
Images and video
Pixel precision, huge volume
Boxes, polygons, keypoints, and masks, plus object tracking across frames. Screened on precision and stamina rather than education. The taxonomy and the occlusion rules are the hard part, not the drawing.
Text and language
Judgment against a rubric
Entities, relations, intents, sentiment, and response rating. Screened on reading comprehension and rubric discipline. Native fluency matters, and so does resisting the urge to label by personal taste.
Audio and speech
Transcription conventions
Verbatim transcription, speaker turns, timestamps, and acoustic tagging. Screened on typing speed, hearing, and accent tolerance. Pay structure needs care if you price by audio minute rather than by hour.
Clinical and specialist
Credential plus privacy
Medical, legal, and financial data annotated by people who already know the domain. Costs several times general annotation and brings its own privacy obligations, which belong in the posting.
Screen for the Trait the Modality Needs
Image and video annotation rewards precision and stamina, so screen with a timed sample batch and look at boundary quality. Text annotation rewards reading comprehension and rubric discipline, so screen with ambiguous items where the guideline gives an answer that differs from common sense. Audio rewards typing speed and accent tolerance, so screen with a genuinely messy recording rather than a clean one. Clinical annotation rewards domain knowledge, so screen with cases a general annotator would get confidently wrong.
What Belongs in the Posting
An annotator job description does four jobs at once: it describes the actual work, it filters people who will not last, it states the classification and data obligations that protect you, and it closes the candidate. Most annotation postings only do the first, which is why they attract volume and not fit.
The parts candidates read first
The modality: text, image, video, audio, or mixed
The tool they will spend all day inside
Whether the work is remote, hybrid, or on site
Hours, shift pattern, and whether volume is steady
The parts that filter applicants
Language, fluency, and any domain or clinical credential
The accuracy target and how it is measured
Whether the queue contains distressing content
Equipment, connection, and workspace requirements
The parts that protect you
Non-exempt status stated on the posting
Confidentiality and data-handling obligations
Timekeeping expectations for remote annotators
Equal opportunity statement and essential functions
The parts that win the hire
An hourly rate or a good-faith range
A paid trial rather than an unpaid sample
Who reviews the work and how feedback arrives
Whether the role can grow into QA or lead work
The omission that costs small teams the most is the accuracy target. Candidates want to know what good looks like and how it will be measured, and a posting that cannot say implies a team that has not decided. Our guide to writing a job description covers the general structure, and the hiring templates library has the rest of the postings a growing data team needs.
6 Data Annotator Job Description Templates to Download
Download all six as one file or copy them individually. Each follows the same structure: company overview, position summary, key responsibilities, required qualifications, a classification and compliance note, an equal opportunity statement, and how to apply. The bracketed fields are the only parts you need to change.
Download All 6 Data Annotator Job Description Templates
General, image and video, text and NLP, audio and speech, clinical, and QA reviewer. All in one download.
Data Annotator (General)
Mixed modality, entry level
The default posting when your projects move between text, images, and audio and you are hiring for accuracy rather than a specialty.
Image and Video Annotator
Computer vision
For bounding boxes, segmentation, and object tracking, with the accuracy target, the pre-label review reality, and the content warning stated.
Text and NLP Annotator
Entities, intents, response rating
For language data: rubric discipline, inter-annotator agreement targets, and rules for the personal information sitting in raw text.
Audio and Speech Annotator
Transcription and tagging
For speech data, with the transcription convention, the timing tolerance, and an honest note about paying by audio minute.
Clinical Data Annotator
Credentialed, HIPAA scope
For medical annotation by clinicians or coders, with license requirements, coding systems, and business associate obligations.
QA Reviewer / Lead Annotator
Guideline and gold set owner
For the person who owns quality: audits, agreement metrics, adjudication, and training, with a hard look at whether the role is really exempt.
Template 1: Data Annotator (General)
The default posting when your projects move between modalities and you are hiring for accuracy rather than a specialty. It sets the non-exempt classification, the confidentiality requirement, and the paid trial up front.
Data Annotator Job Description (General)
DATA ANNOTATOR JOB DESCRIPTION
Company: __ ([City, State] or Remote)
Reports to: [Annotation Lead / Data Operations Manager / ML Engineer]
[Company Name] builds [product / model / dataset] for [industry]. Our models are
only as good as the data behind them, so annotation is treated as core work here,
not as overflow. You will join a team of [number] annotators supporting
[number] active projects.
POSITION SUMMARY
The Data Annotator labels, tags, and classifies raw data so it can be used to
train and evaluate machine learning models. The work follows written annotation
guidelines, meets a defined quality bar, and flags the edge cases the guidelines
do not yet cover.
KEY RESPONSIBILITIES
•Label and classify [text / images / audio / video / sensor] data in
[annotation tool] according to the current project guidelines
•Meet the throughput and accuracy targets set for each project
•Flag ambiguous items and edge cases rather than guessing, and propose
guideline updates when a pattern repeats
•Take part in calibration rounds and resolve disagreements with other
annotators on the same batch
•Re-label or correct items returned by quality review
•Keep all project data confidential and handle it only inside approved systems
•Log hours worked accurately, including any required training and calibration
time
REQUIRED QUALIFICATIONS
•[High school diploma / associate degree / bachelor's degree: set your bar]
•Sustained attention to detail across repetitive work
•Comfortable reading and applying detailed written guidelines exactly
•Basic computer skills and the ability to learn a new annotation tool quickly
•[Language requirement, if any] at a native or fluent level
•Must sign a confidentiality agreement before receiving project data
CLASSIFICATION AND COMPLIANCE NOTE (read before posting)
Annotation is task-based production work. It does not meet the executive,
administrative, or professional duties tests, so a data annotator is almost
always non-exempt: paid hourly and owed overtime at one and a half times the
regular rate past 40 hours in a workweek. Paying above the federal salary
threshold does not change that, because the duties test still has to be met.
If annotators work from home, pay for all hours worked including training,
calibration, and required meetings, and use a timekeeping system rather than an
honor system. Do not classify a worker as an independent contractor when you set
the schedule, the tool, the guidelines, and the method. This is general
information, not legal advice.
EEO STATEMENT
[Company Name] is an equal opportunity employer and provides reasonable
accommodations for the essential functions of this role.
COMPENSATION AND HOW TO APPLY
Compensation: $_ per hour, [benefits summary], [schedule and hours]
To apply, email __ with your resume. Shortlisted candidates
complete a paid [length] annotation trial.
Template 2: Image and Video Annotator
For computer vision work: boxes, polygons, keypoints, segmentation, and object tracking across frames, with the pre-label review reality and the distressing content disclosure both stated plainly.
Image and Video Annotator Job Description (Computer Vision)
IMAGE AND VIDEO ANNOTATOR JOB DESCRIPTION (COMPUTER VISION)
Company: __ ([City, State] or Remote)
Reports to: [Computer Vision Lead / Annotation Manager]
For language data: entities, relations, intents, sentiment, and response rating, with inter-annotator agreement targets and clear rules for the personal information sitting inside raw text.
[Company Name] builds language models and text products for [support
automation / search / document processing / legal or financial text]. We are
hiring text annotators to label the training and evaluation data that decides
how those systems behave.
POSITION SUMMARY
The Text Annotator labels entities, relations, intents, sentiment, and content
categories in written data, rates and compares model outputs where the project
calls for it, and applies a written guideline consistently across thousands of
short judgments.
KEY RESPONSIBILITIES
•Label named entities, relations, intents, sentiment, and topic categories in
[documents / chat logs / support tickets / search queries]
•Rank or rate candidate model responses against the quality rubric
•Write short justifications for difficult judgments so reviewers can audit them
•Apply the guideline exactly, including its tie-break rules, and raise a
question when two rules genuinely conflict
•Meet inter-annotator agreement targets measured against a gold set
•Redact or escalate personal information encountered in raw text
•Take part in guideline reviews and calibration sessions
REQUIRED QUALIFICATIONS
•[Bachelor's degree in linguistics, English, or a related field / equivalent
demonstrated reading and writing ability: set your bar]
•Native or fluent [language], with strong reading comprehension
•Ability to apply a rubric consistently rather than by personal preference
•[Domain knowledge in your field: legal, medical, financial, technical]
preferred
•Comfort with repetitive judgment work and a fast reading pace
•Must sign a confidentiality agreement before receiving project data
CLASSIFICATION AND COMPLIANCE NOTE
Text annotation is non-exempt hourly work. A linguistics degree does not create
an exemption: the learned professional exemption requires that advanced knowledge
be the primary duty and be applied through consistent exercise of discretion and
judgment, and applying someone else's annotation rubric is not that. Raw text
frequently contains customer personal information, so define what annotators may
see, where they may see it, and what they must redact or escalate. If the work
includes reviewing harmful or explicit content for safety training, disclose it
in the posting. This is general information, not legal advice.
EEO STATEMENT
[Company Name] is an equal opportunity employer and provides reasonable
accommodations for the essential functions of this role.
COMPENSATION AND HOW TO APPLY
Compensation: $_ per hour, [benefits summary], [remote or hybrid]
To apply, email __ with your resume. Shortlisted candidates
complete a paid trial on a sample batch with a published rubric.
Template 4: Audio and Speech Annotator
For speech data, with the transcription convention, timing tolerance, and an honest note on paying by audio minute. If the work is purely clinical dictation, the medical transcriptionist templates are the better fit.
Audio and Speech Annotator Job Description
AUDIO AND SPEECH ANNOTATOR JOB DESCRIPTION
Company: __ ([City, State] or Remote)
Reports to: [Speech Data Lead / Annotation Manager]
systems for [industry]. We are hiring audio annotators to transcribe and label
recordings so our models learn real speech, including accents, overlaps, and
noise, rather than studio-clean audio.
POSITION SUMMARY
The Audio Annotator transcribes speech verbatim, marks speaker turns and
timestamps, and tags acoustic events and audio quality problems according to the
project transcription convention.
KEY RESPONSIBILITIES
•Transcribe recordings verbatim following our convention for fillers, false
starts, numbers, and unclear speech
•Segment audio and mark speaker turns, overlaps, and silence boundaries
•Tag acoustic events, background noise, language, dialect, and audio quality
•Time-align transcripts to the tolerance the project requires
•Flag recordings that are unusable or that contain sensitive disclosures
•Meet the words-per-hour and accuracy targets for each project
•Take part in convention reviews and calibration passes
REQUIRED QUALIFICATIONS
•[High school diploma / associate degree / bachelor's degree: set your bar]
•Typing speed of at least [number] words per minute with high accuracy
•Native or fluent [language] and comfort with a wide range of accents
•Good hearing and a quiet work environment; headphones [provided / required]
•Experience with transcription or captioning preferred
•Must sign a confidentiality agreement before receiving project data
CLASSIFICATION AND COMPLIANCE NOTE
Audio annotation and transcription are non-exempt hourly work. Two cautions on
pay. First, if you pay per audio minute or per file, that piece rate still has to
produce at least the applicable minimum wage for every hour actually worked, and
overtime is computed from the resulting regular rate, so you must still track
hours. Second, listening time, loading time, and tool problems are hours worked.
Recordings often capture personal, financial, or health information, so set
handling rules before the first file is assigned and confirm any consent and
recording obligations that apply in the states where the audio was captured.
This is general information, not legal advice.
EEO STATEMENT
[Company Name] is an equal opportunity employer and provides reasonable
accommodations for the essential functions of this role.
COMPENSATION AND HOW TO APPLY
Compensation: $_ per hour, [equipment provided], [schedule]
To apply, email __ with your resume and a sample transcript
of the [length] audio file we provide.
Template 5: Clinical and Medical Data Annotator
For medical annotation done by clinicians or certified coders, with license requirements, coding systems, and the business associate obligations that come with protected health information.
Clinical and Medical Data Annotator Job Description
CLINICAL AND MEDICAL DATA ANNOTATOR JOB DESCRIPTION
Company: __ ([City, State] or Remote)
Reports to: [Clinical Data Lead / Medical Director / Annotation Manager]
for [health systems / payers / device makers]. We are hiring annotators with
clinical background to label medical data, because a general annotator cannot
reliably distinguish [finding A] from [finding B].
POSITION SUMMARY
The Clinical Data Annotator labels medical text, images, or signals using
clinical judgment and our annotation guideline, documents the reasoning behind
difficult calls, and works within our HIPAA safeguards at every step.
KEY RESPONSIBILITIES
•Annotate [clinical notes / radiology images / pathology slides / ECG or other
signals] against the project guideline and class definitions
•Map findings to [ICD-10 / SNOMED CT / RxNorm / internal ontology] where the
project requires coded output
•Adjudicate disagreements between other annotators on the same case
•Document reasoning for edge cases and contribute to guideline revisions
•Work only inside approved, access-controlled systems and never move protected
health information to personal devices or accounts
•Report any suspected privacy incident immediately under our breach procedure
•Complete HIPAA and security training before the first case is assigned
REQUIRED QUALIFICATIONS
•[RN / LPN / MD / medical coder (CPC or CCS) / medical student / relevant
clinical degree: set the credential the project actually requires]
•[Number] years in [specialty] or equivalent coding experience
•Working knowledge of clinical terminology and [coding system]
•Current, unrestricted license where the role requires one
•Must complete HIPAA training and sign a confidentiality agreement before
receiving any data
CLASSIFICATION AND COMPLIANCE NOTE
Classification turns on the actual primary duty. A registered nurse or physician
performing genuine clinical judgment may qualify for the learned professional
exemption, which requires advanced knowledge in a field of science or learning
acquired through prolonged specialized instruction. Registered nurses paid hourly
are treated as non-exempt by the Department of Labor, and licensed practical
nurses generally do not meet the learned professional test. A clinical annotator
applying a fixed rubric to a queue is doing production work regardless of the
credential on the wall, so classify on the day, not the diploma. On privacy: if
you annotate protected health information for a covered entity you are almost
certainly a business associate, which requires a business associate agreement,
a workforce training program, and access controls. Use de-identified data
wherever the project allows it. This is general information, not legal advice.
EEO STATEMENT
[Company Name] is an equal opportunity employer and provides reasonable
accommodations for the essential functions of this role.
COMPENSATION AND HOW TO APPLY
Compensation: $_ per hour, [benefits summary], [project length]
To apply, email __ with your resume and license or
certification details.
Template 6: Annotation QA Reviewer and Lead Annotator
For the person who owns the guideline, the gold set, and the quality bar. Pair it with the quality analyst templates if you are also building a broader review function.
Annotation QA Reviewer and Lead Annotator Job Description
ANNOTATION QA REVIEWER / LEAD ANNOTATOR JOB DESCRIPTION
Company: __ ([City, State] or Remote)
Reports to: [Data Operations Manager / ML Lead]
Employment type: Full-time
FLSA status: Non-exempt unless the duties test is genuinely met (see note)
Compensation: $_ per [hour / year]
ABOUT THIS ROLE
[Company Name] runs [number] annotation projects across [modalities]. As the
team has grown, consistency has become the bottleneck rather than volume. We are
hiring a QA reviewer and lead annotator to own the guideline, the gold set, and
the quality bar.
POSITION SUMMARY
The Annotation QA Reviewer audits completed work against the guideline, measures
and reports quality, maintains the gold standard set, adjudicates disagreements,
and trains new annotators onto each project.
KEY RESPONSIBILITIES
•Audit sampled batches and report per-annotator and per-project quality scores
•Build and maintain the gold standard set used for calibration and onboarding
•Own the annotation guideline: write it, version it, and keep the change log
•Adjudicate disagreements and publish the resolution so the rule sticks
•Run calibration sessions and measure inter-annotator agreement
•Train and onboard new annotators onto each project
•Report quality trends to the ML team and recommend guideline or taxonomy
changes when the data says the rule is wrong
REQUIRED QUALIFICATIONS
•[Number] years of annotation experience across at least [number] projects
•Demonstrated ability to write a guideline others can follow without you
•Comfort with agreement metrics and basic quality reporting in
[spreadsheets / SQL / our dashboard]
•Clear written communication and the patience to teach
•[Domain expertise in your field] preferred
•Must sign a confidentiality agreement before receiving project data
CLASSIFICATION AND COMPLIANCE NOTE
Do not assume this role is exempt because it has lead in the title. A lead
annotator who mostly annotates and reviews is non-exempt. The executive exemption
requires managing a recognized department or subdivision, customarily directing
the work of at least two full-time employees, and having real authority or
influence over hiring and firing decisions. The administrative exemption requires
office work directly related to management or general business operations plus
the exercise of discretion and independent judgment on matters of significance.
Both also require payment on a salary basis at or above the applicable threshold,
and some states set a higher one. If the person spends most of the week in the
annotation queue, treat them as non-exempt and pay overtime. This is general
information, not legal advice.
EEO STATEMENT
[Company Name] is an equal opportunity employer and provides reasonable
accommodations for the essential functions of this role.
COMPENSATION AND HOW TO APPLY
Compensation: $_ per [hour / year], [benefits summary]
To apply, email __ with your resume and a short description
of a guideline you have written or improved.
Overtime, Contractors, and Data Access
A data annotator is a non-exempt hourly employee in essentially every case, owed overtime past 40 hours in a workweek. The exemptions from minimum wage and overtime under section 13(a)(1) each require a duties test that annotation does not meet, and job titles never determine exemption status.
Annotation is non-exempt work, with almost no exceptions
The white-collar exemptions cover executive, administrative, professional, computer, and outside sales employees, and each has a duties test that annotation does not meet. Labeling a queue against someone else’s guideline is production work: it is measured in throughput and accuracy, and it does not involve managing a business function or exercising discretion on matters of significance. That means hourly pay and overtime at one and a half times the regular rate past 40 hours in a workweek. Two traps follow. Paying an annotator a salary above the federal threshold does not create an exemption, because salary is only one half of the test. And the computer employee exemption does not apply either: it covers systems analysis, program design, and software engineering, not people who work inside a labeling tool. This is general information, not legal advice.
Contractor status is the most expensive shortcut here
Annotation is the kind of work companies instinctively push to contractors, because volume swings and the tasks look self-contained. The Department of Labor analyzes the economic reality of the whole relationship rather than the label on the agreement, and the factors that matter most in annotation almost all point toward employment: you set the guideline, the tool, the quality bar, the throughput target, and often the hours, and the worker has no meaningful opportunity for profit or loss and no independent business. Getting it wrong is not a paperwork problem. It exposes you to back wages, liquidated damages equal to those back wages, civil money penalties, and separate tax and state law consequences. If you genuinely need surge capacity, contract with a vendor that employs its own annotators rather than converting your own team to invoices. This is general information, not legal advice.
Annotators see your most sensitive data
Annotation puts raw production data in front of the newest and lowest-paid people on your team: support transcripts with customer names, dashcam footage with faces and plates, call recordings with card numbers read aloud, clinical notes. Treat access as the design problem it is. Sign confidentiality agreements before the first task, not at the end of week one. De-identify or redact upstream wherever the model does not need the identifier. Keep the work inside access-controlled systems and forbid personal devices, screenshots, and re-uploads to outside tools. If you touch protected health information on behalf of a covered entity you are almost certainly a business associate, which brings a business associate agreement and a workforce training obligation. Write the rules into the job description so the expectation is set before the offer.
Pay for training, calibration, and downtime
Onboarding an annotator onto a project takes real time: reading the guideline, working the gold set, sitting in calibration, and re-doing early batches that reviewers reject. All of that is hours worked and all of it is paid. The same is true of waiting time when the queue is empty, the tool is down, or a batch has not been uploaded, if the worker is not free to use the time for their own purposes. Piece rates are legal but they do not remove the obligation: whatever you pay per frame, per file, or per audio minute must still produce at least the applicable minimum wage for every hour worked, and overtime is computed from the regular rate that the piece rate produces. That is why you still need real timekeeping even on a per-task pay model. This is general information, not legal advice.
Most annotation is remote, and remote non-exempt work is where wage and hour claims are born. Training, calibration sessions, guideline reading, tool downtime, and waiting for a batch to load are all hours worked. A per-task pay model does not remove the obligation: whatever you pay per frame or per audio minute must still yield at least the applicable minimum wage for every hour actually worked, and overtime is calculated from the regular rate that piece rate produces. Track hours in a system, not on trust, and say so in the offer.
Screening for Annotation Quality
Screen annotators with a short paid trial scored against a gold standard set, because the trait that predicts success does not appear on a resume. Attention to detail is easy to claim, impossible to interview for, and obvious within one scored batch.
Build the gold set before you post the job. A few hundred pre-adjudicated items covering the ordinary cases and the nasty ones gives you an onboarding tool, a calibration tool, and a hiring screen at the same time. Then read the pattern of a candidate’s errors rather than the count.
Signal
How to test it
What a weak result looks like
Guideline discipline
Include items where the rule contradicts common sense
Labels by instinct and defends it as obvious
Boundary precision
Timed sample of occluded and truncated objects
Boxes drawn loosely to protect throughput
Consistency over time
Repeat five items at the start and end of the trial
Same item labeled two different ways
Edge case handling
Seed genuinely ambiguous items with no clean answer
Guesses silently instead of flagging
Agreement with peers
Overlap part of the batch with a current annotator
Systematic drift from the team on one class
Communication
Ask for a written note on the hardest three items
Cannot explain why a call was made
Domain knowledge
Cases a generalist would get confidently wrong
Confident and wrong, which is the worst pattern
Two operational notes. Pay for the trial: unpaid work samples are a bad look and a legal risk for non-exempt work. And keep the trial short, because a two-hour scored batch tells you almost everything a full day would, and long unpaid exercises drive away the careful candidates first.
Keep the scored trials somewhere you can compare side by side rather than scattered across a reply thread, because the whole point of a gold set is that two candidates get measured against the same answers. Applicant tracking is coming soon to FirstHR.
What to Pay a Data Annotator
There is no dedicated federal occupation code for data annotator, so no official median exists for the title. The honest approach is to benchmark against the nearest classifications the Bureau of Labor Statistics does track and adjust for the taxonomy complexity and domain knowledge your project actually requires.
Nearest Occupational Benchmarks
According to the Bureau of Labor Statistics Occupational Employment and Wage Statistics survey (May 2025), data entry keyers had a national median wage of $19.88 per hour ($41,340 per year), ranging from $15.00 at the 10th percentile to $28.26 at the 90th. Statistical assistants had a median of $24.20 per hour ($50,330 per year), reaching $39.42 at the 90th percentile (U.S. Bureau of Labor Statistics, OEWS).
Benchmark occupation
National median (BLS OEWS, May 2025)
How it maps to annotation work
Data entry keyers
$19.88 per hour / $41,340 per year
The floor anchor for simple, high-volume general annotation
Medical transcriptionists
$19.43 per hour / $40,410 per year
Closest match for audio transcription and time alignment
Inspectors, testers, sorters, samplers, weighers
$23.35 per hour / $48,570 per year
Structurally similar to QA review against a defined standard
Statistical assistants
$24.20 per hour / $50,330 per year
Best fit for structured judgment work and complex taxonomies
Data scientists
$57.80 per hour / $120,230 per year
Not an annotation benchmark; the ceiling this work supports
Three adjustments to those anchors. Taxonomy complexity pushes pay up, because a 200 class taxonomy with occlusion rules is not the same job as a binary flag. Domain credentials push it up sharply: a clinical annotator is paid as a nurse or a coder, not as an annotator. And a QA reviewer who owns the guideline and trains the team sits above the annotators they review. Publish a good-faith range where pay transparency laws require it, and remember that most annotation hiring is remote, so our guide to hiring remote employees covers the multi-state wage questions that follow.
Companies Using FirstHR Onboard 3x Faster
Join hundreds of small businesses who transformed their new hire experience.
Building an Annotation Team Without an HR Department
Annotation teams fail in three predictable places: the guideline was never written, hiring optimized for speed instead of agreement, and the team scaled in a burst that the paperwork could not follow. Each has a fix that costs hours rather than months.
The guideline lives in one person’s head, so quality collapses the moment you hire a second annotator
One annotator is always self-consistent, which hides the fact that no rule has ever been written down. Add a second person and the dataset splits in two, because every ambiguous case was resolved by habit rather than by a rule. Write the guideline before the posting goes out, not after the first bad batch. It needs class definitions, worked positive and negative examples, explicit tie-break rules, and a change log with dates. Then build a gold set of a few hundred pre-adjudicated items and use it for onboarding, for calibration, and for measuring agreement. This is the single highest-return hour of work in the whole annotation pipeline, and it is the one small teams skip.
You hire for speed, then discover that speed without agreement produces an unusable dataset
Throughput is easy to measure and easy to game, so it is what unmanaged annotation teams optimize. Accuracy against a gold set and agreement between annotators are the numbers that actually predict model performance. Screen for them at hire: run a short paid trial on a sample batch with a published rubric, then score the trial against your gold answers and look at the pattern of the mistakes rather than the raw count. Someone who is consistently wrong in the same direction has misread a rule and can be corrected in ten minutes. Someone who is randomly wrong is not reading carefully, and no amount of feedback fixes that. The paid trial costs a few hours of wages and saves months of contaminated data.
Annotation teams scale in bursts, and the paperwork arrives with them
Annotation headcount tracks project wins, so you hire one person in March and nine in June. Each one needs the same sequence: offer letter, confidentiality agreement, data-handling policy sign-off, security and privacy training, tool access, guideline walkthrough, gold set trial, and a payroll setup that reflects non-exempt hourly status. FirstHR was built for that kind of burst. The onboarding wizard runs the identical sequence for every new annotator, built-in e-signature handles the confidentiality agreement and the policy acknowledgments, training modules cover the security and privacy walkthrough before the first task is assigned, and document management keeps the signed agreements and any certifications on the employee profile with renewal dates attached. Applicant tracking is coming soon to FirstHR. Note that FirstHR is an onboarding and HR platform, not a payroll provider.
Once the offer is signed, the work becomes a repeatable sequence, and a shared onboarding checklist is what keeps the confidentiality agreement, the security training, and the tool access from being remembered three weeks late. For the annotation side specifically, treat the guideline walkthrough and the gold set trial as onboarding steps with owners and dates, not as something the new hire picks up.
Key Takeaways
A data annotator applies a written guideline rather than personal judgment, so the guideline, the gold set, and the agreement metric have to exist before the posting does.
Data annotator is one title covering at least four jobs, and image, text, audio, and clinical annotation screen for different traits and carry different pay benchmarks.
Annotation is non-exempt hourly work in essentially every case, because it fails the duties test for every white-collar exemption regardless of what you pay.
Contractor agreements are the most expensive shortcut in annotation hiring, since you control the guideline, the tool, the quality bar, and usually the hours.
No dedicated federal wage code exists for the title, so benchmark to nearby occupations: $19.88 per hour for data entry keyers up to $24.20 for statistical assistants (BLS OEWS, May 2025).
Screen with a short paid trial scored against your gold set, and read the pattern of the errors rather than the raw count.
Annotation headcount arrives in bursts, and confidentiality agreements, data-handling sign-offs, and security training arrive with it. FirstHR runs the same onboarding sequence for every new annotator, with e-signature for agreements and policy acknowledgments, training modules delivered before the first task is assigned, and document storage that keeps every signed form on the employee profile. Applicant tracking is coming soon to FirstHR.
Frequently Asked Questions
What does a data annotator do?
A data annotator labels raw data so a machine learning model can learn from it. In practice that means drawing bounding boxes and segmentation masks on images, tagging entities and intents in text, transcribing and time-aligning audio, tracking objects across video frames, or rating model outputs against a rubric. The work is governed by a written annotation guideline that defines every class, every edge case, and every tie-break rule, and the annotator’s job is to apply it consistently rather than to use personal judgment. On mature projects a growing share of the day is reviewing and correcting pre-labels a model produced instead of labeling from scratch. Good annotators also flag ambiguous cases rather than guessing, because a guessed label silently teaches the model the wrong thing.
Is a data annotator exempt or non-exempt under the FLSA?
Non-exempt, in essentially every case. The white-collar exemptions each carry a duties test, and annotation does not meet any of them: it is task-based production work measured by throughput and accuracy, not the management of a business function or the exercise of discretion on matters of significance. That means hourly pay and overtime at one and a half times the regular rate for hours past 40 in a workweek. Paying an annotator a salary above the federal threshold does not create an exemption, because the salary test is only half the analysis. The computer employee exemption does not help either: it covers systems analysis, program design, and software engineering, not people working inside a labeling tool. A lead annotator who still spends most of the week in the queue is also non-exempt, whatever the title says. This is general information, not legal advice.
Can I hire data annotators as independent contractors?
Usually not, and this is the most common expensive mistake in annotation hiring. The Department of Labor looks at the economic reality of the whole relationship rather than the label on the agreement, and the facts of annotation work almost all point toward employment: you set the guideline, the tool, the quality bar, the throughput target, and frequently the hours, while the worker has no real opportunity for profit or loss and no independent business serving other clients. Misclassification exposes you to back wages, liquidated damages equal to those back wages, civil money penalties, and separate tax and state law consequences. If you genuinely need surge capacity for a project spike, contract with a vendor that employs its own annotators rather than converting your own team to invoices. This is general information, not legal advice.
How much should I pay a data annotator?
There is no dedicated federal occupation code for data annotator, so benchmark against the nearest classifications instead of inventing a number. According to the Bureau of Labor Statistics Occupational Employment and Wage Statistics survey (May 2025), data entry keyers had a national median wage of $19.88 per hour, or $41,340 a year, with the 10th percentile at $15.00 and the 90th at $28.26. Statistical assistants, a closer match for annotators doing structured judgment work, had a median of $24.20 per hour or $50,330 a year, reaching $39.42 at the 90th percentile. Medical transcriptionists sat at $19.43 per hour and quality inspectors, testers, and sorters at $23.35. For general annotation the data entry range is the honest anchor; for credentialed clinical or legal annotation, benchmark to that profession instead, because it costs several times more. Publish a good-faith range where pay transparency laws apply.
What qualifications should a data annotator have?
For general annotation, far less formal education than most postings ask for and far more discipline than most postings screen for. The real requirements are sustained attention across repetitive work, the ability to read a detailed guideline and apply it exactly rather than by instinct, and the honesty to flag an ambiguous case instead of guessing. Native or fluent language ability matters for text and audio work. Beyond that, requirements are modality-specific: precision and stamina for image and video, reading comprehension and rubric discipline for text, typing speed and accent tolerance for audio. Only specialist annotation genuinely needs a credential, and there the credential is a clinical license, a coding certification, or domain experience rather than anything about annotation itself. Screen with a short paid trial on a sample batch, because the trait you care about does not appear on a resume.
What is the difference between a data annotator and a data entry clerk?
The difference is that a data entry clerk transcribes information that already exists while a data annotator creates a judgment that did not exist before. A clerk moves a value from an invoice into a field, and the correct answer is knowable by looking at the source document. An annotator decides whether a blurred shape at the edge of a frame counts as a pedestrian, or whether a support ticket expresses frustration or a neutral request, and the correct answer comes from a guideline that a human wrote and that will keep changing. That is why annotation needs a gold set, agreement metrics, and a guideline owner, and data entry mostly needs a verification pass. Pay reflects it too: annotation benchmarks above general data entry once the taxonomy gets complicated, and well above it once domain expertise is involved.
How do I measure data annotation quality?
Use two numbers together: accuracy against a gold standard set and agreement between annotators working the same items. The gold set is a few hundred items you have pre-adjudicated and hold fixed, used for onboarding, for periodic calibration, and for scoring new hires during a paid trial. Agreement tells you whether the guideline is clear, because when two careful annotators disagree the usual cause is an ambiguous rule rather than a careless person. Sample completed batches rather than reviewing everything, publish per-project and per-annotator scores, and treat a drop as a signal to re-read the guideline before blaming the team. Look at the pattern of mistakes, not just the count: consistent errors in one direction mean a misread rule and take minutes to fix, while random errors mean the work is not being read carefully at all.
