Per-Label vs Hourly vs Managed: Annotation Pricing Models Compared

Why Pricing Model Choice Matters More Than Rate

When procurement teams evaluate annotation vendors, they typically focus on the headline rate — AUD $0.12 per label vs AUD $0.18 per label. The rate matters, but the pricing model determines the total cost more than any quoted rate. A per-label contract at AUD $0.12 with 25% rework costs more than a managed-service contract at three times the apparent unit price.

A 2024 Cognilytica survey of enterprise AI teams found that annotation budget overruns averaged 34% above initial estimates — and the primary driver in 61% of cases was a mismatch between pricing model and project characteristics, not a bad rate. The wrong model creates misaligned incentives between buyer and vendor that compound over the project lifecycle.

The three models on the market today — per-label, hourly, and managed service — each transfer risk differently. Understanding which party bears throughput risk, quality risk, and scope-change risk under each model is the foundation for making the right choice. Our annotation pricing page gives transparent rate ranges for each task type across all three structures.

Per-Label Pricing: What It Is and When It Works

Per-label (also called per-unit or per-task) pricing charges a fixed rate for each completed annotation unit. The unit definition varies by task: a bounding box, a sentiment label, a named entity tag, a transcribed word, a semantic segmentation polygon, or an image-level classification. Rates vary significantly by task complexity and annotator expertise required.

Typical Australian market rates for common task types:

• Simple binary text classification: AUD $0.02–$0.06 per label
• Multi-class text classification (5–10 classes): AUD $0.06–$0.15 per label
• Image bounding boxes: AUD $0.10–$0.40 per box
• Semantic segmentation polygons: AUD $0.50–$2.00 per image
• Speech transcription: AUD $0.80–$2.50 per audio minute
• Medical image annotation (board-reviewed): AUD $1.50–$8.00 per image

Per-label pricing works best when task complexity is uniform, scope is fixed, and quality can be measured objectively per unit. It incentivises vendors to be efficient — which is a benefit when your guidelines are clear and a problem when your guidelines have ambiguity that vendors will resolve by choosing the fastest interpretation.

The primary risk with per-label pricing is rework. If 15–25% of labels fail QA and require reannotation, your effective cost-per-label can be 30–45% higher than the quoted rate. Negotiate rework inclusion — a vendor who quotes AUD $0.15/label with rework included is often cheaper than a vendor quoting AUD $0.11/label with rework billed at full rate.

Hourly Pricing: When Time-Based Billing Is the Right Call

Hourly pricing bills for annotator time rather than label output. It suits tasks where annotation time is highly variable across records — typically where annotator judgement, domain research, or multi-step reasoning is required per item.

The practical cases where hourly pricing makes sense:

• Open-ended content moderation with escalation paths where borderline cases require multi-annotator review
• Legal or medical expert annotation where annotators must research before labelling
• Complex NER with dense documents where annotation time per record varies 5–10x based on document length and entity density
• RLHF preference annotation where evaluators must read and assess long response pairs
• Pilot projects with undefined scope where per-label pricing would require speculative unit estimates

Hourly rates for annotation work in Australia typically range from AUD $25–$45/hour for general-domain annotators, AUD $55–$95/hour for expert annotators (clinical, legal, financial), and AUD $80–$150/hour for bilingual expert annotators in low-resource languages.

The main risk with hourly pricing is throughput uncertainty. Without output benchmarks per annotator per hour, there is no natural ceiling on total cost. Best practice is to require vendors to provide throughput benchmarks from comparable tasks upfront, and to build output-per-hour minimums into the contract. A vendor who refuses to commit to minimum throughput on an hourly contract is transferring all throughput risk to you.

Managed Service Pricing: Paying for Outcomes, Not Labour

Managed-service pricing bundles everything — annotators, QA workflows, project management, tooling, reporting, and compliance documentation — into a single deliverable fee. The buyer pays for a QA-passed, delivered dataset, not for the inputs that produced it.

This model suits enterprise buyers who:

• Lack internal annotation infrastructure (tooling, QA processes, annotator management)
• Need compliance assurances (SOC 2, ISO 27001, HIPAA) that require vendor-side controls
• Are running multi-phase or ongoing annotation programmes where renegotiating per-label rates across phases creates unnecessary overhead
• Want accountability for dataset quality, not just labour delivery

Managed-service contracts are typically structured as project-based (fixed fee for a defined dataset) or retainer-based (monthly fee for a defined annotation throughput commitment). Project-based pricing shifts scope-change risk to the buyer; retainer-based pricing shifts throughput-variation risk to the vendor.

The effective per-label cost under managed service is typically 2–4x the per-label rate for comparable raw annotation. That premium buys project management, QA, and accountability — and for buyers who have already experienced a per-label project where internal QA overhead consumed more time than the vendor saved, the premium is usually worth it. For a transparent breakdown of what managed annotation projects cost at different scales, see our annotation pricing guide.

Project Case Study: Pricing Model Switch That Saved AUD $210K

An Australian insurance company was annotating 180,000 claim documents for a fraud detection model. They had contracted a per-label vendor at AUD $0.08 per entity tag, with an estimated 4.2 million entity annotations across the dataset.

At the six-week mark, the project had produced 800,000 labels — roughly 19% of the target — at a cost of AUD $64,000. Internal QA review identified an 18% rejection rate, meaning the effective cost after rework was tracking at AUD $0.097/label rather than $0.08. More critically, the QA work was consuming 1.8 FTE of internal ML engineering time per week.

The company renegotiated to a managed-service structure at a fixed project fee of AUD $380,000 for the full delivered dataset. Apparent per-label cost: AUD $0.091 — higher than the original quoted rate. Actual outcome:

• Delivered dataset QA rejection rate: 2.1% (vs 18% under per-label)
• Internal QA overhead: 0.2 FTE/week (down from 1.8 FTE/week)
• Project completed in 14 weeks (vs 28-week projection under the per-label trajectory)
• Total internal QA cost saving over project: approximately AUD $130,000 in engineering time
• Rework cost eliminated: approximately AUD $80,000

The managed-service contract cost AUD $316,000 more at the headline rate but saved approximately AUD $210,000 in total cost of delivery. The effective saving only became visible once internal QA time was costed at market rate — which most procurement processes omit from vendor comparisons.

The Decision Framework: Which Model Fits Your Project

The right pricing model depends on four variables: task complexity uniformity, internal QA capacity, compliance requirements, and your team's annotation infrastructure maturity.

Choose per-label pricing when: task complexity is uniform across records, your guidelines are complete and unambiguous, you have internal QA capacity to review output, and scope is defined before work begins. Per-label pricing at clear task definition is the lowest cost model when all of these hold.

Choose hourly pricing when: annotation time varies significantly across records, you are running a pilot or definition phase where scope is still emerging, or you need expert annotators who are paid for judgement rather than throughput. Require throughput benchmarks in the contract.

Choose managed service when: your team lacks annotation infrastructure, compliance documentation is required, quality accountability matters more than headline rate, or you are running a programme large enough that internal QA overhead is a material project cost. Model the internal overhead before comparing managed-service quotes to per-label rates.

For complex or ambiguous tasks — medical image annotation, legal document NER, Arabic dialect NLP — managed service is the default recommendation regardless of team maturity, because quality failure cost in these domains typically exceeds any rate difference. See our overview of custom annotation services for how task-specific pricing is structured for non-standard workflows.

For a deeper look at what drives annotation costs by task type, see our honest 2026 annotation pricing breakdown and the true cost of cheap annotation vendors analysis.

Negotiating Annotation Contracts: Key Clauses by Pricing Model

Regardless of the model, three contract clauses determine real-world cost more than the quoted rate:

Rework policy. For per-label contracts, clarify whether rework is billed at full rate, discounted rate, or included in the original price. For managed service, define the rejection threshold — labels failing the agreed QA standard should be reworked at vendor cost, with no buyer billing.

Scope change mechanism. Define what constitutes a scope change and how it is priced. For per-label, new label classes or task additions typically trigger renegotiation. For managed service, specify whether scope changes are priced at a day-rate, a per-label add-on, or absorbed within an agreed variance band (typically ±10%).

Quality threshold and remedy. Define the minimum acceptable quality level (typically 95%+ accuracy or κ ≥ 0.80 for classification tasks), the measurement method, and the remedy if the threshold is missed — whether that is vendor-funded rework, a rate reduction, or contract termination rights. A contract without an explicit quality clause puts all quality risk on the buyer. Our annotation project scoping checklist covers all the contract and scoping questions in detail.

Hybrid Models: Combining Pricing Structures Across Project Phases

Most large annotation programmes benefit from a hybrid approach: hourly pricing for the definition and pilot phase, per-label or managed service for production. The pilot phase should explicitly inform the pricing model for the production phase — if the pilot reveals high annotation time variance, hourly or managed service is the right choice; if annotation time is uniform and guidelines hold up well, per-label pricing offers the best value.

A common structure for enterprise programmes:

• Phase 1 (definition + pilot): Hourly, 40–80 hours, goal is to produce guidelines, calibration set, and throughput benchmark
• Phase 2 (production batch 1): Per-label with rework included, using Phase 1 benchmarks to validate vendor estimates
• Phase 3+ (ongoing): Managed service retainer for continuous annotation programmes, or per-label for discrete batch projects

This structure keeps the high-uncertainty phase (definition) under hourly pricing where cost scales with actual time spent, and moves to outcome-based pricing once scope and throughput are understood. For teams building or scaling their annotation function, our guide to annotation team management and scaling covers how to structure these multi-phase programmes operationally.

FAQ