Short answer. In-house annotation frequently looks cheaper because most in-house budgets count only annotator wages. Add recruiting, training, management, QA, tooling, infrastructure and idle capacity and the gap narrows or reverses. CVAT's published 2024 case study is instructive precisely because it does not flatter outsourcing: for an illustrative 100,000-image, 2.3-million-object project it estimated $122,220 in-house before software licences against roughly $225,400 outsourced. Outsourcing is usually justified by speed, flexibility and avoided operating overhead — not by a guaranteed lower direct price. Decide on total operating cost and on whether annotation should become a permanent internal capability.
This is the one annotation decision that is genuinely a build-versus-buy question, and it deserves to be argued honestly in both directions. There are real programmes for which in-house is correct and real programmes for which it is a two-year detour. The determining factors are utilisation, permanence and data sensitivity — not price per label.
The published cost example, read carefully
CVAT published a 2024 case study for an illustrative 100,000-image project averaging 23 objects per image, producing 2.3 million annotation objects. Its in-house scenario estimated $122,220 before software-licence costs. Its outsourcing scenario estimated approximately $225,400 for the same object count.
Three observations, in order of importance:
- The outsourced figure is higher. Anyone selling outsourcing on direct price alone is arguing against a published example. The honest case for outsourcing is elsewhere.
- The in-house figure excludes software. It also, in the way most such models do, assumes the team exists and is fully utilised for the duration.
- It is one illustrative scenario. A 2.3-million-object single-modality project is not a multi-year, multi-language, multi-modality programme, and the economics diverge sharply as those variables enter.
What each side actually owns
| Cost category | In-house | Outsourced managed service |
|---|---|---|
| Annotator recruitment | Buyer owns | Vendor owns |
| Training and calibration | Buyer owns | Vendor manages |
| Workforce utilisation | Buyer carries idle capacity | Vendor absorbs more capacity planning |
| Annotation management | Buyer hires managers | Included or managed, depending on contract |
| QA and validation | Buyer builds the process | Provider supplies the process |
| Software and infrastructure | Buyer acquires and maintains | May be included or priced separately |
| Flexibility to ramp down | Often difficult | Usually easier contractually |
| Direct unit rate | Can be lower | Often includes service overhead |
The row that decides most cases is workforce utilisation. An internal annotation team is a fixed cost against a variable workload. If your ML roadmap produces labelling demand in bursts — and most research-driven roadmaps do — you are paying for the troughs.
The hidden cost problem
TELUS Digital has argued publicly that a simple labour-hours calculation misses hidden setup and engineering costs in data labelling. Appen has published a similar argument specifically about tooling: building an internal annotation tool introduces development, maintenance and opportunity costs that rarely appear in the original business case.
Both points are strongest where annotation is not the company's core business. The engineering hours spent building a labelling interface are hours not spent on the model, and that opportunity cost does not appear on any line item.
A more complete in-house model includes:
- Recruiting and onboarding, per annotator, including the ones who leave in month two
- Team leadership and project management headcount
- QA design, gold-set construction and ongoing calibration
- Annotation tooling: licence, integration, or build plus maintenance
- Storage, transfer and compute
- Security controls, access management and audit
- Workflow engineering as the taxonomy evolves
- Training refreshes after every guideline change
- Employee turnover, and the learning curve paid again each time
- Unused capacity during model-training and evaluation cycles
When in-house genuinely makes sense
Five conditions. If three or more hold, build.
- The annotation workload is permanent, predictable and strategically core. Not "we will always need labels" but "we will need roughly this much, every month, for years".
- Sensitive data cannot leave a tightly controlled internal environment, and no vendor's residency or facility controls satisfy the requirement.
- The organisation already has annotation managers, QA specialists and suitable tooling. Most of the cost of building is building; if it is already built, the arithmetic changes.
- Specialist employees must annotate as part of their normal roles — clinicians labelling clinical data, engineers labelling engineering data. Here the labour is not substitutable at any price.
- The team can be kept highly utilised over a long period. Utilisation below roughly 70% erodes the direct-cost advantage that motivated the build.
When outsourcing is clearly the better economics
- Demand is variable, seasonal or project-driven.
- The programme needs rapid scale that internal hiring cannot match.
- Multiple languages are required, particularly outside the top ten.
- Several modalities are in scope, each needing different tooling and different reviewer skills.
- Annotation is explicitly not something the organisation wants to become good at.
- The cost of being late exceeds the cost of the premium.
The hybrid most mature programmes end up with
The binary framing is usually wrong at scale. The common mature structure is:
- A small internal team owning the taxonomy, the gold set, adjudication and acceptance. This is the part that must not be outsourced, because it is where your definition of correct lives.
- External capacity for volume production, calibrated against the internal gold set.
- A specialist vendor, retained, for the narrow high-risk workstream where a controlled benchmark shows a meaningful advantage.
This keeps the strategic capability in-house and the fixed cost out. It also gives you a credible answer to the question that ends most in-house business cases: what happens when the person who understood the ontology leaves?
How Lifewood approaches this
Lifewood's model is designed for the buy side of this decision: the buyer purchases a managed annotation capability rather than recruiting and training an annotation organisation.
Three things do the work. The quality framework is built in rather than assembled by the client — trained annotators, senior review, automated consistency checks and client feedback loops, against a 95%+ accuracy SLA, with below-threshold batches reworked at Lifewood's cost. A distributed workforce across 40+ delivery centres in 30+ countries absorbs volume changes that an internal team would carry as idle capacity or overtime. And coverage across 50+ languages and multiple modalities means capability can be redeployed across changing workloads rather than hired for each one separately.
Lifewood has operated in AI data since 2004, with 56,788 registered contributors and 414,120 training hours delivered in 2025 — figures that describe the operating layer a buyer would otherwise have to build. Whether building it is the right decision still depends on the five conditions above.
Sources and further reading
- CVAT published in-house and outsourcing cost case studies for a 100,000-image / 2.3-million-object project at cvat.ai and cvat.ai.
- TELUS Digital decision framework for data labelling strategy at telusdigital.com.
- Appen on the build-or-buy decision for annotation tooling at appen.com.
- Lifewood delivery figures (50+ languages, 40+ delivery centres across 30+ countries, 95%+ accuracy SLA, 56,788 registered contributors, 414,120 training hours in 2025) published on lifewood.com.