Building an in-house team or entrusting the work to a provider is often decided by setting an hourly rate against a salary. That comparison ignores most of what each option commits, and it proves false in both directions depending on the nature of the project.
This article sets out that choice. It extends the article on labeling modalities.
What in-house work brings
Five advantages accompany an in-house team in data labeling. Business knowledge is built and then stays inside the company. Adjudications are given in minutes rather than in days. The data never leaves the organisation’s perimeter. The reference evolves at the very pace of understanding the need. And the same team can serve several successive projects. One important practical consequence follows. The first advantage weighs more than the others over time, an in-house team accumulating a knowledge of the field no provider reconstitutes quickly.What in-house work costs
Five charges accompany an in-house data labeling team. Recruitment as well as the operators’ initial training. Supervising the work as well as managing the schedules. Tool licences as well as the storage infrastructure. The cost of periods with no production to absorb. And the difficulty of varying capacity quickly. One important observation follows for a data labeling project. The fourth charge often decides the calculation, an in-house team being paid every month whereas annotation needs arrive in waves, which makes this option expensive as soon as the volume becomes irregular.What outsourcing brings
Five advantages accompany the use of a data labeling provider. Capacity adjusts to the volume genuinely needed each month. Recruitment and training do not weigh on the client. Experience gained on other projects benefits their own. The tools as well as the infrastructure are already in place. And the cost becomes a variable expense rather than a fixed charge. One practical consequence follows. The first advantage alone justifies this option on irregular projects, a team forming and shrinking without the client bearing the quiet periods.What outsourcing costs
Five counterparts accompany this second option. The reference must be written since it no longer passes on orally. Adjudications go through a circuit and take time. The data leaves the organisation’s perimeter. Business knowledge builds up at the provider’s. And the provider’s margin is added to the production cost. One observation follows. The first counterpart often proves beneficial, the obligation to write down what stayed implicit improving the corpus’s consistency regardless of who produces it.What the mixed solution permits
Five arrangements are met in practice. Scoping kept in-house and production at the provider’s. The routine volume outside and the sensitive cases in-house. A small in-house team absorbing peaks through outside capacity. Quality control kept entirely in-house on outsourced work. And a provider to start with before a progressive move in-house. One important observation follows for a data labeling project. The first arrangement gives the best results, business expertise staying with whoever holds it while production capacity stays adjustable, which avoids having to choose between the two decisive advantages.What volume and duration change
Five situations direct this choice differently. A low and purely occasional volume clearly favours outsourcing. A high and perfectly regular volume brings the two options closer. A permanent need over several years favours in-house work. A very irregular volume makes an in-house team hard to justify. And a fast rise in load exceeds in-house recruitment capacity. One important practical consequence follows for a data labeling project. The last situation is met more often than foreseen, a project moving from a few thousand to a few hundred thousand elements within a quarter, which makes the mixed solution more frequent than the exclusive choice.What comparing costs must include
Five items are often missing from the in-house calculation. Social charges as well as the full cost of a position. Supervision time taken from other functions. Tool licences and data storage. Periods of under-activity paid for regardless. And the recruitment delay during which nothing is produced. One important practical consequence follows for a data labeling project. The second item almost always escapes the calculation, supervising an annotation team occupying a significant share of a project manager whose salary appears elsewhere, which distorts the comparison without anyone having intended it.What confidentiality imposes here
Five constraints sometimes limit outsourcing. A regulatory prohibition on leaving the territory. An industrial secret a contract does not always suffice to protect. Personal data requiring a regulated treatment. A certification requirement weighing on the provider. And data access limited to controlled workstations. One important observation follows for a data labeling project. The first constraint settles matters with no discussion possible, no economic consideration making lawful a transfer the regulation forbids, which places this check before any comparison of costs.What arbitration requires of the client
Five conditions make an outsourcing practicable. A clearly identified person able to settle a business case. A real availability, measured in hours and not in intentions. A deliberately short response time on the cases reported. A decision recorded in writing rather than an oral answer. And a continuity of that function throughout the project. One important observation follows for a data labeling project. The second condition is missing more often than the first, a designated business expert monopolised by their own priorities letting cases pile up, which blocks a production the provider cannot unblock alone.What changing provider costs
Five charges accompany a transfer to another supplier. Passing on the reference as well as its already settled cases. The full training of a new team in the field concerned. A period during which consistency with the existing work stays uncertain. A markedly reinforced control during the very first batches. And a possible rework of the data produced in the meantime. One important practical consequence follows for a data labeling project. The first charge depends entirely on an earlier decision, a reference written and handed over making the transfer practicable within weeks whereas knowledge kept oral makes it nearly impossible, which makes that document the real measure of dependence on a provider.Three errors proper to this choice
Three data labeling defects belong to this decision. An hourly rate compared with a gross salary and nothing else. An in-house move decided on a need that will prove occasional. And an outsourcing launched with no available arbiter on the client side. Those three defects are committed before any start, they are observed only after a few months, and their common point is to bear on the duration of the need rather than on its volume.What this subject has that is particular
Four traits distinguish this question from the course’s others. It is settled on hypotheses rather than on measurements. It commits the company well beyond the current project. It is hard to revise once the team is assembled. And it mixes financial and regulatory considerations. One important observation follows for a data labeling project. The first trait explains the frequency of poor choices, the future regularity of the need staying an estimate at the moment of deciding, which makes a reversible option preferable to an optimal one grounded in an uncertain forecast.What the first project must establish
Four observations direct the definitive decision. The volume genuinely consumed against the volume forecast. The frequency of new cases once the reference has stabilised. The arbitration time effectively mobilised on the client side. And the regularity observed in demand across several months. One important practical consequence follows for a data labeling project. The third observation often surprises and weighs on the choice, an outsourced project consuming more business-expert time than foreseen, which narrows the gap with an in-house team once believed to cost more in supervision.What this chapter shares with the previous ones
Three findings recur from the start of this course. What is not written stays in the head of whoever knows it. A decision taken before production determines what it will reach. And the visible cost covers only part of the real expense. One important observation follows for a data labeling project. The first finding takes on a new reach here, knowledge left oral making the organisation dependent on the people who hold it, which holds for an in-house team exactly as for a provider.What the provider’s location changes
Five parameters vary with the location adopted. The hourly cost, which explains most of the observed rate gaps. The time difference, which lengthens the arbitration circuits. The working language, determining on textual data. The regulatory framework applying to the data handled. And team stability, which varies greatly from country to country. One important practical consequence follows for a data labeling project. The last parameter weighs more than the first over time, a high turnover forcing constant retraining of operators and the dispersion that accompanies each arrival, which cancels part of the gain the hourly cost had promised.What the contract must provide for here
Five clauses deserve writing down before the start. Ownership of the reference as enriched during the project. The response time expected on the business cases reported. The arrangements for returning the data at the mission’s end. The possibility of varying the volume upwards as well as downwards. And the conditions of a transfer to another provider. One important observation follows for a data labeling project. The first clause decides future dependence, a reference staying the provider’s property locking the client in far more surely than any exclusivity clause, which justifies treating it as a question of intellectual property rather than as a formality.What automation changes about this calculation
Four effects modify the balance between the two options. The manual volume falls on the corpus’s common cases. The remaining share becomes harder and more expert. Control takes a proportionally greater place. And production capacity counts for less than judgement. One important observation follows for a data labeling project. Those four effects progressively favour a small and competent team over a large capacity, which brings in-house work within reach of more situations than it covered when raw volume dominated the expense.What this choice does not resolve
Four difficulties persist whichever option is adopted. The reference must be written in both cases. Omission stays invisible in both cases. Business arbitration falls to the client in both cases. And documentation conditions future value in both cases. One important practical consequence follows for a data labeling project. Those four difficulties put this choice’s importance in perspective, a rigorous organisation obtaining good corpora by either route whereas an approximate one fails equally by both, which shifts attention towards the decisions handled in the preceding chapters.What a rise in load requires
Four delays condition a fast increase in capacity. Recruitment, which is counted in weeks including in-house. Training in the reference, highly variable by modality. The period during which dispersion stays high after an arrival. And the reinforcement of control during that period. One important observation follows for a data labeling project. The third delay is almost always neglected in schedules, a doubled team not immediately producing double since newcomers annotate slowly and diverge more, which spreads out a capacity gain the calculation assumed to be immediate.What the provider brings here
Four contributions distinguish a data labeling partner from a mere supplier. Help in writing the reference rather than plain execution. An adjustable capacity announced with its real lead times. A control arrangement described with its exact coverage. And a return of the enriched reference to the end client. One important practical consequence follows. The last contribution runs against the provider’s apparent interest, a reference handed to the client making a change of supplier easier, whereas it designates precisely what separates a service from a lock-in and grounds the lasting relationship the client was looking for.What this chapter presupposes of the provider
Four practices distinguish a partner from a mere executant. Saying when going in-house would serve the client better. Handing over the enriched reference rather than keeping it. Announcing a realistic lead time for a rise in load. And accepting a volume reduction without disproportionate penalty. One important practical consequence follows. The first practice runs squarely against immediate interest, a provider recommending that a client build their own team giving up a contract, whereas that advice is often worth several recommendations and lastingly separates an advisory relationship from a supply one.Approaching this choice
Five questions scope this choice in data labeling. Is the need regular. Waves cost dearly in-house. Is the field specific. Expertise builds slowly. Can the data leave. Some of it cannot. Must capacity vary. In-house adjusts badly. And is an arbiter available. Without one, outsourcing stalls. Those five answers determine the suitable option. Asking them before comparing rates avoids a calculation that is accurate on false assumptions.The question that frames this choice
One question determines the suitable option in data labeling. Will the annotation need be permanent or tied to a project. A permanent need, a chain that retrains regularly or a field that evolves, justifies an in-house team whose expertise accumulates and pays for itself over time. A need tied to a project, a corpus to build once for a given model, sits badly with a fixed charge that will outlive the need that created it. That question belongs to the company’s trajectory and not to the immediate volume, it is asked before any comparison of rates, and it separates two calculations neither of which means anything for the other situation.Three decisions before settling
Three decisions commit this choice in data labeling. Establishing whether the need will be regular or come in waves over two years. Checking the regulatory constraints weighing on the data. And designating the person who will settle the new cases. Those three decisions cost a few meetings, they precede any consultation of providers, and their absence leads to comparing options on criteria that are not the right ones.Three checks before comparing
Three controls make the comparison honest. The full cost of an in-house position, supervision included. The foreseeable regularity of volume over the next two years. And the existence of a regulatory constraint on the data. Those three controls are conducted in a few hours, they require no technical competence, and their absence produces a comparison that is arithmetically right and practically wrong.Three questions before consulting
Three questions precede any consultation of providers. Will the need outlive the project that gives rise to it. Who will settle the business cases and with what availability. And can the data lawfully leave the company. Those three questions are asked internally before any tender, they require no technical expertise, and their absence leads to consulting on a need nobody has yet defined.What this chapter teaches
One cross-cutting observation deserves closing this examination. The choice is settled by the regularity of the need rather than by the rate. Three findings compose it. An in-house team is paid every month, the needs arrive in waves. Business expertise builds slowly and stays where it is built. And outsourcing forces the writing of what stayed implicit. That finding extends the preceding chapters, comparing costs first presupposing knowing what each option covers.What this chapter leaves to the next
Both options rest on the same human competences. Three questions stay open in data labeling. What a trained annotator can do that a beginner cannot. How long that competence takes to build. And how a team is structured around this work. Those three questions belong to the annotator’s role, which constitutes the subject of the following chapter.Why recommending in-house work sometimes wins the account
One point about how this advice lands belongs at the close of this chapter. A provider who says the client should build their own team loses that contract. Three reasons follow in data labeling. The client came looking for a supplier, not for an opinion. A competitor will happily take the work without the caveat. And nothing obliges anyone to volunteer that the need is permanent. One important practical consequence follows for a data labeling provider. Saying it anyway pays over a longer horizon, since a client advised into a badly fitting arrangement discovers it within a year and remembers who sold it to them, whereas one advised correctly returns with the projects that genuinely suit outsourcing.What the mixed arrangement asks of both sides
Four commitments make a shared setup work. The client keeps the reference and the arbitration. The provider keeps the capacity and the throughput. Both agree on how a new case travels between them. And neither treats the boundary as fixed for the project’s whole life. One important observation follows for a data labeling project. The last commitment matters most as the work evolves, the boundary between what stays in-house and what goes out moving as the reference stabilises, which makes an arrangement reviewed twice a year more durable than one settled at the outset.Common mistakes
These failures recur often enough that naming them is usually enough to avoid them.- Comparing an hourly rate with a salary without the associated charges.
- Bringing an irregular need in-house out of concern for economy.
- Outsourcing without designating an available arbiter.
- Neglecting the cost of non-production periods in-house.
- Assuming an in-house team is assembled within weeks.
- Forgetting licences and infrastructure in the in-house calculation.
- Entrusting a very specific field without passing on the reference.
- Ignoring the regulatory constraints on data leaving the company.
- Treating the choice as final when it can be mixed.
- Comparing the two options on production cost alone.
