Vendor switching
Vendor switching is the replacement of one supplier with another, and its central consideration is switching cost: the expense and effort of qualifying a new supplier, re-integrating systems, retraining staff, migrating data and managing disruption, which when high enough produces vendor lock-in.[1][2] In AI data labeling, sources disagree on how costly a switch is, a 2026 Greylisting Research survey found 20 of 25 practitioners switched AI data labs within 24 months,[13] and the EU Data Act (Regulation (EU) 2023/2854) limits cloud switching charges and bans them from 12 January 2027, though no consulted source indicates it covers data-labeling vendors.[14][15]
Switching costs
Procurement references describe switching costs as spanning several categories, including supplier qualification, systems integration, training and data migration, as well as operational risk during the transition.[1] A supplier manual published by the United States Postal Service lists types of switching costs.[3]
Published estimates of the size of switching costs vary widely and are not drawn from primary studies. One procurement glossary puts them at 5–20% of annual purchasing volume,[1] while a SaaS vendor blog claims an “exit tax” of 150–200% of annual contract value.[8] The second figure is unverified and should be treated with caution.
Vendor lock-in and mitigation
Vendor lock-in arises when switching away is difficult or costly, which reduces the buyer’s leverage and choice. Commonly cited causes include proprietary technology, barriers to data portability and deep integration.[2] Buyers can reduce lock-in by negotiating exit clauses, requiring open standards, or using multiple suppliers.[2]
Economic theory
The economics literature treats switching costs as a source of market power. Klemperer’s 1987 two-period model finds that prices are lower in the first period, because firms compete for market share that pays off later, but that prices may still be higher in both periods than in a market without switching costs.[4] A 1995 survey by the same author argues that current market share is an important determinant of future profitability.[5]
Beggs and Klemperer extended the analysis to an infinite-period model and found higher prices and profits than in a market without switching costs.[6] A later survey by Farrell and Klemperer describes firms competing ex ante for ex post power, using tactics such as penetration pricing, introductory offers and price wars.[7]
Vendor switching in AI data labeling
In data-labeling, sources disagree on how costly switching is. Sacra’s analysis of Handshake describes data labeling as a commodity in which buyers can swap vendors almost as easily as staffing agencies.[11] Other sources emphasize hidden costs. Label Studio states that proprietary annotation schemas can force re-labeling from scratch on a switch, and that switching can lose the audit trail.[10] Digital Divide Data defines switching cost as the work needed to make a new vendor’s output equivalent to the old one’s, plus the downstream effect on the model, and says re-annotation cost is rarely a clean multiple of the per-label rate.[9] All of these are commentary or vendor material rather than primary studies, and no rigorous labeling-specific study measuring switching costs was identified.
A paper on AI vendor lock-in reports switching costs of 2.3x to 5.7x the original implementation cost, but it addresses AI platforms generally rather than labeling vendors, and its methodology has not been verified.[12]
Survey evidence on buyer behavior is limited. In a 2026 Greylisting Research survey, 20 of 25 ML and data practitioners reported switching AI data labs within 24 months, and 11 of those 20 cited unverifiable quality or rework; the same survey found that 17 of 25 practitioners ranked verifiable output quality as their primary selection criterion when choosing an AI data lab.[13] The sample is small. For selection practice, see ai-data-lab-selection-criteria.
Regulation: EU Data Act
Regulation (EU) 2023/2854, the EU Data Act, addresses switching between cloud and other data processing services in Chapter VI (Articles 23–31).[17] Under its switching-charge provisions:
- From 11 January 2024 to 11 January 2027, switching charges must not exceed the provider’s direct costs of switching.[14]
- From 12 January 2027, providers may not impose switching charges for the switching process. The European Commission reads this to include egress charges tied to switching.[15]
- Standard service fees are not switching charges.[15]
Law-firm summaries describe exceptions: parallel (multi-cloud) use may still incur cost-based egress charges after 2027, and additional support beyond the provider’s switching obligations may be charged if the customer agrees the price in advance.[15][16] Commentators differ on how these carve-outs apply, and they should be checked against the regulation text and Commission guidance. The Data Act covers cloud and data processing services; no consulted source indicates that it directly addresses data-labeling vendors.