How to Write an RFP for AI Training Data Services

Why AI training data RFPs fail

Most AI training data RFPs fail to get useful responses because they are structured around the wrong unit of comparison. An RFP that asks only for per-unit rates, service listings, and certifications will get responses that cannot be meaningfully compared -- every vendor will quote a low rate for a generic task, claim to cover all relevant languages and modalities, and present their certifications. The buyer cannot distinguish capability from marketing.

A useful AI data RFP asks vendors to demonstrate specific capabilities against the actual requirements of the program. It specifies task types precisely enough that per-unit quotes are comparable across vendors. It asks for evidence of past performance on comparable work. And it structures the evaluation criteria to reward actual capability rather than the ability to write a polished proposal.

This guide gives you a structured RFP template organized into six sections, with the specific questions that distinguish capable vendors from well-marketed ones.

Section 1: Program specification

The program specification section describes what you are buying in enough detail that vendors can quote accurately. It should include: modality (image, video, text, audio, LiDAR, multimodal); task types with specific descriptions (not just 'annotation' but '3D cuboid annotation of LiDAR point clouds for AV perception'); volume estimates and ramp timeline; language and dialect requirements with specific dialects if applicable; turnaround requirements; output format specifications; and integration requirements (tooling platform, API format, delivery method).

Be specific about the annotation ontology if it is defined. If the ontology will be co-developed with the vendor, say so explicitly. Ontology co-development is a scope item that should be priced separately from annotation volume -- vendors who do not separate these will either underprice the annotation or overprice the ontology development in a way that is invisible to the buyer.

Include a sample task description -- ideally with 10 to 20 example items -- that vendors can use to calibrate their quotes. Without sample tasks, quotes for complex annotation tasks will vary by 300 to 500% based on different assumptions about complexity. Sample tasks make the comparison tractable.

Section 2: Workforce and delivery model

The workforce section asks vendors to describe who will actually do the work. The key questions: Are annotators direct employees, long-term contractors, or anonymous crowd workers? Where are they physically located? What is the average tenure of annotators on comparable projects? How do you handle staff turnover mid-project?

For multilingual requirements: What is the language background of annotators for each required language? Are they first-language native speakers? What dialect communities do they represent? How do you calibrate for dialect-specific tasks?

Ask for a description of the project team structure: who is the named project manager, what is their role, and what is their availability for client communication? For complex, long-running programs, the project management relationship is often more important than any other operational variable.

Section 3: Quality architecture

The quality section asks vendors to describe their quality process in specific, auditable terms. Ask for: the QA process step-by-step for the specific task type in your RFP (not a generic description of their process); how inter-annotator agreement is measured for this task type; what IAA scores they have delivered on comparable projects (ask for the distribution, not the average); what the escalation process is when a batch misses threshold; and what rework policy applies when delivered accuracy falls below the contracted threshold.

Ask for a sample QA report from a comparable project. This reveals whether the vendor actually documents QA in the structured way they describe. A vendor who cannot quickly produce a sample report is not running the QA process they describe.

Specify your accuracy requirements explicitly in the RFP: minimum acceptable accuracy threshold, measurement method, and consequences of missing threshold. This forces vendors to price their QA architecture into the quote rather than underbidding and arguing about threshold definitions after the fact.

Section 4: Security, compliance, and data handling

The compliance section asks for specific documentation commitments, not just policy statements. Required items: list of current certifications with issue dates and scope (SOC 2 Type II, ISO 27001:2022, HIPAA BAA if applicable); jurisdiction-specific compliance alignment (GDPR DPA + SCCs for EU data; CCPA service-provider addendum for California; PIPEDA for Canada; LGPD for Brazil; POPIA for South Africa); and data handling procedures (encryption at rest and in transit, access controls, audit logging, endpoint controls for remote workforce).

For any program involving personal data: request the subprocessor list and ask whether all subprocessors who handle personal data are covered by appropriate DPAs. For programs involving regulated data (healthcare, financial): request the specific regulatory compliance commitments in writing, not just a general statement of compliance.

Ask vendors to describe their breach notification process: what triggers notification, who is notified, and within what timeframe. The answer should reference specific regulatory requirements (GDPR's 72-hour obligation, HHS HIPAA notification timelines) rather than generic language.

Section 5: Pricing and commercial terms

The pricing section should ask for a total cost breakdown for the program specification you provided in Section 1, not just a per-unit rate. Ask specifically for: setup and calibration costs; per-unit rates by task type; QA and project management included vs. separate; rework policy (included in rate, or billed separately); volume tiers and minimum spend commitments; rush pricing for expedited delivery; and what causes scope to change and how changes are priced.

Ask for a worked example: take a representative batch of 10,000 items from your Section 1 specification and ask vendors to show the total cost for that batch including all ancillaries. This makes per-unit rates comparable across vendors with different bundling approaches.

Request a reference from at least one comparable project with a contact you can call. Ask the reference specifically about cost predictability: did the final invoice match the quoted rate, and what caused any variance?

Section 6: Evaluation criteria

A well-structured RFP gives vendors the evaluation criteria upfront. This serves two purposes: it tells vendors what matters (so they address it specifically rather than using a generic proposal template), and it commits the buyer to a structured evaluation that reduces post-RFP scope disputes.

A reasonable weighting for an enterprise AI data program: technical capability and workforce quality (35%); quality architecture and QA evidence (25%); security and compliance posture (20%); pricing and commercial terms (15%); project management and communication (5%). For programs involving regulated data or safety-critical applications, weight security and compliance higher.

Require a technical demonstration or pilot as part of the evaluation. A paid pilot of 500 to 2,000 items from the actual program -- not a demonstration dataset -- is the most efficient way to evaluate capability. Vendors who refuse to participate in a paid pilot on reasonable terms are signaling something about their confidence in their own quality.

Frequently asked questions

What should an AI training data RFP include?

An AI data RFP should include: program specification (modality, task types, volume, languages, turnaround); workforce model questions; quality architecture requirements; security and compliance documentation requirements; pricing structure (total cost, not just per-unit rate); and evaluation criteria. Include sample tasks to make quotes comparable.

How do I compare AI annotation vendor quotes?

Require vendors to quote on a worked example from your actual program specification -- a defined batch of items with specific task types, not just a per-unit rate. Ask each vendor to break out setup, QA, rework policy, and project management separately. Compare total delivered cost, not headline rate.

Should I run a pilot before awarding an AI data contract?

Yes. A paid pilot of 500 to 2,000 items from your actual program is the most effective evaluation tool. It reveals real quality, realistic turnaround, and communication quality under production conditions. Vendors who refuse reasonable pilot terms should be viewed skeptically.

What compliance documentation should I require in an AI data RFP?

Request: SOC 2 Type II report (under NDA), ISO 27001:2022 certificate, GDPR DPA + SCCs for EU data, subprocessor list, and applicable regulatory overlays (HIPAA BAA, CCPA addendum, LGPD DPA, etc.). Ask for issue dates and scope, not just a claim of certification.

How much should AI training data cost?

Costs vary by modality and complexity. Simple image annotation runs $0.05 to $0.15 per item; RLHF preference pairs run $0.50 to $25 per pair; LiDAR 3D cuboid annotation runs $2 to $20 per frame. Managed offshore providers deliver at 50 to 70% below US-domestic rates. Get total cost quotes including setup, QA, and rework policy, not just per-unit rates.