Procurement & Vendor Management (Deep Dive)
In one line: A new AI vendor at a 500+ engineer enterprise takes 3–9 months to onboard — Security, Privacy, Legal, AI/Responsible-Use, and Procurement each have their own review — and the right play is to start all five clocks the same week you start the proof-of-concept, not after.
At home, buying software means typing in a credit card. At a big company, buying an AI tool means five separate teams — Security, Privacy, Legal, the responsible-AI group, and the purchasing department — each examine the vendor before anyone can use it. That examination takes 3 to 9 months, and it can quietly become the longest part of your whole project. The trick this page teaches: start the paperwork the same week you start trying out the tool, so both finish around the same time.
The Architecture page lists the vendors a typical enterprise stack uses (Bedrock, Azure OpenAI, Vertex, Portkey, Vespa, Snowflake, Datadog, etc.). This page is the operational view: what it actually takes to add a new vendor to that list, why it takes so long, and what an engineer can do to make procurement not the bottleneck on a launch.
The reality
A new AI vendor — even a small one, even one you trust — typically requires:
- Security review (SOC 2 Type II, ISO 27001, SIG questionnaire, pen test reports).
- Privacy review (DPA, data flow diagram, sub-processors, residency, GDPR/CCPA assessment).
- Legal review (MSA, indemnification, IP terms, AI-specific clauses like training-data warranties).
- AI / Responsible-Use review (training data sources, model cards, bias testing posture, opt-out controls).
- Procurement (commercial negotiation, contract, PO, vendor master data setup).
End-to-end: 3–9 months. For sensitive or high-spend deployments, longer (12+ months for new model providers with significant data exposure).
What each function actually looks at
Security review
- SOC 2 Type II report (current, ideally within 12 months).
- ISO 27001 cert (preferred but not always required).
- HIPAA BAA (if PHI is in scope).
- SIG (Standardized Information Gathering) questionnaire — long, detailed, painful for vendors but the standard.
- Pen test summary (third-party tested).
- Vulnerability disclosure program.
- Sub-processor list.
- For AI vendors specifically: how is the vendor itself securing its model traffic, its training data, its model weights, its customer prompts.
Privacy review
- DPA (Data Processing Agreement) — usually a back-and-forth on terms.
- Data flow diagram — what data goes from us to them, where it sits, how long.
- Sub-processor list — every party that touches the data.
- Residency commitments — EU data stays in EU?
- GDPR Article 28 obligations, CCPA service-provider language.
- For AI: opt-out from training on customer data, content-retention policies, deletion-on-request.
Legal review
- MSA (Master Service Agreement) — many clauses, often a multi-month negotiation.
- Indemnification (who pays if their model causes harm).
- IP ownership of outputs.
- Warranty / liability caps.
- Termination terms, exit / data portability.
- AI-specific: model-warranty language, training-data warranties, IP-infringement indemnity (for model outputs that arguably reproduce training data).
AI / Responsible-Use review
- Training data sources — what was the model trained on; provenance.
- Model cards — capabilities, limitations, known biases, intended uses.
- Bias and fairness testing posture.
- Safety evaluations — what alignment / red-teaming the vendor has done.
- Versioning and EOL policy.
- Content provenance / watermarking (relevant under EU AI Act).
This function may sit in Legal, in Responsible AI, or as a virtual review by the AI Risk Committee.
Procurement
- Commercial negotiation — list price → discount.
- Multi-year commit terms.
- Vendor master data setup.
- PO and invoicing.
End-to-end timeline (typical)
A new mid-sized AI SaaS vendor (say, an eval platform):
| Phase | Elapsed time | Active work |
|---|---|---|
| Initial vendor discovery + POC | Week 1–4 | Engineering pilot, in parallel below |
| Security review (kickoff to closeout) | Week 1–10 | SIG response, doc review, follow-ups |
| Privacy review (DPA) | Week 4–14 | DPA back-and-forth |
| Legal review (MSA) | Week 6–18 | MSA negotiation |
| AI / Responsible-Use review | Week 6–12 | Model cards, training data discussion |
| Procurement / commercial | Week 12–18 | Discount, commit terms, PO |
| Contract signature + vendor onboarding | Week 18–20 | |
| Production go-live | Week 20+ |
End-to-end: 4–5 months for a relatively simple AI SaaS vendor. A model provider with deep data exposure routinely takes 6–12 months.
What accelerates it
- Vendors with existing approvals at your company. Adding a new SKU from an already-approved vendor (e.g., enabling a new Bedrock model on existing Bedrock contracts) can take days.
- Vendors with SOC 2 Type II, ISO 27001, HIPAA BAA already in place. Drastically shortens Security review.
- Vendors with private deployment options (BYO cloud, single-tenant, contractual data-residency).
- Buying through a hyperscaler marketplace (AWS Marketplace, Azure Marketplace, GCP Marketplace). The marketplace often inherits an existing contract path, shortens Procurement, and applies existing approved vendor status.
- Pre-approved AI vendor list. The Responsible AI / Procurement function maintains a curated shortlist of pre-vetted AI vendors. Engineers steered to this list ship faster.
- Vendor liaison engagement. Larger vendors have enterprise customer-success teams who shepherd their own review process; use them.
- Standard contracting templates for AI vendors (the company's preferred clauses already drafted).
What slows it down
- Sub-processor list with surprising entries (e.g., training data labeled by a third party in a country your residency policy disallows).
- Training-on-customer-data terms without strong opt-out.
- No EU residency option for a tool you'll need in EU.
- New ownership / unclear funding — vendors recently acquired or with funding question marks.
- "Beta" or "preview" SKUs. Legal hates these because the warranty language is often weak.
- Vendor that refuses to sign your DPA without significant edits. Long back-and-forth.
- Vendor without a HIPAA BAA when you need one. Often a deal-breaker for healthcare-adjacent companies.
- Vendor without a SOC 2. Possible to navigate but adds months of additional security review.
The pragmatic playbook
Working patterns:
- Start vendor onboarding the same week you start the POC. Both clocks run in parallel. By the time engineering says "this works, let's ship," the legal/security clocks are already half-done.
- Maintain a pre-approved AI vendor shortlist. New AI work goes to the shortlist first; off-shortlist requires a deliberate "we need this" justification.
- Use the gateway to abstract providers. Adding a new model provider behind the gateway is a smaller change than adding a new app-level integration; the engineering surface is smaller, so the review surface is smaller.
- Lean on hyperscalers when timelines are tight. Bedrock-mediated access to Anthropic is often much faster than a direct Anthropic enterprise contract, because the contract is already in place via AWS.
- Use marketplace SKUs. AWS Marketplace + Azure Marketplace + GCP Marketplace cover much of the AI ecosystem in 2026; the inherited contract path is real value.
- Get engineering to the negotiation table. Pure-procurement negotiations get list-minus-30%; engineering input on load shape, commit horizon, model mix gets list-minus-50% or better.
The deepest reason every mature enterprise eventually has a central AI gateway isn't observability or cost — it's that the gateway shortens the procurement surface for new providers.
Without a gateway, adding "let's try Mistral Large for our French support feature" means a new app-level vendor integration, new auth, new observability wiring, new exception handling — and a procurement review focused on a complete application stack.
With a gateway, the same "let's try Mistral Large" is "add this model to the registry, route this feature's calls to it." Procurement reviews a much smaller surface; engineering ships in days instead of months once approval lands.
This is one of the highest-leverage operational benefits of central platform investment, and it almost never shows up in the platform team's stated value proposition.
Vendor ongoing management
Approval isn't the end. Ongoing:
- Annual SOC 2 refresh. Vendor's SOC 2 report needs to be re-reviewed annually.
- DPA + sub-processor updates. Vendors update their sub-processor lists; you need to be notified and re-approve.
- Contract renewals. Multi-year commits hit renewal; renegotiate based on actual usage and roadmap.
- Incident notifications. Vendor security incidents have to be communicated within contractual SLAs.
- EOL handling. Vendor announces a model EOL; your model-EOL playbook kicks in (see Release Management).
A "Vendor Relationship Manager" role often exists for major vendors at large enterprises — the named human who keeps the relationship healthy.
What changes vs. startup vendor procurement
| Startup | Enterprise | |
|---|---|---|
| Onboarding time | A credit card | 3–9 months |
| Reviews required | None | Security + Privacy + Legal + AI + Procurement |
| Contract type | Click-through TOS | MSA, DPA, BAA, custom AI clauses |
| Procurement | Doesn't exist | Multi-month commercial negotiation |
| Hyperscaler advantage | Doesn't apply | Often the fastest path |
| Marketplace advantage | Marginal | Major time-saver |
Common mistakes
- Starting procurement after the POC succeeds. You've just added 3–9 months to the timeline. Start both clocks the same week.
- Picking a vendor that doesn't have a SOC 2 because "the product is so much better." The product being better is rarely worth 6 extra months. Pick the SOC-2-having competitor when you can; push the better one to get a SOC 2 if you must.
- Skipping the hyperscaler / marketplace path when it's available. The contract overhead is the largest hidden cost in enterprise AI vendor onboarding. If the hyperscaler version exists, take it.
- Negotiating without engineering input. Procurement alone gets list-minus-30%; engineering at the table gets list-minus-50%+ with better commit terms. Be in the room.
- Letting the pre-approved vendor list go stale. A list updated annually is a list nobody uses by Q3. Quarterly refresh, owner named.
- Forgetting that vendor approval includes ongoing obligations. A signed contract isn't "done" — annual SOC 2 refreshes, sub-processor updates, contract renewals all need owners.
- Engineering treating procurement as someone else's problem. The engineer who shepherds their vendor through procurement gets months faster than the engineer who waits passively. Procurement is part of the project.
What's next
→ Next: Regulatory landscape — the regimes that drive much of what procurement and governance are actually trying to satisfy.