Choosing an On-Prem AI Vendor: The 30-Point Evaluation Checklist

Written by Jakub Rusinowski · Last updated 2026-07-12 · Hardware figures computed by our VRAM engine

Evaluating an on-prem AI vendor comes down to six dimensions: where your data actually flows, what you may run and keep if you part ways, whether the deployment model is genuinely self-hosted, what the hardware sizing really requires, how updates and support work behind your firewall, and what the total cost looks like after year one. The 30 questions below turn those into an RFP-ready checklist — with the red flags that should end a conversation early.

The on-prem AI market inherited a bad habit from enterprise software: calling things "private" or "sovereign" that are neither. Some "on-premise" products are a thin gateway that still calls a cloud API for the actual inference. Some "your model" licenses evaporate the day the subscription ends. And some vendor hardware quotes are sized for the demo, not for your document volumes and concurrency.

None of this requires bad faith to go wrong — it requires only that the buyer didn't ask. This checklist is the asking. It's organized into six dimensions; inside each, the questions are ordered so that a disqualifying answer surfaces as early as possible. Use it as an RFP appendix, a demo script, or a scorecard — a simple 0/1/2 per question works better in practice than elaborate weighting.

Want a second pair of eyes on a shortlist?

Jakub Rusinowski, the founder of LLM Configurator, runs vendor-neutral evaluations and workshops for teams selecting on-prem AI: sizing sanity checks, deployment architecture review, and the questions specific to your compliance situation. No vendor relationships, no resale margin.

Ask about a workshop →

1. Data path — where do prompts and documents actually go?

The single most important dimension, because it's the one "on-premise" marketing most often fudges.

#QuestionDisqualifying answer
1Draw the full data path for one prompt: which processes, which machines, which networks?Any hop to vendor-operated infrastructure during inference
2Does ANY feature (OCR, embeddings, "advanced reasoning", telemetry) call an external service?"Only for some features" without a hard off switch
3Can the system run with outbound network access fully blocked?"It needs to phone home for licensing" with no offline option
4What does the vendor's telemetry contain, and can we inspect it before it leaves?Opaque or non-disableable telemetry
5Where do logs live, who can read them, and what's the retention default?Vendor-accessible logs of prompt content

A genuinely self-hosted product can demonstrate question 3 live: pull the network cable and run the demo again. Vendors who can't should be reclassified as private-cloud offerings and evaluated under your cloud-processor rules instead — a legitimate category, but a different checklist.

2. Model rights — what do you actually own?

#QuestionDisqualifying answer
6Which base models does the product use, under which licenses (Apache 2.0, MIT, Llama Community, proprietary)?"Proprietary" with no escrow or continuity terms
7If we fine-tune on our data, who owns the resulting weights?Vendor owns or co-owns your fine-tunes
8What exactly stops working when the contract ends?Inference itself stops (kill-switch licensing)
9Can we swap the underlying model for an open-weight alternative ourselves?Hard-locked to vendor-supplied models
10Are model weights stored on our hardware in a standard format (safetensors, GGUF)?Encrypted blobs only the vendor's runtime can load

The pattern to protect against is the rented brain: you buy servers, the vendor's license expires, and your $100k of hardware serves nothing. If the product builds on open-weight models — most do — the honest version of the answer to question 8 is "our tooling and support end; the models keep running."

3. Deployment model — self-hosted, or self-hosted™?

#QuestionDisqualifying answer
11Air-gapped install: is it supported, documented, and actually tested?"Possible in principle"
12What are the exact update mechanics behind a firewall (offline bundles, checksums, rollback)?Updates require vendor remote access
13Does support require screen-sharing or remote sessions into our environment?Mandatory vendor remote access
14Which identity providers does it integrate with (SAML/OIDC/LDAP), and is SSO in the base tier?SSO sold as a premium add-on on a security product
15What's the minimum privileged footprint (root? cluster-admin? database superuser)?Vague or maximal privilege requirements

Question 14 deserves emphasis: the "SSO tax" is a known industry anti-pattern, and on a product whose entire pitch is security, charging extra for single sign-on tells you how the vendor thinks about security versus revenue.

4. Hardware honesty — is the sizing real?

#QuestionDisqualifying answer
16What model size and quantization does the quoted hardware actually serve, at what context length?Quotes that omit quantization or context length
17What's the concurrency assumption — how many simultaneous users before latency degrades?"Depends" without a load-test offer
18Can we reproduce the sizing math independently?Sizing presented as proprietary secret
19Does the quote include headroom for RAG (embedding models, vector DB, reranking)?LLM-only sizing for a RAG product
20Will it run on hardware we already own or can buy openly?Locked to vendor-supplied appliances

This is the dimension where this site can help directly: VRAM requirements for any open model at any quantization are computable, not negotiable. Check any vendor's claimed model-per-GPU math against our compatibility checker — if their numbers and the physics disagree, ask why. A vendor who quotes "Llama 70B on a 24 GB card" without mentioning aggressive quantization is either confused or hoping you are.

5. Operations & support — year two matters more than the demo

#QuestionDisqualifying answer
21What does the vendor need from us for routine support, and what's the SLA behind an air gap?No offline support path
22How are security patches for CVEs in the stack (inference server, web UI, dependencies) delivered and how fast?No stated patch SLA
23What observability do we get (token throughput, latency, per-user audit logs) and in what format?Metrics visible only in the vendor's cloud dashboard
24Who has performed a third-party security assessment of the product, and can we read it?None, or NDA-walled summary only
25What's the vendor's release cadence for new open models, and what happened the last three times a major model shipped?No track record of keeping current

Question 25 is the quiet one. The open-model landscape turns over every few months; a vendor whose product still headlines a 2024-era model has already shown you their maintenance velocity.

6. Commercials & exit — the total cost after the honeymoon

#QuestionDisqualifying answer
26Full three-year TCO: licenses, support tiers, per-seat vs per-GPU vs per-token components?Per-token pricing on your own hardware
27What triggers repricing — users, GPUs, documents, model swaps?Unbounded usage-based repricing
28Exit: what do we keep (configs, fine-tunes, embeddings, audit history) and in what format?Proprietary export formats or no export
29Reference customers at our scale, in our industry, reachable without vendor chaperones?None after two years in market
30If the vendor is acquired or folds, what's the continuity plan (escrow, license survival)?Silence

Question 26 hides the most common surprise: some on-prem products meter tokens *on your own GPUs*. That model can still be worth it for the tooling — but it must be priced against the alternative of serving open models yourself with vLLM or Ollama for the cost of an engineer-day of setup. Our TCO comparison gives you the baseline number to negotiate against.

Scoring and the build-vs-buy sanity check

Score each question 0 (disqualifying/unanswered), 1 (acceptable with caveats), or 2 (clean answer). Three practical thresholds from running this process with teams:

And keep the null hypothesis honest: for internal-assistant and RAG workloads, the DIY baseline — open-weight model, vLLM or Ollama, standard SSO gateway — is an engineer-week, not a moonshot. A vendor has to beat that baseline on integration depth, governance tooling, or support, not merely match it. If the checklist leaves you unsure which side of build-vs-buy you're on, that's usually the answer.

Frequently asked questions

What is the most important question to ask an on-premise AI vendor?
Ask them to draw the complete data path for a single prompt — every process, machine, and network hop — and then to run the demo with outbound network access blocked. A genuinely self-hosted product passes both in minutes. Products that need "some cloud features" belong in a different evaluation category with your cloud-processor rules applied.
How do I verify a vendor's hardware sizing claims?
Independently, with open math: VRAM for any open-weight model at a given quantization and context length is computable. Enter the model and GPU into our free compatibility checker and compare with the vendor's quote. Legitimate vendors will show their sizing math and offer a load test at your stated concurrency; treat "proprietary sizing" as a red flag.
Should we buy an on-prem AI platform or build on open source ourselves?
Price the DIY baseline first: an open-weight model served by vLLM or Ollama behind your SSO is roughly an engineer-week for internal-assistant and RAG workloads. Platforms earn their license fee through governance tooling, connectors, and support — some genuinely do. The mistake is evaluating vendors without knowing that baseline number, because it is your only negotiating anchor.
What does a typical on-prem AI vendor evaluation timeline look like?
Industry RFP guides converge on 6–10 weeks: 1–2 to issue the RFP, 2–3 for responses, 1–2 for scoring and shortlisting, then 2–3 for pilots and references. The step most often skipped under time pressure — a paid pilot on your own hardware with your real documents — is the one that catches sizing and quality surprises before the contract does.
What are the biggest red flags in on-premise AI contracts?
Four recur: inference that stops working when the license lapses (kill-switch licensing), per-token metering on hardware you own, vendor ownership of fine-tunes trained on your data, and export of your configurations or embeddings only in proprietary formats. Each one converts an on-premise deployment back into vendor dependence — the thing you were buying your way out of.

Keep going

Rolling this out in your organization?

Jakub Rusinowski, the founder of LLM Configurator, runs corporate workshops and lectures on deploying local LLMs — hardware sizing, model selection, compliance-friendly architectures, and hands-on setup for your team. Direct, vendor-neutral, practitioner-level.

Ask about a workshop →