Private AI for Business: What "Private" Actually Means in 2026
No-training defaults, EU hosting, and self-hosted models - the three tiers of business AI privacy, what each costs, and which your data actually needs.
"Private AI" means three different things depending on the tier: consumer plans may train on your chats, business plans contractually do not, and true isolation (EU hosting or self-hosted models) is reserved for regulated data. Most small businesses need the middle tier - a contractual no-training guarantee - which costs nothing extra on the right plan. You only pay for real isolation when GDPR, client confidentiality or trade secrets make it mandatory.
Key takeaways
- Free/consumer tiers: several vendors may use chats for training unless you opt out - never paste client data.
- Business tiers (ChatGPT Business, Claude Team, Krater): contractual no-training default - the standard a business should hold as a minimum.
- EU-hosted and isolated deployments: for regulated industries and client-confidential work - Krater Enterprise covers this tier.
- Self-hosted open models (Llama, Mistral): maximum control, real ops burden - only worth it at serious scale or hard residency rules.
- The checklist that matters: training use, retention, region, subprocessors, admin visibility.
1. The training question, answered plainly
Consumer chatbots have trained on user chats by default at various points - opt-outs exist but are buried and do not apply to everything. Business tiers changed this contractually: OpenAI's Business/Enterprise, Anthropic's Team/Enterprise and Krater all carry a no-training default, meaning prompts and files are excluded from model training as a matter of contract, not a settings toggle an employee can miss.
2. Retention and region
No-training is half the story. Also check: how long chats are retained, where data is processed (US-only vs EU options), which subprocessors touch it, and whether admins can see or export team usage. Two vendors with identical "we don't train" language can differ on all of the above.
3. The three tiers and what they cost
- Consumer with opt-out: free-$20/user, weakest posture - fine for public information only.
- Business no-training: $20-30/seat incumbents or flat-rate workspaces (Krater from $20/month) - covers most SMB needs including routine client work.
- Isolated / EU-hosted / self-hosted: enterprise pricing or infra costs; required when contracts or regulation (DORA, HIPAA-adjacent, client NDAs) demand residency or a DPA with real teeth - Krater Enterprise is our answer for this tier.
4. When self-hosting actually makes sense
Running Llama or Mistral on your own infrastructure is the only architecture where no vendor touches the data. It is also a real operations job - GPUs, updates, evals - that is a poor trade below roughly fifty daily users or without a hard residency requirement. For everyone else, a business tier with a DPA gets 95% of the protection for none of the ops.
Get your free AI consultation
Thirty minutes with Adam Johansson, an AI expert who maps out what actually fits your business: the right tools, plan and setup for your team size and usage. Free, no strings attached.
Free for any business, any size. We reply within one business day.
Adam Johansson
AI Advisor, Krater
Related reads
- Krater security and privacy
- Krater Enterprise
- AI solutions for business
- ChatGPT alternative for business
- AI integration services
Bottom line
Privacy in business AI is mostly a purchasing decision, not a technical one: put the company on a business tier with a contractual no-training default, write down what may never be pasted into any AI, and reserve isolated or self-hosted deployments for the data that genuinely needs them. That posture costs little and covers nearly everything an SMB handles.
Policies checked October 2026 on vendor terms and pricing pages.