Quite a lot of AI startups today are not building models. They are packaging OpenAI, Anthropic or other foundation models behind a thin UI or workflow. That is not automatically wrong. But it creates material risks for buyers that many companies ignore until it is too late.
Below is a clear view of the issues, what they mean for a business, and what you should check before buying from one of these vendors.
The core problem
If the “AI startup” does not control the model, they also do not control:
- cost
- availability
- data handling
- roadmap
- compliance
- security posture
You are relying on two suppliers. One owns the UX. The other owns the intelligence. When the middle company adds no real IP, you carry unnecessary risk.
Cost risk
Startups built on top of another model provider have no pricing power. They pass through API charges. When OpenAI or Anthropic adjust rates, the startup has to follow or take a loss. Impact on you:
- unpredictable long-term pricing
- sudden cost increases when usage scales
- hard to forecast ROI or budget
If the vendor is subsidising costs to win customers, expect a big correction later.
Reliability and continuity risk
Most wrappers have no redundancy. If their upstream provider goes down or rate-limits them, your service stops. They usually lack alternative model integrations because that requires more engineering than UI work.
Impact on you:
- unplanned outages
- slow performance during load spikes
- no SLA they can actually enforce
- service disruption if the upstream changes terms or blocks their account
Small startups also face survivability risk. If their margin is thin, a change in API rules can kill the business overnight.
Data protection and compliance issues
The vendor often assumes that sending your content to a foundation model API is acceptable. They rarely provide:
- processing agreements with the actual model provider
- clear data flow maps
- country/region processing information
- retention rules
- auditability
- proof of isolation for sensitive sectors (health, finance, government)
Impact on you:
- regulatory gaps
- audit failures
- unclear cross-border transfers
- risk of non-compliance with ISO 27001, SOC 2, NIS2, GDPR
Many startups cannot answer basic questions such as “where is inference processed” or “are prompts stored”.
Security weaknesses
Wrappers usually rely on the model provider’s security posture. They rarely have:
- secure prompt handling
- encryption at rest with key ownership
- proper authentication and RBAC
- tamper-proof logs
- model-layer access controls
- secure fine-tuning pipelines
Impact on you:
- exposure to data leakage
- insider misuse risk
- weak audit trails
- dependence on a vendor with minimal internal security capability
If you are in a regulated sector, this is a non-starter.
Weak IP and vendor lock-in
If the vendor owns no proprietary model, all differentiation is UI and workflow logic. That means you can recreate the product internally or with another integrator.
Impact on you:
- no defensible value
- limited roadmap
- slow innovation
- entire product depends on what OpenAI or Anthropic release next
- hard lock-in because they may not let you export workflows, prompts or training data
You end up tied to a vendor with shallow capability.
Ethical and governance gaps
Wrappers rarely implement:
- model risk assessments
- bias evaluation
- safety constraints
- red-team testing
- content filters
- monitoring for harmful outputs
Impact on you:
- higher risk of harmful or incorrect outputs
- potential brand damage
- legal exposure if outputs cause harm
- no way to demonstrate responsible AI use
Compliance frameworks like ISO 42001 expect this discipline. Most wrappers are nowhere near it.
No real “off switch”
If the startup closes, changes terms or is acquired, you lose your automation layer. Without access to the underlying prompts, logic or tuning data, migrating becomes painful.
Impact on you:
- operational disruption
- forced vendor change under pressure
- expensive re-implementation
- loss of historical learning or fine-tuned data
What smart buyers should do
If you want to use a vendor that is effectively a wrapper, you need to run a tight evaluation.
Ask these questions:
- What models do you use and who processes the data?
- Do you have contracts in place with those providers that cover privacy and security requirements?
- Can you run your product on multiple model providers?
- Can we export workflows, prompts and fine-tuning data?
- Do you provide a full data-flow diagram?
- Do you store prompts or responses? If so, for how long and where?
- What happens if your upstream provider changes their policies, pricing or rate limits?
- Do you have your own safety, audit and validation layers or do you rely on the model provider?
- Are you ISO 27001 / SOC 2 / ISO 42001 aligned or on a roadmap?
- What is your margin structure? Is your business dependent on third-party cost reductions?
You will filter out weak vendors quickly.
What vendors should offer if they want credibility
A serious AI startup should:
- support multiple models or at least have a migration capability
- invest in their own safety, filtering and orchestration layers
- provide transparent data flows and compliance docs
- implement security controls equal to other SaaS providers
- deliver clear SLAs with measurable uptime
- show actual IP: agents, frameworks, fine-tuning pipelines, retrieval layers, not just UI
- offer export and portability options
If they can’t do these things, they are not enterprise-ready.
Final view
Most “AI startups” today are resale layers. Some provide value through UX and vertical workflow optimisation. Many do not. The real risk sits with businesses that assume these vendors are building genuine technology when they are not.
If you buy from a vendor without understanding the upstream dependency, you take on extra operational, compliance and financial risk. The smart approach is simple. Check the architecture. Ask the hard questions. And avoid vendors whose only asset is someone else’s model with a thin interface on top.