Check the boundaries
Licensing, data flows, hosting arrangements and security requirements.
Safe adoption of global AI models
Choose a model with a clear view of the whole system.
Evaluate hosted and open-weight models against your workload, data requirements and full operating costs.
Same taskModel optionsShared criteria
A starting point
A focused model comparison.
Define one workload and a shortlist. We compare task performance, data handling, licence terms and operating costs against agreed criteria.
Discuss a model comparisonWhen it helps
You need to choose a provider, test a lower-cost model or decide whether self-hosting is worth the work.
Bring what you have
A workload, representative tasks and any shortlisted models. We agree data access and test conditions before testing.
The work
Compare performance, control and cost.
Evaluate your task
Representative examples, failure cases and a fair comparison of model options.
Understand the cost
Infrastructure, integration, human review, maintenance and supplier dependencies.
What you leave with.
- A task-based model shortlist
- Findings and limitations
- A view of full operating costs
- An adoption decision record
Methods and supporting detail
Begin with the task
We define the work the model needs to support and the errors that would matter. A customer-support assistant, an internal document tool and a product feature require different evidence, even when they use the same underlying model.
This creates a basis for a useful shortlist and a fair comparison with the existing approach.
What we assess
Suitability and evaluation
Compare candidates using representative examples, agreed acceptance criteria and the intended system configuration. We examine difficult cases, unsupported answers, refusal behaviour and the human work needed to check results.
Data flows and deployment
Compare hosted services, dedicated infrastructure and self-managed deployment. Identify what leaves the environment, where logs and supporting services run, how access is controlled and how data is removed.
Running a model locally can change the data flows, but it does not remove the need to examine the rest of the application.
Licensing and supplier dependencies
Identify the terms attached to the exact model version, supporting code and intended use. Surface restrictions, distribution requirements and questions needing legal review before the organisation commits.
We also consider update policies, version availability, service continuity and the practical effort involved in changing provider.
Security and operations
Examine permissions, connected tools, external content, logging and incident handling. Define how the system will be tested before release and reassessed after a model, prompt, integration or supplier change.
Cost and maintainability
Compare the cost of completing the task to an acceptable standard: model usage, hosting, integration, failed attempts, monitoring and human review. We record assumptions about volume and workload so the comparison can be revisited as usage changes.
An international perspective
Our team brings Mandarin capability and experience across UK, European and Chinese technology environments. That helps us understand documentation, supplier conversations and commercial context across markets.
We assess the actual provider, terms and deployment arrangement. Country of origin is one part of due diligence, rather than a substitute for it.
Before we begin.
Are open-weight models always cheaper?
No. The model is one part of the cost. Infrastructure, integration, support and human review can change the result.
Can you assess models from China?
Yes. We use the same task, data, licensing and security questions regardless of origin, with Mandarin capability where it helps the assessment.
A useful place to start
What decision is ahead of you?
Tell us what you are building, adopting or trying to understand.