Evaluate
Will private AI work for your workload?
Free tools for the decisions that come before buying anything: what it costs compared with your current hosted usage, what hardware the workload actually needs, and where each part of the system runs. Results appear in your browser with no sign-up.
Cost
Hosted vs self-hosted cost comparison
Purchase cash flow and annualised ownership cost, kept separate. Includes power, admin time, support and any residual hosted usage. Shows a break-even month, or tells you there isn’t one.
Open the tool →Hardware
Workload-based hardware sizing
Memory for weights and KV cache from model architecture, context length and concurrency, then which GPUs fit and the estimated per-stream decode speed.
Open the tool →Deployment-boundary checklist
“Can this run entirely inside our infrastructure?” has a precise answer only when every component is traced. For each item, answer two questions: where does it run? and does it need internet access?
| Component | What to establish | Runs where? | Needs internet? |
|---|---|---|---|
| Source documents | Where are they stored, and does the assistant read them in place or copy them? | On-device / on-prem / hosted | Yes / no / for updates only |
| Extraction and OCR | Does parsing PDFs, scans or Office files call a hosted service? | On-device / on-prem / hosted | Yes / no / for updates only |
| Embeddings | Is the embedding model local, or an API call per chunk and per query? | On-device / on-prem / hosted | Yes / no / for updates only |
| Vector store | Self-hosted, embedded, or a managed cloud index? | On-device / on-prem / hosted | Yes / no / for updates only |
| Prompts and retrieved context | Which process assembles them, and where is it logged? | On-device / on-prem / hosted | Yes / no / for updates only |
| Inference | Which runtime and model? Are cloud models, web search or plugins disabled? | On-device / on-prem / hosted | Yes / no / for updates only |
| Logs and telemetry | Do the app, runtime or extensions send usage data or crash reports anywhere? | On-device / on-prem / hosted | Yes / no / for updates only |
| Outputs | Where are answers stored, exported or forwarded (chat history, tickets, email)? | On-device / on-prem / hosted | Yes / no / for updates only |
| Authentication | Does sign-in depend on an external identity provider that must be reachable? | On-device / on-prem / hosted | Yes / no / for updates only |
| Updates and model downloads | How do models, containers and patches arrive in an offline or restricted network? | On-device / on-prem / hosted | Yes / no / for updates only |
| Support access | Who can reach the system for maintenance, and from where? | On-device / on-prem / hosted | Yes / no / for updates only |
Three deployment boundaries
On-device
Processing happens on the user's own computer or phone. Nothing is shared between users unless you build that in.
On-premises
Processing happens on infrastructure your organisation owns and operates, in your building or your colocation space.
Private hosted or hybrid
Explicitly approved external infrastructure, or a defined mix, with every external flow named in writing.
Want the trace done against your actual stack?
A fit assessment maps every dependency above for one workload, tests quality and permissions on agreed examples, and ends with a go/no-go recommendation — including “stay with your hosted solution” when that is the right answer.
Frequently Asked Questions
Do I need to give an email address to use these tools?
No. Both calculators run entirely in your browser and show their results immediately. Nothing you enter is sent anywhere unless you choose to forward a summary with an assessment request.
How accurate are the results?
They are first-order estimates, labelled as such. They are useful for ruling options in or out and for framing a budget. They do not replace a benchmark of your workload on candidate hardware, which is what a fit assessment provides.
Is running the model locally enough to make a deployment private?
Not by itself. Document extraction, embeddings, the vector store, logging, authentication and updates can each send data outside your boundary. The checklist on this page traces every component, not only the model.