Model Pilot
You write the request -- the pilot picks the best model. Code, image, research or fast answer: routed to fit.
How the pilot routes
How Model Pilot works
Five steps from prompt to answer -- with transparent routing rationale in the audit log.
- 1
Just write the request -- no model picker
Write as usual. Model Pilot is on by default and takes the call. Switching to manual selection is one click away.
- 2
Pilot analyses the request in real time
Based on key signals (attachments, code, keywords, desired tone) Pilot classifies the request into a category: code, vision, reasoning, research, fast answer.
- 3
Routing to the optimal model
Each category has its preferred model -- e.g. Claude Opus for reasoning, Gemini for vision, GPT with web search for research. The company admin can adjust per team.
- 4
Answer with routing rationale
The answer comes back and Pilot shows a short rationale: 'Reasoning detected -> Claude Opus'. The choice stays transparent and auditable.
- 5
Audit entry with model rationale
Every routing decision lands in the audit log: detected category, chosen model, alternative models, token usage.
Who benefits from Model Pilot?
Four roles for whom automatic model choice saves time and tokens.
Generalists & power users
Anyone juggling different tasks all day -- email, research, code, image analysis -- saves the per-task model decision.
IT & engineering
Pilot routes code requests to a coding model, architecture discussions to a reasoning model -- without each engineer needing to know the models.
Marketing & content
Tone tasks to creative models, fact-checks to research models, image descriptions to vision models.
Onboarding new staff
New colleagues do not need to learn 100+ models -- Pilot takes over while they focus on the task.
Why auto-routing?
Frequently asked questions on Model Pilot
Answers on routing logic, configuration, cost and override options.
How does Model Pilot pick the right model?+
Pilot analyses the request using key signals: language, attachments (image, code, PDF), task type (reasoning, creative, research, fast answer). From these signals a category is derived that selects the matching model.
Which categories does Pilot distinguish?+
Currently five: reasoning & analysis, vision & image, web research & live data, code & tech, fast answer & routine. Each category has a default model that is overridable per tenant.
Can I override the model choice anytime?+
Yes. The chat window has a 'Pilot' / 'Manual' toggle. Manual lets you pick from 100+ EU models directly. You can switch per request even within the same chat.
Does Pilot really save tokens and cost?+
Yes. Routine requests get routed to smaller, cheaper models (e.g. Claude Haiku instead of Opus, Mistral Small instead of Large). Savings average 30-60 percent versus 'always the largest model'.
Who decides which model is used per category?+
The company admin in the superadmin panel per tenant. Defaults are Claude Opus 4.7, Gemini 2.5 Pro, GPT-5.x, Mistral Large -- depending on plan configuration.
Does Pilot work with all data types?+
Yes -- text, image, PDF, Excel, CSV, code, voice input. If Pilot detects e.g. an image in the upload, it auto-routes to a vision model.
Can Pilot also poll multiple models for the same request?+
Currently one model is picked per request. For parallel answers from multiple models use Model Comparison -- a separate mode.
Is the routing decision documented?+
Yes -- every Pilot decision lands in the audit log with detected category, selected model and rationale. This helps with compliance audits and model optimisation.