Features / Platform
AI controls and spend
Know what it is doing, and what it costs
Every piece of AI work is metered and attributed. You can bring your own provider key, cap what may be spent, and stop it entirely without breaking support.
- Sort what arrivedtriagegpt-5.5-codexOpenAI
- Work out what they askedresponder_classifygpt-5.5-codexOpenAI
- Draft the replyresponder_authorgpt-5.5-codexOpenAI
- Write the fixauthorgpt-5.5-codexOpenAI
- Review the fixreviewerclaude-sonnet-4.6Anthropic
What it does
- Spend metered per action and attributed per product
- Bring your own provider key on Free, Scale and Enterprise
- Budgets, caps and a kill switch
- At zero the AI pauses, the inbox, help center and boards keep working
Without it
Handing AI your customers is a decision, not a toggle
One model for everything is wrong twice
Pick the strong model and you pay reasoning prices to work out that a message is a password reset. Pick the cheap one and it gets a production incident with a stack trace and confidently invents a cause.
The invoice is one number
Your AI bill arrives as a single line. You cannot tell whether it went on classifying easy tickets, on drafting replies people rewrote anyway, or on the two investigations that actually saved someone a day.
The model that wrote the fix says the fix is fine
You ask an AI to write a patch, and then you ask the same AI whether the patch is safe. It says yes. That is not a review, and everybody in the loop knows it.
How it works
Work is sorted by kind before a model is chosen
The router names five kinds of work and they are the ones in the code, not a marketing grouping: triage what arrived, work out what a customer message is, draft the reply, write the fix, and review the fix. Which model runs is a decision per kind of work.
The number that matters
5kinds
Fixed by the product, not by whoever is on shift.
Each kind goes somewhere it makes sense
A fast, cheap model sorts intent, priority and owner. A stronger reasoning model reads logs, traces and code. A coding model writes the patch. If a provider is down or refuses, the fallback chain takes the next one rather than dropping the work.
What that includes
- A fast
- cheap model sorts intent
- priority
- owner
A different vendor checks the fix
The reviewer must come from a different vendor than the model that wrote the code. Tore records the author model and vendor, the reviewer model and vendor, and whether the review was genuinely independent. If no independent reviewer is available, it records the review as unavailable instead of quietly passing it.
What that includes
- Tore records the author model
- vendor
- the reviewer model
Every call is metered before it is made
Tore reserves an estimated cost against the organization's budget for the period, calls the model, then settles the real cost with token counts and the price list version it used. Each entry carries the model, the provider, the feature and whether your own key was used.
What that includes
- Each entry carries the model
- the provider
- the feature
- whether your own key was used
At the ceiling it stops, and stops cleanly
The budget check and the kill switch are read before the model is called, not reconciled afterwards. Over the line, the run is refused and the investigation parks in a paused state; the inbox, the help center and the boards never depended on it and keep working.
What that includes
- Over the line
- the run is refused
- the help center
- the boards never depended on it
- keep working
What it means for you
You can see what the AI cost you, by kind of work
Spend is attributed to the thing it did, so the question stops being whether AI is expensive and starts being which part of it earns its keep.
No model marks its own homework
The check on a proposed fix comes from somewhere with no stake in the answer, and the record says which vendor did it. If nothing independent was available, you see that too.
A bad night has a ceiling
A loop that would have burned through your month gets refused on the call it cannot afford, not discovered in next month's invoice.
The honest limit: The routing runs today, but the screen where you would change it is marked preview and cannot be saved yet. Model per work type, the fallback chain, daily and monthly caps, the per-model cap, and choosing between Tore-managed keys and your own are all shown and none of them stick. Until they do, changes go through us.