control · cogshift enterprise · ai routing & spend control
Control the AI.
Understand the change.
Seven in ten requests don’t need the expensive model, and few organisations can say which. Cogshift is a gateway between your people and the providers you already pay: a small judge reads each request and routes it to the cheapest model that can do it, under route names that never move, with budgets, policy and identity under your administrators’ control.
And because every prompt passes through it, you get eyes into the largest change to office work in a generation: adoption, behaviours and changing patterns of work as your organisation builds a new AI culture, the spend being wasted, and whether it is work, personal, or something you have banned. Never a prompt read.
three ways to buy ai
Three ways to buy AI.
Pay the strongest model for everything. Pay the cheapest model for everything. Or decide on every request.
$4.50 per million tokens, every request
- Capable on every request, including the hard ones.
- You pay frontier prices for routine work a cheaper model would finish just as well.
Overpaid on the routine majority
$0.18 per million tokens, every request
- Cheap on every request.
- The hard minority is answered by a model that cannot do it, and nobody can tell which requests those were.
Wrong on the hard minority
$1.48 blended, at a seven-in-ten cost-effective split
- Routine → cost-effective model, at $0.18.
- Hard → strongest model, escalated, at $4.50.
- You decide the routes, budgets and providers; nothing changes at a desk.
Right model for each request
understand the change · adoption · behaviour · insight into use
The biggest change to office work since the electronic office. We give you eyes into it.
Understand the culture shift as it happens. Your organisation is adopting AI faster than anyone can survey, and Cogshift sits on the line every prompt travels, so it can tell you what is really happening: who has taken it up, what they use it for, what habits are forming, whether the training is landing, and where the money is going. Every prompt tagged as it passes, above spend, before the invoice.
Every prompt tagged by context.
A parent-and-leaf taxonomy of topics: engineering, then code review; sales, then proposal drafting. Each prompt lands under work, personal, or a banned topic, so you see the shape of the work and the share that isn’t.
Adoption by department, team, person, role and time of day.
Which teams lean on it, which haven’t started, which roles have made it a habit, and when in the day the work happens. Adoption as a curve you can watch, not a survey you commission.
Habits, patterns, and the wasted spend inside them.
Who reaches for the frontier model for a two-line edit. Which team spends most on the routine work the judge would have sent down a tier. Where a model is being used as a search engine. Behavioural insight into LLM use, with the cost of each habit attached.
Is the training taking effect?
Run a session on prompting, or on which tool for which job, and watch the behaviour change: the topics a team reaches for, the habits it drops, the roles that moved and the ones that didn’t. Who needs the next session, and which team should be teaching it. Skills adoption, measured in use rather than attendance.
Work, personal, banned: a number, not a suspicion.
The share of spend going to personal or unrelated use, by team. Banned topics surfacing as a topic and a count the day they appear. Credential-shaped content screened on the way out. The evidence you hand an auditor is the evidence you looked at.
adoption by department
share of staff active this month
what it is used for · engineering
- engineering100%
- ├─ code review31%
- ├─ debugging24%
- ├─ documentation18%
- ├─ test writing11%
- └─ other16%
parent and leaf topics, tagged per prompt
work, personal, banned
- work 89%
- personal 8%
- banned 3%
share of spend, this scope. banned topics surface the day they appear.
training effect · prompting workshop, week 6
routine prompts sent to the right tier, before and after the session
the boundary
Insight, not surveillance.
Leadership sees topics, totals and trends, never prompt content. A model does the tagging; no person reads a prompt. That is what lets the people team, the DPO and the CFO all say yes to the same dashboard.
- what leadership sees
- topics, totals, trends, by scope
- what leadership never sees
- a prompt, a completion, a document
- who reads a prompt
- nobody. a model tags it; people see the totals
- small teams
- rolled up, so no individual is singled out
control the ai
Every request takes one managed road.
A gateway between your people and the providers you already pay. Route names that never move; behind them, everything is yours to decide.
- Route names that never move. Repoint a model or vendor at 9am; nobody reconfigures a tool.
- Your providers, your contracts. Any provider, added from the dashboard; a self-hosted model without a rewrite.
- Resolution by scope. The same name resolves differently by company, department, team or person.
- Budgets that hold. A ceiling per person and team, checked before a token is spent.
- Single sign-on. Your directory, your leavers process, your access reviews.
- Credential screening. Key- and password-shaped content screened on the way out.
- Failover. A provider outage routed around behind the same name.
- Canary and A/B. Five percent of a route to a new model; promote on your own data.
Every decision here is made centrally and is invisible at the desk. Everything included is listed below.
How a request is routed
Routine → cost-effective model
$0.18 · 96% less- Summarise a document
- Edit a function
- Draft an email, rewrite a paragraph
- Answer from a policy
Hard → strongest model, escalated
$4.50 · only where the judge escalates- Debug a failing system
- Plan a strategy
- Reason across many documents
- Review a contract
behaviour has a cost
Four things multiply into your AI bill. One dial is yours.
Don’t slow AI adoption to control the AI bill. Control the model decision, and see the habits that drive it.
scale ↑
grows with adoption. good.
price per request ↓
the only stage you can actually move
unit price
set by the provider
- scale is not the problem
- More people using AI is the outcome you paid for. Nobody should be throttling adoption to control a bill.
- list price is not yours
- The provider sets it. You can change vendor, but you cannot argue the number down.
- the middle stage is yours
- Which model answers which request is a decision. Today it is made by habit; Cogshift makes it per request.
what moving one dial is worth
- 26×
Price gap, frontier against cost-effective, same provider.
- 67%
Less model spend when seven requests in ten answer on the cheap tier.
- 568×
Widest price gap between two live models in the catalog.
- 0
Config changes on staff machines when a route is repointed.
Published list prices, Cogshift catalog, 30 August 2026, at a 3:1 input-to-output token mix. Widest pair GPT 5.5 Pro $67.50 against GLM 5.3 Flash $0.12 (unrounded $0.11875). List prices and a modelled mix, not customer results.
a route name’s year
The name is static. Everything behind it moves.
The names your staff configure, code, draft,
review, never change. Behind them, your administrators change
models, vendors, tiers and budgets all year, and every request is judged,
costed and reported as it passes.
- janConfigured onceStaff type one name into their tools. No desk ever changes again.
- marResolved for Legal onlyThe same name resolves to a stronger tier for one department. Staff still type one name.
- mayCanary on 5% of one routeA new model on live traffic; compare, then promote behind the same name.
- julRepointed at 9amThe vendor behind the name changes. No configuration edited, live within a minute.
- sepFailed over to a second providerAn outage routed around automatically. Zero desk changes, nobody noticed.
- novBudget ceiling reachedVisible before it was reached, adjusted in one action.
- decSpend by team, model and taskReported as it happened, not when the invoice arrived.
Leave your details. We’ll be in touch.
Cogshift Enterprise is in a free closed trial for organisations. Leave your contact details and we’ll get back to you to set it up: your own tooling, pointed at a Cogshift route, with routing, cost and the dashboard on one screen, and your own split in the first month.
- Free for the length of the trial. No card, no commitment.
- Your providers, your contracts, your keys. Any provider your administrators add.
- One binary on your infrastructure, or run by us.
- Your own figure from your own traffic, in the first month.
This sends us one email with what you typed. It is not stored in a database and there is no mailing list. A person reads it and replies.
Everything included — 16 capabilities
cost
Compound routes
26×A judge model reads each request and picks the cheapest capable tier; a cost-effective model answers the routine majority, and a stronger model is held in reserve for the requests that genuinely require it. Per request, not per person or project.
Canary and A/B testing
5%Send five percent of a route to a new model, or run two side by side on live traffic, and compare cost, acceptance and speed on your own work before committing. Decided on your data, not a vendor’s benchmark.
Individual and team budgets
holdsA spending ceiling per person and per team, enforced before a token is spent, visible before it is reached, adjustable in one action. A control, not a post-mortem.
Live cost reporting
liveSpend by team, model and task, at published rates, as requests happen. Not reconstructed from the provider’s invoice weeks later. You see the bill while there is still time to act on it.
control
Static route names
onceTools are configured once against code, draft,
review. Change the model or the vendor behind a name at 9am; it
is live for everyone within a minute, and nobody edits a configuration.
You control the routes
centralWhich models, which providers and which tiers serve your organisation are decided by your administrators. A change is one reviewed, reversible action. Individual users do not choose their own upstream, and do not need to.
Resolution by scope
4 scopesThe same route name resolves to different models by company default,
department, team or one named person. Legal’s draft can
resolve to a stronger tier than Support’s; one reviewer can be pinned to
a model for an evaluation.
Your provider contracts
allCogshift supports all providers. Administrators add the ones they already have contracts with, from the dashboard, with each provider’s published prices, sources and dates. A self-hosted model is a route configuration, not a rewrite.
Failover across providers
autoIf a provider is down, slow or rate-limited, the route fails over to a second provider or model behind the same name. Availability becomes a property of the route, not of any one vendor.
No reconfiguration of your estate
0Every staff tool points at a route name that never changes. Every decision on this page is made centrally and is invisible at the desk. No rollout to schedule, no rollback plan to write.
safety
Single sign-on
same dayAccess follows your directory. Joiners get access through the usual process, leavers lose it the same day, and your access review already covers AI. A static key inherits its owner’s identity.
Credential screening
on the way outContent shaped like an API key, a token or a password is screened out of outbound requests and reported, so a credential does not reach a provider in the first place. It screens credential-shaped content; it does not claim to catch every secret.
Audit lane
siemEvery administrative action, and every refused one, maps onto a closed catalogue and lands in one audit chain you can feed to your SIEM. “Who changed what” has an answer.
insight
Insight from tagging
every promptEvery prompt is tagged by context as it is routed: a parent-and-leaf topic, work, personal or banned, department, team, person, role and time of day. Leadership sees topics and totals, never prompt content.
Adoption and behaviour
trendsHow fast AI is being taken up, where, by which roles, at what time of day, and for what. The habits forming, the wasted spend inside them, and whether training changed any of it, as a trend you can watch rather than a snapshot.
How much of the bill is company work?
cfos ask firstWork, personal, banned: the split sums to the bill, by team. Banned topics surface as a topic and a count. No person reads a prompt to find out.
Part of Cogshift READ — Telltail · ROUTE — Personal · CONTROL — Enterprise