Platform | Intent Router

Every request, the right model. Or no model at all.

Intent Router reads what each agent request is trying to do, then sends it to the cheapest model that can do it well. Security and budget checks happen in the same step, before any model is called.

Trust & Security for your entire Cloud and AI Ecosystem

Agents are scaling faster than budgets

A single agent task can make dozens of model calls, and most agents send every one of them to the same costly frontier model. Even reading a file or classifying a ticket.

of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls.

Two of the three causes now sit in one decision path.

Escalating costsUnclear business valueInadequate risk controls

Four demands now land on every enterprise running agents

01

A trust framework

Clear rules for agent actions, and safety nets for when something goes wrong.

02

Verified intent

What is the agent really trying to do, and does it hold the permissions it claims?

03

Accountable agents

Every agent recognized, bounded and tracked, and stoppable at runtime if it goes rogue.

04

Affordable agents

Agents loop, retry and call frontier models for routine steps at machine speed. Now it's a CFO question.

New · The CFO demand

Routing decides which model sees your data

It also decides what you pay and what your agents can do. So it can't be a black box that answers to someone else's business model, or a side system your security team never sees.

Today · Separate systems work against each other

Router picksbefore inspection
Model call paidcost already spent
Security inspectsafter the fact
Block or alertwork already wasted
Spend reconciledon next month's invoice

Operant · One decision, before any model is called

Read intent → safe? · which model? · within budget?Semantic Firewall · Intent Router · Spend Meter
BLOCK · $0CACHE · NO CALLROUTE · BEST FIT

One decision, before any model is called

ONE DECISION PATH · EVERY REQUEST FRONTIER REASONING FAST · LOW-COST SPECIALIST MODALITY YOUR PRIVATE MODELS INTENT CACHE Same intent already answered. No model call. BLOCKED Never reaches a model, so it never costs. Agent request Read intent Policy boundary Route Spend Meter Response
Read intent

Classifies each request by task, reasoning depth, latency needs and data sensitivity.

Check policy

Semantic Firewall blocks unsafe requests and keeps each agent inside its approved providers, models and regions.

Route

Picks the cheapest approved model that can do the job well, public or private. Or answers from cache.

Meter spend

Spend Meter enforces live budgets by agent, team and task while tokens are being used.

Three ways to spend less

One coding task can make 24 model calls. Here's where they go once each one is routed by intent.

One coding task · 24 model calls · Illustrative

3frontier calls|4never reach a model

Frontier Fast, low-cost Specialist Intent cache Blocked

Actual routing depends on your policy, your approved models and the intent of each request.

The right model

Frontier reasoning for architecture and judgment calls. Fast, low-cost models for file reads, extraction and triage.

The right modality

Call audio goes to a transcription model, scans to a document model, text to an embedding model. Not everything through one multimodal model.

No model at all

The intent cache matches on what's being asked, not how it's worded. Repeat requests are answered with no model call.

01 · The right model

Premium reasoning only where it's needed

Architecture decisionLarge refactor planFile readsLint fixesTest runsDocumentation FRONTIER FAST · LOW-COST

A coding agent: frontier for two steps, fast models for the rest.

02 · The right modality

Each workload to the model built for it

Today: every input through one frontier multimodal model

Call audio▶Transcription model
Screenshots, scans▶Image & document model
Uploaded documents▶Document model
Text to index▶Embedding model
Long video▶Async video model
Final resolution▶Reasoning model

Matched to each model's API contract and its input and output types.

03 · No model at all

Cache on what's being asked, not on how the words look

 
Similarity cachematches on wording
Intent cache · Operantmatches on what's asked
Look alike · different intentSummarize the Q3 revenue memo for the board.Summarize the Q3 revenue memo for the press.
HIT

Serves the board summary to a press workflow. Wrong answer, and data exposed.

MISS

Different audience, different sensitivity. A fresh, policy-checked call.

Look different · same intentHow do I reset my password?Locked out. What's the reset process?
MISS

The wording differs, so it pays for another model call.

HIT

Same intent. Served from cache with no model call, faster.

Cached answers are served inside the same security boundaries as live traffic.

Routing lowers the cost of each call. Spend Meter holds the total.

Live budgets by agent, team and task, enforced as tokens and dollars are used. Not reconciled on next month's invoice.

Spend Meter · live budgets · illustrativeLimit
Coding agents
$0Within budget
Support agents
$0Within budget
SOC agents
$0Within budget
Research agents
$0Within budget

One team hits its cap and gets throttled. The rest keep running.

Your policy draws the boundary. Intent Router optimizes inside it.

One catalog: 1,000+ public models plus the private, fine-tuned, platform-hosted and in-cloud frontier models you already run.

Eligible models are set before any routing happens. Intent Router never optimizes outside your boundary.

  • Public models (1,000+)
  • On your data platforms
  • Private and fine-tuned
  • Frontier, in your own cloud
  • Approved for this agentproviders · credentials · segmentation
  • If the request carries regulated dataonly where it's allowed to go

Sovereign AI, on your terms

Sovereign by policy, without giving up savings

Residency rules, approved regions and approved providers decide which models each agent can reach. Regulated data goes only where it's allowed. Everything else stays free to use cost-effective public models.

Works alongside your AI gatewayOpenAI-compatible interfacesNative provider APIs

Switch providers without rewrites

Agents reach Intent Router through OpenAI-compatible interfaces and native provider APIs. Change models or providers in policy, not in your agents' code.

Keep the gateway you already run

Your gateway keeps access, keys and failover. Intent Router adds routing, security and spend control on top, in one decision.

Learn about Semantic Firewall

Secure but unaffordable won't scale. Cheap but unsafe won't last.

Intent Router makes reducing spend and securing agents the same decision, on every call.

Why gateways, auto-routers and DIY aren't enough

Each one answers part of the problem. None of them decides safety, model choice and spend in the same step.

Approach 01

AI gateways

Good at
Access, keys, traffic handling, failover and observability.
Where it stops
Routes on rules someone set by hand. It can't read intent, so it can't tell an attack from a legitimate request.
With Operant
Keep it. Intent Router adds routing, security and spend control on top.

Approach 02

Hosted auto-routers

Good at
Sensible zero-setup defaults for individual developers.
Where it stops
Sorts prompts into fixed task buckets and picks what the crowd spends on, from the models its service hosts.
With Operant
Your request's intent and your approved pool decide, including private models.

Approach 03

Build it yourself

PROXYCLASSIFIERCACHEPOLICY
Good at
Full control in a prototype, plus a budget service.
Where it stops
Five systems to build and keep in sync, retuned for every model release, and a supply chain you own. A core open-source AI gateway was backdoored in 2026.
With Operant
One decision path, kept current as models ship. No tiers to retune.

Three questions on every request. Who answers them?

ApproachShould this request happen?Which model, or no model?Is spend held live?All three in one decision?
AI gateway
Hosted auto-router
Build it yourself
Operant
In the decision pathPartial or a separate add-onYou build and maintain itNot addressed

Category-level view based on public product descriptions. Capabilities vary by vendor and configuration.

One understanding of intent, every agent you run

Security decides what may happen. Routing decides how it runs best. Spend control decides what it may spend.

Frontier reasoning Fast, low-cost Specialist modality Private · in your cloud Intent cache Blocked

Coding agentsAcross engineering

ArchitectureRefactor planFile readsLint fixesTest generationDocs

Firewall blocks malicious code suggestions and secrets in prompts · Spend Meter budgets by team and project

Customer supportVoice, chat, documents

Call audioScreenshotsUploaded files"Where's my order?"Password resetResolution

Firewall keeps personal data from models that shouldn't see it · Spend Meter caps cost per case

Security operationsAlert triage, 24/7

EnrichmentClassificationHigh-severity caseInstruction planted in logs

Firewall stops instructions planted in log data and alert payloads · Spend Meter keeps 24/7 triage predictable

Document processingFinance, legal, claims

Scanned pagesField extractionRisky clausesExceptions & disputes

Firewall keeps regulated data on residency-compliant endpoints · savings grow with every document

Research and analystsInternal and external sources

Deep analysisQuick lookupsSummariesRepeat lookups

Firewall keeps confidential plans and client data inside · Spend Meter budgets each research task

Start anywhere. Converge on one decision.

Start with routing, security or spend control, or run all three from day one.

FAQ

Does Intent Router replace our AI gateway?
No. It runs alongside your gateway and adds routing, security and spend control in a single decision.
Will cheaper models hurt output quality?
Each request goes to the cheapest model that can do it well. Architecture and judgment calls still go to frontier models.
Who decides which models an agent can use?
You do. Your policy sets approved providers, models and regions before any routing happens. Intent Router never picks outside that pool.
Do we need all three controls on day one?
No. Start with routing, security or spend control and add the rest later. If you already run Semantic Firewall, routing plugs into it.

Scale agents without scaling the bill.

Start with routing, security or spend, or run all three from day one. Part of the Operant 3D AI Defense Platform.