Lock your automation to one model vendor and their price list becomes your cost structure — and you'll pay frontier rates for extraction work a cheap model does fine. Here the model is a layer: six providers plus Azure OpenAI on your own keys, chosen per agent, swapped without rebuilds, and metered to the message under hard budget caps. Enterprise mandates — an existing Azure agreement, data residency, BYOK — are the design, not the exception.
In a pack, model choices arrive made: every agent ships with a working default you can swap, on your own keys.
See Solution Packs →
Model lock-in compounds quietly: every workflow you build on a hard-coded model is another workflow that vendor's pricing, outages and roadmap now own. And most enterprises carry mandates — an existing Azure OpenAI agreement, data-residency terms, keys-stay-ours — that most AI products simply can't meet.
So intelligence here is a layer, not a dependency. Providers connect on your keys, each agent picks its own engine, a swap is a setting, and every call is metered to the message with hard caps above it. The model decision stays a decision — revisable, priced, and yours.
Pick a job. Each keeps the choice yours: any provider, per agent, on your own keys, with the meter and the ceiling attached.
A finance department's invoice pipeline has two very different jobs in it: pulling fields off PDFs, and deciding what to do about exceptions. One price should not fit both.
Field extraction is high-volume, low-judgment work. That agent is configured with a fast, low-cost model — its accuracy needs are structural, not creative.
The exception-handling agent — mismatch reasoning, escalation decisions — runs a frontier model, because being wrong there costs more than the tokens do.
Per-message cost accounting shows each agent's spend separately, so the mix is tuned on evidence.
A cheaper model with comparable quality just launched. On most platforms that's a migration ticket. Here it's a setting change on the agents that should move.
If the provider is connected, its new model is available to pick. A new provider is one key away.
The extraction agents switch models; their inputs, tools, tables and permissions are untouched.
Next window's per-model cost roll-up shows the actual saving in dollars — or tells you to switch back.
The company already has an Azure OpenAI agreement with negotiated terms and a residency commitment. The mandate: AI traffic runs on that agreement or it doesn't run.
IT connects the org's Azure OpenAI endpoint, API version and deployments — an org-scoped connection with the org's own key.
Deployments appear in the model menu like any provider's models; workers run on them without special-case plumbing.
Traffic stays on the company's Azure tenancy and region. Nothing was renegotiated to adopt one more product.
A provider is degraded and part of the workforce runs on it. You need those agents on a healthy engine now — not after a re-engineering sprint.
Per-model accounting shows exactly which agents and how much traffic ride on the degraded provider.
The affected agents switch to comparable models on a healthy connected provider — a setting change per agent.
Workflows, tools and permissions never moved. When the provider recovers, moving back is the same one change.
The fear isn't the average month — it's the outlier: a looping agent on a frontier model, discovered on the invoice. The ceiling should exist before the loop does.
Hard daily and monthly caps per workspace, employee or team — governance-enforced, not advisory.
Period-end forecasting projects spend against the cap and surfaces the breach date early.
A cost threshold trips and the circuit breaker pauses the worker mid-loop. The invoice stays boring.
Connect providers with your own keys once; every agent, team and employee runs on the model you choose — swappable, metered, capped.
Six properties that keep the intelligence layer yours — the providers, the keys, the choice, and the bill.
Hard-code one vendor's model into your automation and you've signed up for its pricing, its outages and its roadmap, with no exit. Here the model is a configuration layer: connect OpenAI, Anthropic, Google, Mistral, DeepSeek and xAI — plus Azure OpenAI as its own connection type — and every model becomes pickable across the workspace. No single vendor ever owns your intelligence layer.
Paying frontier prices for extraction work is the quiet waste in most AI bills. Model choice here is per agent, not a global default: the agent doing document extraction runs a fast, cheap model; the agent making the judgment call runs a frontier one. Both live in the same workflow, and each choice is one setting on that agent — mixing providers freely across a team.
Because the model is configuration, a swap is not a migration. When a cheaper or better model ships, point the agent at it — the workflow, tools, tables and permissions stay exactly as they were. A provider outage works the same way in reverse: move the affected agents to a comparable model on a healthy provider and resume. Model churn stops being your re-engineering problem.
Keys are the organization's own. A connection is scoped to the org or to a person, and at call time the platform resolves the exact configured connection first — then org-level keys, then provider-level, then environment. No reselling, no markup, no pooled tenant key: your spend lands on your vendor agreement, in your region, under your negotiated terms. For enterprises with an existing Azure OpenAI agreement, that agreement — endpoint, deployments and all — is the connection.
Cost control starts with knowing where the money goes. Every LLM call carries its token counts and USD cost at the message level — not a monthly invoice surprise, a per-message record that rolls up by session, worker and model. "What are we spending on AI" stops being a guess, and "which steps belong on a cheaper model" becomes something you read off, not estimate.
Metering shows the spend; governance caps it. Hard budget limits — daily and monthly, per workspace, employee or team — sit on top of the per-message accounting, with period-end forecasting and circuit breakers that pause a worker when cost or error thresholds trip. A runaway loop on an expensive model burns to a ceiling, not through a quarter. That's the governance layer, applied to the model bill.
An AP pipeline runs extraction and judgment as separate agents on separate models — chosen per agent, on the org's own keys. Here's how the system behaves across a quarter.
OpenAI, Anthropic, Google, Mistral, DeepSeek and xAI behind one menu — no model hard-coded into any workflow.
A first-class connection type: your endpoint, API version and deployments — your existing enterprise agreement, honored.
Each agent picks its own model. Cheap for extraction, frontier for judgment, mixed freely in one team.
A model change is a setting. Workflow, tools, tables and permissions are untouched by the swap.
Org- or personal-scoped keys, resolved configured-connection-first. No markup, no pooled tenant key.
Token counts and USD cost on every call, rolled up by session, worker and model.
Hard daily and monthly caps with breach forecasting; breakers pause a worker when thresholds trip.
Model calls happen inside runs that are logged, sanitized and pinned to the config version they executed under.
End-to-end recruitment: screen candidates, draft outreach, schedule interviews.
Accounts-payable automation: OCR-extract invoices, run 3-way matching, route for approval.
Outbound prospecting engine: ICP → TAM → contacts → personalized messages.
SaaS customer support team: triage, technical, billing, onboarding with collaborative routing.
Insurance claims processing: intake FNOL, assess risk, draft settlement recommendations.
Healthcare RCM: categorize denials, detect payer patterns, draft appeals.
We use analytics cookies to see which pages help and which don’t. Nothing loads until you choose. Cookie Policy
