For Enterprise
The reasoning layer
your AI estate is missing.
Every request reaches the model cold — no classification of intent, no organizational context assembled for this specific request, no policy evaluated, no session attributed, nothing compounding. The models are running. The layer that makes them reason better, cost less, and answer for themselves never shipped.
grāmatr is that layer. Classification, context delivery, governance, cost attribution, and compounding intelligence — in the request path, before the model answers. Context delivery is the core: the right context, assembled just in time, so the model returns a better answer for fewer tokens. Governance is one of the five — see the governance architecture.
The spending is climbing. The ledger is empty.
The problem is not the models. Models work. The problem is what runs before them.
Your AI estate processes requests with:
- No classification of intent
- No organizational context assembled for the specific request
- No policy evaluated before the model responds
- No session attributed to an actor at the time it happens
- No signal compounding into institutional intelligence
The models generate responses. The responses are real. But they belong to no one, they are attributable to nothing, and they leave no audit trail you can give a regulator.
That is not a model problem. It is an infrastructure problem. One layer is missing from every request your organization runs.
Five functions. One request path. Before the model answers.
grāmatr operates in the window between your organization's intent and the model's response. Every request is classified in the request path — intent classification in under 100ms (~26ms typical), the full contextual packet assembled in ~135ms — before any model token is generated.
Classification
Intent is classified before anything else loads. The patent-pending pre-classifier routes each request to the right context, the right model endpoint, and the right policy in under 100 milliseconds. Your AI estate does not process requests blindly. It classifies them first.
Context delivery
The right organizational context is assembled for this specific request — just in time. Not a static system prompt loaded on every call regardless of relevance. Exactly what this request needs, retrieved from your organization's accumulated intelligence and delivered before the model responds. The model returns a better answer. The request costs fewer tokens. Not because the task simplified — because the model received what it needed to answer it well.
Governance
Policy is evaluated before the model responds — not as a moderation layer on the output, but in the request path, before the model sees the request. Every policy evaluation is logged. Every exception is recorded. The audit trail is the request path, not a post-hoc reconstruction of it.
Cost attribution
Every session is attributed to its actor at the moment of inference. Not aggregated after the fact. Not estimated from billing line items. Not assembled from log exports. The attribution record is written when the session occurs — which means the CFO can run the number at any point in the billing cycle and it is accurate.
Compounding intelligence
Every routed request trains the classifier on your organization's patterns. The system gets more accurate with use — not because it fine-tunes a shared model, but because the classification architecture learns what this organization does and routes future requests accordingly. The intelligence accumulates under your audit trail. It does not expire between sessions. It does not reset when the billing cycle turns.
Real applications. Same architecture. In production.
These numbers come from production, not a demo:
- Real production, not a test — from Anneal, the first application built on grāmatr, running in production on the platform, with CPP (Covington Place Partners) — the second application built on grāmatr, in active development — behind it. Full record on the proof page.
- The architecture under the conditions being sold — the audit trail from those sessions exists; it is the same ledger enterprise buyers are being asked to build, and it holds in production.
- A different category of vendor — a team that builds infrastructure and runs it in production before selling it stands apart in procurement, due diligence, and any auditor conversation about whether the architecture has been stress-tested.
Now consider what those numbers look like across an enterprise estate.
The numbers your CFO will run.
One architecture review produces the inputs for a budget-ready business case. These are the anchors.
Every deployment tier. Your security posture drives the decision.
Enterprise security requirements are not uniform. grāmatr supports every deployment tier from the same architecture. The governance layer, attribution records, and audit trail are identical across all four. The deployment boundary is yours to define.
| Tier | Deployment boundary | Availability |
|---|---|---|
| grāmatr Cloud | Managed multi-tenant — tenants isolated at the database row level, never commingled. SOC 2 Type I: targeting Oct 15, 2026; not yet certified. Anneal runs here out of the box. | Available now |
| Private cloud | Within your infrastructure boundary. Your network, your key management. Same architecture inside your environment. | September 1, 2026 |
| On-premises | Full local deployment. No data leaves your network. | September 1, 2026 |
| Fully air-gapped | On-premises with no external connectivity at all. | January 1, 2027 |
Licensing and install path ready by each date. SOC 2 Type I: targeting October 15, 2026 — not certified today. Book an architecture review to scope your environment ahead of availability.
Where we stand. What is next. Nothing inflated.
The compliance program runs as a versioned code repository — every control mapping, every evidence link, every policy version-controlled and tracked through pull requests. The evidence pack is not assembled the night before the audit. It exists continuously, maintained as code.
We will not claim certifications we do not hold. Current compliance status and the full evidence pack are available for enterprise due diligence under NDA — not as a PDF, as a working repository.
Deploy Anneal now. Build on grāmatr underneath.
If your organization needs an AI application in production in weeks, not quarters: Anneal is the first application built on grāmatr — running in production today. License Anneal on grāmatr Cloud — available now — and deploy in weeks. The private cloud, on-premises, and air-gapped tiers follow the availability dates in the deployment table above. Book an architecture review to establish the path. The intelligence layer underneath is the same one that runs your broader AI estate. You are not deploying a separate product — you are deploying the fastest path to a production AI application on the infrastructure you are already evaluating.
Ask about Anneal in your architecture reviewProcurement-grade questions. Direct answers.
What compliance certifications does grāmatr hold today?
The SOC 2 program is active, targeting Type I by October 15, 2026 and Type II by April 15, 2027 — after the required observation period following Type I completion. Every control mapping and evidence link is version-controlled as code, not assembled before the audit. GDPR and CCPA: aligned today. The evidence pack is available under NDA for enterprise due diligence. We do not claim certifications we do not hold.
Can we deploy on-premises?
Yes. On-premises (connected) has a committed availability date of September 1, 2026; the fully air-gapped variant — no external connectivity at all — has a committed availability date of January 1, 2027. Licensing and the install path are ready by each date. The full grāmatr architecture runs inside your network: classification, context delivery, governance, attribution, and audit trail, with no data leaving your boundary. Book an architecture review to scope your environment ahead of availability.
Does grāmatr train on our session data?
No. grāmatr trains on routing performance signals — classification accuracy, context relevance scores, and session outcomes — not on the content of your sessions. Your organizational intelligence accumulates within your deployment and belongs to you. It is not used to train shared models or surface to other tenants.
How does cost attribution work?
Every session is attributed to its actor at the time of inference — not reconstructed from log exports after the fact. Attribution records are queryable per-user, per-team, per-project, and per-time-range — and resolve down to the commit, the pull request, the story point, and the line of code. The CFO can run the number at any point in the billing cycle and it is accurate, because the attribution happened in the request path when the session occurred.
What is the integration surface?
For AI tools in the request path: grāmatr operates as the intelligence layer before the model call. Integration is a single baseURL change — no rip-and-replace, no new infrastructure. Point any OpenAI- or Anthropic-compatible tool at the grāmatr gateway:
# OpenAI-compatible tools (Cursor, ChatGPT, VS Code)
export OPENAI_BASE_URL=https://gateway.gramatr.com/v1
export OPENAI_API_KEY=<your-grāmatr-key>
# Anthropic-compatible tools (Claude Code, Claude SDK)
export ANTHROPIC_BASE_URL=https://gateway.gramatr.com
export ANTHROPIC_API_KEY=<your-grāmatr-key>For platform and API integrations: a standard REST API with API key authentication. Beyond the gateway, a grāmatr plugin marketplace — in private preview today, with a public marketplace on the roadmap — adds tool-native support for the specific tools your teams already run. The integration surface is intentionally small; the intelligence layer is what has weight, not the integration cost.
What happens to our data if we end the deployment?
Your data is yours. On cancellation, you receive a full export of your organization's data — classification records, governance history, session attribution logs, and the full audit trail. All raw data is permanently deleted from grāmatr systems within 30 days of export confirmation. A certificate of deletion is available on request.
How does the 99.9% SLA work in practice?
The 99.9% SLA covers the grāmatr classification and routing layer. The fail-open design means that if grāmatr experiences an outage, requests flow through to the underlying model without the grāmatr context layer — your AI workflow does not hard-fail. The SLA details and incident response commitments are in the service agreement.
Your AI estate is running. Make it yours.
Your organization has AI deployed. The layer that makes it governable, attributable, and auditable is one architecture decision. One conversation establishes what the deployment looks like in your environment.
Book an Architecture ReviewOr explore: Platform overview · Pricing · For Partners · System Status