Azure OpenAI puts OpenAI's models in an EU region under a Microsoft enterprise contract. That solves latency and procurement. It does not make the provider European, and it leaves the governance layer for you to build. Akumi is the governed EU-sovereign platform: EU-resident models, a fail-closed egress guard, a PII firewall, and cited retrieval, behind the same OpenAI-compatible call.
Azure OpenAI is the compromise most enterprises reach first, and it is a reasonable one. You get OpenAI's models inside a Microsoft enterprise agreement, deployed to a European region, on paper your organization has already signed. For a lot of workloads that is enough.
It stops being enough when the objection is about the provider rather than the postcode. Choosing an EU region changes where the compute runs; it does not change who processes the data or under which jurisdiction. And the governance layer, pseudonymization, retrieval that isolates one customer from another, an audit trail that answers where a specific request ran, is still yours to assemble out of separate Azure services with separate bills. Akumi answers on EU-sovereign models by default and ships that layer as part of the same request.
An EU-sovereign AI platform behind one OpenAI-compatible endpoint. EU-resident models answer by default and non-EU routing is blocked at a guard that fails closed. The PII firewall, guard, governed retrieval, response cache and per-request observability run on the same call, each metered separately, with residency recorded per request in a metadata-only audit trail.
Microsoft's hosted deployment of OpenAI's models, provisioned as a resource inside your Azure subscription and pinned to a region you choose. It brings the frontier line-up, Microsoft's enterprise agreement and support, and integration with the rest of Azure, including Azure AI Search if you want retrieval and Azure Monitor if you want telemetry.
| Dimension | Akumi | Azure OpenAI |
|---|---|---|
| Model provenance | EU-sovereign models served in the EU, with external providers available only under an explicit policy. | OpenAI models, hosted by Microsoft. The region is European, the provider and the model are not. |
| Non-EU egress | A fail-closed egress guard. External routing does not happen unless the request is firewalled or a recorded acknowledgment exists. | Region pinning and contractual commitments. There is no request-level guard that fails closed. |
| Personal data before the model | PII firewall on by default: personal data is pseudonymized before egress and restored in your response. | Not included. Content filtering exists, but pseudonymize-and-restore is yours to build. |
| Retrieval | Conversation memory and your documents share one knowledge graph, so an answer follows the link from a user to their organization to the document that answers them. Each collection is its own graph partition, namespaced to your organization and workspace before the query is built, and every response carries the sources it used. | Azure AI Search: a separate service to provision, wire up, tune and pay for, with the tenant isolation between customers left to your own application code. |
| Model choice | EU-sovereign models plus external providers behind one policy, one endpoint and one bill. Swap models with a string. | OpenAI models only. A second provider means a second contract, a second integration and a second invoice. |
| Compliance evidence | One DPA, published sub-processors, and a per-request record of the model, provider and region that served it. | A Microsoft enterprise agreement, with the per-request residency evidence left for you to assemble. |
| Getting started | Self-serve signup, published pricing, one base_url, first request in minutes. | An Azure subscription, resource provisioning, and in many organizations a quota request. |
Region pinning is a real control and it solves real problems: latency, some data-locality commitments, and a box on an internal form. What it does not do is change the identity of the processor. Personal data sent to a US-headquartered provider is still a transfer requiring a lawful basis, a data processing agreement and standard contractual clauses, and it stays a Schrems II discussion however European the endpoint address looks.
If your DPO's objection is that the data reaches a US provider at all, an EU region does not answer it, and no amount of configuration will. Akumi removes the transfer from the default path instead: EU-sovereign models answer the request, and reaching anything else requires an explicit decision enforced by a guard that fails closed rather than a setting that can be left wrong.
| Dimension | Akumi | Azure OpenAI |
|---|---|---|
| What the region controls | Where EU-sovereign models run, with the provider EU-based too. | Where Microsoft runs OpenAI's models. The provider is unchanged. |
| Transfer paperwork | Not triggered on the default path. | Still required for personal data, region notwithstanding. |
| Enforcement | A fail-closed guard in code. A misconfiguration refuses the request. | Contractual commitments and a region setting. |
Azure OpenAI is a completion endpoint. A production application needs more than that, and on Azure each piece is a separate service you provision, wire together, tune and pay for: Azure AI Search for retrieval, Azure Monitor or Application Insights for telemetry, your own code for pseudonymization, and whatever you build for prompt-injection defence. The integration work is real, and so is the ongoing ownership.
Akumi runs those as modules on the same request. Retrieval, the PII firewall, the guard, the cache and the trace all fire on one call, each metered on its own line so you can see what governance costs without reconciling five invoices. The isolation question is the one worth checking closely: Azure AI Search has no built-in multi-tenancy, so keeping one customer's documents out of another customer's answers is your application's filtering logic to write and to get right.
| Dimension | Akumi | Azure OpenAI |
|---|---|---|
| Retrieval | Included: a governed knowledge graph, partitioned per workspace before the query is built. | Azure AI Search, provisioned, wired up, tuned and billed separately. |
| Tenant isolation in retrieval | Enforced in code from the authenticated caller, not from anything the request supplies. | Your application's filters or an index per tenant, designed and maintained by you. |
| Personal data | PII firewall on by default, pseudonymized before egress and restored after. | Content filtering exists. Pseudonymize-and-restore is yours to build. |
| What you reconcile | One vendor, one bill, usage attributed per request and per module. | Several Azure services, each with its own meter. |
Getting to a first Azure OpenAI call is rarely just a key. There is a subscription, a resource to provision, a region and deployment name to pick, role assignments, and in many organizations a quota request that waits on someone else's queue. None of it is hard; all of it is calendar time, and most of it happens before you learn anything about whether the thing works for your use case.
Akumi is self-serve with published pricing. Add a payment method, create a key, change one base_url, and you have a real answer from a real model against your own workload the same afternoon. If it does not fit, you have lost an afternoon rather than a sprint.
| Dimension | Akumi | Azure OpenAI |
|---|---|---|
| Before the first call | Signup, a payment method, a key. | A subscription, a provisioned resource, role assignments, often a quota request. |
| Knowing the price | Published per plan and per meter. | Calculable, once you know which resources and tiers you need. |
| Trying a second model | A string change behind the same endpoint, with router fallback. | A second deployment, and a second provider entirely if it is not an OpenAI model. |
It helps with data locality but it does not settle the transfer question. The processor is still US-headquartered, so personal data sent to it needs a lawful basis, a DPA and standard contractual clauses. Akumi answers on EU-sovereign models by default, so that transfer does not occur on the default path, and each request records the region that served it as evidence.
For the chat completions surface, yes. Akumi is OpenAI-compatible, so a client pointed at Azure OpenAI moves with a base_url and a key. There are also first-party SDKs for PHP, TypeScript and Python and an MCP server.
Yes, under an explicit routing policy. External models are off by default; reaching one requires a firewalled payload or a recorded compliance acknowledgment, and the egress guard fails closed so it never happens by accident.
A governed knowledge graph built into the platform. Conversation memory and your documents share one substrate, each collection is its own partition namespaced to your organization and workspace before the query is built, and every response carries the sources it used. There is no separate service to provision or isolate.
The partition key is derived server-side from the authenticated caller, never from an identifier the request supplies, so a request can only ever address partitions inside its own organization and workspace. This matters because Azure AI Search has no built-in multi-tenancy, leaving that boundary to your application code.
No. Akumi is self-serve: a payment method, a key and one base_url. No subscription, no resource provisioning and no quota request.
No, and we do not claim to be. Akumi is built around GDPR and EU AI Act obligations with DORA and NIS2 supported, a DPA governs our role as processor, and the sub-processor list is published. ISO 27001 is in progress and not yet certified, which the trust page states plainly.
From the audit trail. Every request records the model, the provider and the region that served it, as metadata only, so residency is answered per request rather than by pointing at a region setting.
Change one base_url and send a real request. You will know within an afternoon whether it fits, which is faster than most organizations can provision a resource group.
We use privacy-friendly, EU-hosted analytics to understand how Akumi is used. Accept analytics cookies, or continue without them. See our cookie policy.
