# All your company's AI goes through one door. > Atlasberg Platform governs every AI request in your company: identity, policy, DLP, cost and an immutable audit trail. Runs in the cloud, in your VPC, on-premises or fully air-gapped. Atlasberg Platform is the control layer between your organization and every AI model. Every request authenticated, filtered, routed, metered and written to an immutable audit trail. Inside your perimeter, under your rules. ## AI is already inside your company. It just never came through the door. Teams buy tools on their own. API keys circulate in spreadsheets. Customer data leaves the network with no one keeping record. When the regulator asks who sent what, to which model, on which day, today the answer does not exist. It is not a tooling problem. It is an architecture problem. ### You don't know what leaves. Prompts carrying personal data, contracts and source code travel outside with no policy and no record. ### You don't know what it costs. The invoice arrives at the end of the month, aggregated, with no owner, no project, no ceiling. ### You don't know who used it. A shared key is nobody's identity. Without identity, there is no accountability. ### You can't prove anything. Audits demand evidence. Application logs are not evidence, they are notes. ## Eight controls. One single layer. Zero changes to your application. The applications you already run talk to Atlasberg Platform as if it were the AI provider itself. Swap the URL, and governance starts to exist. ### Governance Corporate identity via SSO, virtual keys per department, team and user. Every request has an owner before it leaves the network. ### Audit Immutable hash-chain trail, verifiable by a third party. It records the decision, never the raw content. Evidence that survives an auditor. ### Guardrails Prompt injection, exfiltration and forbidden topics blocked before the first byte of the response. Prevention, not an alert after the fact. ### Brazilian DLP CPF and CNPJ with check-digit validation, cards, emails, phone numbers and confidential files. Mask or block, by policy. ### Cost management Budgets per team, project and model, with ceilings that actually cut. Token cost attributed to the right cost center, same day. ### Multi-LLM routing The right model for each use case, with automatic fallback. Switching providers becomes a policy decision, not a migration project. ### Governed contexts Each team sees only the knowledge base it is allowed to see. Context is sensitive data and gets the same control as a credential. ### Policies per team Legal, engineering and support have different rules. One policy per department, enforced at the same point, no manual exceptions. ## Pay only for the context that still matters. Before each request leaves for the provider, the gateway decides, call by call, which tool results the model will still need. The rest goes. It does not summarise and does not rewrite: whatever a person or the assistant wrote stays untouched. Works with any model, without changing a single line of your agent. Source: our own measurement, published on 23 Sep 2026, on our own traffic. Your number may be higher or lower; measurement mode exists for that. ### 17.01% measured savings Across 30 real agent sessions and 6,892 turns. Real traffic, published assumptions. ### US$ 0.26 to save US$ 101.58 Total cost of the classifier in the measurement. Every dollar of engine gave back 390. ### 98.8% of requests with zero added latency Pruning fires on a threshold and gives up when it is not worth it. Only 83 of 6,892 turns waited 330 ms. ### 22x gain over loss in the worst case Everything that goes is recoverable. Even re-fetching every excerpt cited later, US$ 97 of balance remain. ### The input bill drops the same day No migration, no new prompt, nobody to retrain. Turn it on in the gateway and the history reaches the provider lighter. ### Measure before you switch it on Measurement mode computes what compaction would save on your traffic without pruning anything. The 17% on this page came straight from it. ### Two engines, one protocol Jev, remote, sees the whole conversation. Laya, local, runs on your machine and no excerpt leaves the network. Switching is changing the address. ### No request breaks to save a token It ships off. Turning it on is deliberate. If the engine fails, the original history goes through. The worst possible mistake costs one repeated tool call. ## Six steps between your team and the model. All of them yours. Every AI request in the company follows the same path, in order. No side doors, no exceptions. Policy decisions happen in memory, before the first byte of the response. Not in an end-of-month report. ### Identity Corporate SSO and virtual keys. The request is born with a name, a department and a project. ### Ingress DLP Confidential files never upload. The block happens before the upload, not after. ### Content DLP Personal data is masked or blocked according to the department's policy. ### Guardrails Injection, exfiltration and forbidden topics are intercepted before the response begins. ### Routing Policy picks the model, the provider and the fallback. Local LLMs are a first-class option. ### Metering and audit Tokens, cost and decisions enter the immutable trail. The content does not. ## Your data does not have to leave your network to have AI. A single, self-contained Docker artifact. The same Platform across all four topologies, no stripped-down edition. Your data never leaves your perimeter. Prompts, files and audit trail are processed inside your infrastructure, under your keys. Priced by capacity, never by seat. Governance that charges per user punishes the company for adopting AI. Ours charges for the volume that flows, and everyone is in. ### Atlasberg Cloud Go live today, no infrastructure required. ### Your VPC Your account, your network, your access control. ### On-premises Your own datacenter, integrated with your SSO. ### Air-gapped No internet, with a signed offline license and local LLMs. For those who cannot negotiate this. ## Built for companies that answer to regulators. LGPD, AI Act, BACEN resolutions and SOC 2. Atlasberg Platform does not promise compliance: it produces the evidence compliance demands. ### Verifiable record Every policy decision signed and chained. Any third party can verify the integrity of the trail. ### Residency under your rule Retention, location and access to data defined by you, not by the vendor. ### Audit-ready reporting Exportable evidence in the format the auditor asks for, with no internal project to assemble it. ## The future of the corporation will not be measured in headcount. It will be measured by the density of intelligence of each person. Small teams with governed AI will outperform entire divisions without it. We build the layer that makes this possible without giving up control. Excellence has no size. It has impact. ## The questions every security committee asks. ### Do I need to change my applications? No. The Platform exposes an API compatible with the market standard. Your application swaps the provider URL for the Platform URL and keeps working, now with policy, metering and an audit trail. ### How much latency does it add? Milliseconds. Policy decisions happen in memory, on the request path, before the first byte of the response. There is no queue and no batch processing. ### Which providers are supported? OpenAI, Anthropic, Google Gemini, Azure OpenAI, AWS Bedrock and any compatible local endpoint, including Ollama. All behind the same policy. ### What exactly gets recorded? Who made the request, to which model, under which policy, which decision was taken, how many tokens and at what cost. The raw content of the prompt and the response never enters a log. ### Does it work without internet? Yes. The air-gapped topology runs fully isolated, with a signed offline license and local LLMs. It is the same artifact as the other topologies, with nothing stripped out. ### How does this satisfy LGPD and BACEN requirements? The DLP identifies and treats Brazilian personal data before it leaves the perimeter, and the immutable trail produces the evidence of treatment that supervision demands. Compliance becomes a by-product of the operation, not an annual project. ### What if the license expires? Renewal is a signed license file: importing the new one takes seconds, with no reinstall. And the console warns you well before anything expires. ### How is it priced? By capacity and modules, never by user. The whole company gets in with no per-seat cost. Tiers are defined in the first conversation, together with the topology. ## See the Platform running your traffic, not our slides. Thirty minutes with the engineering team. You bring a real use case and leave with the design of the topology, the policy and the cost. No SDRs, no email sequence. The first conversation is with the people who build. ## More from Atlasberg - [How the platform works](https://atlasberg.com/docs) - [Live demo in your browser](https://atlasberg.com/demo) - [Context compaction: 17% measured, not 60%](https://atlasberg.com/blog/compactacao-de-contexto-17-por-cento) - [The Atlasberg team](https://atlasberg.com/time) - [Talk to the engineering team](https://atlasberg.com/contato) - [Partner program](https://atlasberg.com/parceiros) --- Source: https://atlasberg.com/ Company: Sobimann Tecnologia da Informacao LTDA (Atlasberg), Porto Alegre, RS, Brazil. Contact: contact@atlasberg.com - answered within one business day. Index for agents: https://atlasberg.com/llms.txt