The enterprise AI platform between your data and a decision

BlackGust doesn't replace your systems. It connects to them, builds a shared model of the organization and gives AI agents safe, audited access to it. Every layer can be inspected, configured and deployed wherever your security policy requires: our cloud, your cloud, your servers or an air-gapped network.

  • DeploymentCloud · private · on-prem · air-gapped
  • ModelsOpenAI · Anthropic · local open models
  • AccessRead-only by default · SSO · full audit
  • First resultPilot in 4–8 weeks
Design principle

Your systems stay the source of truth. The platform reads them, connects them and lets AI agents act only within the rights you grant, with every step on the record.

Architecture

Enterprise AI architecture in five layers

Data moves left to right: from the systems you already run, through connectors, into one model of the organization. Agents work on that model and deliver results where people already work. Governance runs underneath every layer, not bolted on at the end.

01 · Sources02 · Connectors03 · Organization model04 · Agents05 · InterfacesERP · accountingCRMDatabases · warehousesDocuments · archivesEmail · chatTelephony · camerasAPICDC replicasETL · exportsOCR · documentsSpeech to textOne modelCounterpartyContractInvoicePaymentOwnerProjectDirectiveAnalystRecords clerkDeadline monitorRequest handlerModel routerfrontier · localExecutive consoleWeb appTeams · SlackVoicePlatform API⟂ GovernanceAccess rightsAudit logHuman approvalLimitsPII maskingQuality evaluation

Layer by layer

Bottom to top: from data sources to the people making decisions.

05 · Interfaces

Where people work

Executive console, web app, corporate chat, voice, and embedding into your own systems.

ExecutiveWebTeams · SlackVoiceAPI
04 · Agents

Who does the work

AI agents with a role, instructions and a set of permitted actions: find, calculate, draft, send for approval.

AnalystRecords clerkDeadline monitorRequest handler
03 · Organization model

What connects to what

Objects and relationships: counterparty → contract → invoice → payment → owner. One source of truth for every department.

ObjectsRelationshipsMetricsVector search
02 · Connectors

How data gets in

APIs, database replicas, exports and document processing. Synced on a schedule or in real time.

APICDCETLOCRSTT
01 · Sources

Your systems, unchanged

ERP and accounting, CRM, databases and data warehouses, documents, email, telephony, corporate chat, cameras.

SAPOracleDynamicsSalesforcePostgreSQLSnowflakeMS SQLExcel
⟂ Governance

The cross-cutting layer

Access rights, audit log, human approval, limits, masking of personal data.

RBACAuditHuman-in-the-loopPII masking
Organization model

One model of the organization, shared by every agent

Most AI pilots stall because each one sees a single system. The organization model links records across systems into business objects and the relationships between them. A question that crosses finance, legal and operations becomes one query instead of three exports and a spreadsheet.

signsbills undersettled bysourceowned byfeedsworks inassigned toaccountable forCounterpartytax ID · rating · limitsContractterms · amount · datesInvoiceamount · due dateSource documentscanned PDF · page 4Contract owneremployee · rolePaymentstatus · overdue daysDirectivedeadline · assigneeDepartmentbudget · headMetricoverdue receivables
Example question · Example · illustrative data

“Which overdue payments sit on contracts owned by my department?”

The highlighted path is how the agent answers: department → owners → contracts → invoices → payments. Each step is a link you can open.

Objects
Business entities your people already talk about: counterparties, contracts, assets, cases, citizens' requests, projects. Defined with you, not imposed from a template.
Relationships
Who signed what, which invoice belongs to which contract, who owns the follow-up. Links come from keys in your systems and from documents read by OCR.
Metrics
Overdue receivables, cycle time, plan versus actual. Calculated once, on the model, so every department sees the same number.
Search
Vector search over documents and correspondence, tied back to the objects they mention. Answers quote the page they came from.
Permissions
Rights are applied to objects and fields. A user, and any agent acting for that user, sees only what the directory allows.
Modules

AI modules on one data model

We start with one module aimed at the most expensive problem. Later modules plug into the model that's already built, so each one ships faster than the last.

First module

Executive

A leadership console: plain-language questions, a daily digest of deviations, tracking of directives, answers linked to their sources.

Typical first use
  • Morning digest of what moved off plan
  • Directive tracking with owners and deadlines
Operations

Operations

Agents for routine processes: reconciliations, requests, deadline tracking, drafting letters and reports, task routing.

Typical first use
  • Supplier reconciliations
  • Request intake and routing
Documents

Documents

Reading scans and PDFs, extracting key fields, checking contracts against a checklist, searching the archive.

Typical first use
  • Contract review against your checklist
  • Archive search with page citations
Knowledge

Knowledge

A corporate knowledge base of policies, procedures and regulations. Staff get answers quoted from the source document.

Typical first use
  • Policy and procedure assistant
  • Onboarding answers for new staff
Voice

Voice

Speech recognition and synthesis in major languages, including English, Spanish, Portuguese, German, French, Chinese, Japanese and Korean. Voice agents for contact centers, hotlines and mobile apps.

Typical first use
  • Call summaries and quality review
  • First-line voice agent for a hotline
Vision

Vision

Video analytics on your existing cameras: people counting, zone occupancy, process and safety monitoring.

Typical first use
  • Safety-zone monitoring
  • Queue and occupancy counts
AI agents

An AI agent is an employee with instructions and limits

An agent isn't a smart chat. It has a role, access to specific data, a list of permitted actions and rules for when to ask a human. Its configuration is readable by the process owner, the CISO and the auditor alike.

Agent · Contract reviewConfiguration
role: review incoming supply contracts
reads: contract registry, counterparty directory, ERP
can: extract terms · check against checklist · raise a comment
cannot: sign · change amounts · write to external systems
asks a human when: amount > $500,000 · new counterparty · deviation from template
log: every step with source and timestamp
  • An agent's rights never exceed those of the user it acts for
  • Actions that change data go through approval
  • Every answer links to a document or record
  • When data is insufficient, the agent says so rather than inventing an answer
  • Quality is checked against a reference question set before every update
Agent lifecycle

Design, evaluate, approve, monitor

Agents are managed like any other production system: specified, tested, signed off and watched. Nothing reaches users without passing each gate, and every version can be rolled back.

Step 1 · with the process owner

Design

Role, data sources, permitted actions and escalation rules are written down with the person who owns the process. Security reviews the access list.

Gate passed whenThe owner signs off a one-page agent specification
Step 2 · on real cases

Evaluate

The agent runs against a reference set of real questions and documents. We measure correctness, citations and refusals, and test it with adversarial prompts.

Gate passed whenResults meet the thresholds agreed in advance
Step 3 · with security and the owner

Approve

Permissions are checked against your directory. The version is recorded in the change log and released to a pilot group before everyone else.

Gate passed whenNamed approvers sign the release
Step 4 · continuously

Monitor

Live metrics, sampled human review and alerts when answers drift. The reference set is re-run before every model or prompt change.

Gate passed whenMonthly quality report; one-step rollback

Guardrails on every agent

  • Least-privilege access inherited from your directory
  • Spending, volume and rate limits per agent
  • Human approval above thresholds you set
  • Personal data masked before any external model call
  • Answers without a source are flagged, not shown as fact

What an agent never does

  • Act outside its written role
  • Sign, pay or commit on the organization's behalf
  • See data its user isn't allowed to see
  • Change its own permissions or instructions
Integrations

Connects to the systems you already run

Read-only by default, through the cleanest path each system offers: an API, a database replica, an export or a document feed. Nothing has to be migrated or replaced.

ERP · finance

Enterprise resource planning

SAPOracleMicrosoft DynamicsExcel models
CRM

Customers and sales

SalesforceHubSpotMicrosoft Dynamics
Data

Databases and warehouses

PostgreSQLOracleMS SQLSnowflakeBigQueryKafka
Documents

Documents and archives

ECM / DMSFile sharesScans and PDFs (OCR)Email attachments
Communications

Email, chat and voice

IMAP emailMicrosoft TeamsSlackVoIP telephony
Identity

Single sign-on and roles

Active DirectoryLDAPSAMLOIDC
Security

Monitoring and audit

SIEM exportLogsMetricsTracing
Vision

Cameras and sensors

Existing CCTVIP cameras

Other systems, such as Workday, ServiceNow or in-house applications, connect via REST or GraphQL APIs, webhooks or exports. Anything non-standard is assessed during the diagnostic.

AI models

Frontier intelligence or sovereign models. Your choice.

The platform is model-agnostic. Choose frontier models through enterprise APIs, run open models trained on your own data inside your perimeter, or combine both. A model can be swapped without rebuilding the solution.

Frontier

OpenAI, Anthropic and other leading models

Through enterprise APIs whose terms exclude training on your data. The strongest reasoning for analysis, documents and agents.

  • Fastest to start
  • Best general reasoning
  • Data masked before it leaves
Sovereign

Local models trained on your data

Open models such as Llama, Qwen or Mistral, deployed on your servers or in an air-gapped network and fine-tuned on your documents, terminology and processes. Nothing leaves your perimeter.

  • Fine-tuning on your corpus
  • Our own security architecture
  • Works with no internet access
Hybrid

The right model for each task

Routing sends sensitive or routine work to local models and hard reasoning to frontier models, within the rules your security team sets.

  • Policy-based routing
  • Cost control per request
  • Full audit of every call

How routing decides

Every request is classified before it reaches a model. Your security team writes the policy; the router applies it and logs the decision.

  • Data class first: confidential and personal data stay on local models unless policy allows masking
  • Task second: hard reasoning goes to the strongest permitted model, high-volume routine work to a smaller one
  • Fallback: if a model endpoint is unavailable, the router uses the next model the policy allows
  • Every call is logged with model, version, cost and reason
Model router · decisionsExample · illustrative data
RequestData classRouteReason
Summarize board minutesConfidentialLocalPolicy: stays in perimeter
Draft a reply to a supplierInternalFrontier · maskedNames and amounts masked
Classify incoming emailInternalLocal · smallRoutine, high volume
Assess three-year contract riskInternalFrontierComplex reasoning
Citizen case filePersonalLocal onlyAir-gapped deployment
Deployment

Four ways to deploy: cloud to air-gapped

The choice depends on data sensitivity, data-residency rules and your IT policy. Platform functionality is the same in every option; what differs is where data lives and which language models are available.

BlackGust CloudPrivate cloudOn-premisesAir-gapped
Where the data livesOur data center, in-regionYour cloud accountYour serversIsolated network, no internet
AI modelsFrontier and localFrontier and localLocal; frontier by agreementLocal only
Isolation
Infrastructure run byBlackGustYour cloud team, with our engineersYour IT, with our engineersYour team on site, with our engineers
UpdatesScheduled maintenance windowsScheduled with your ITThrough your change processDelivered offline, through your change process
Time to deployDays1–2 weeks2–4 weeks4–8 weeks
Best forBusinesses that want a fast startCompanies with a cloud policyBanks, large holdingsGovernment, critical infrastructure
Available inFoundation, EnterpriseFoundation, EnterpriseEnterprise, SovereignSovereign

Timelines cover platform deployment and depend on the client's infrastructure readiness. Air-gapped deployments usually need GPU servers; we prepare the specification during the diagnostic.

Data residency: pick the option that matches your rules, whether that is the GDPR and the EU AI Act in Europe, CCPA/CPRA or sector rules in the US, LGPD in Brazil, PIPL in China, APPI in Japan, PIPA in Korea or the Privacy Act in Australia. We design each deployment to support compliance; see security.

Performance and reliability

Built to run in production, not in a demo

The platform is engineered for many departments and thousands of users on one model. How fast it answers depends on the models, the hardware and your data, so we size it for your load instead of quoting generic benchmarks.

Scale

Scales horizontally

Services run in containers on Kubernetes, or Docker Compose for small environments. Capacity grows by adding nodes; sizing is done during the diagnostic.

Resilience

Degrades gracefully

If a source system is down, answers show the time of the last sync instead of silently using stale data. If a model fails, routing falls back to another permitted model.

Recovery

Recovers to contract

Scheduled backups and recovery to the RPO and RTO agreed in the contract, documented in the runbook your team receives.

Observability

Visible end to end

Metrics, logs and tracing for every agent request: which sources it read, which model it called, how long each step took. Export to your SIEM.

Quality

Measured, not assumed

Reference question sets are re-run before every update. A drop in correctness blocks the release, the same way a failing test blocks code.

Support

Backed by an SLA

Business hours on Foundation, 24/7 for critical incidents on Enterprise, and a 1-hour response with a named duty engineer on Sovereign.

What we will not do: publish latency or throughput figures measured on someone else's data. During the pilot we measure them on your scenarios and write the targets into the SLA.

Platform by package

What comes with each package

The platform is licensed as an annual package. Each one includes forward-deployed engineers, not just software. Before the subscription comes a diagnostic and a pilot on your own data.

FoundationEnterpriseSovereign
Price per year$90,000$250,000$490,000
Scope
Business functionsOne3–5Organization-wide
ProcessesUp to 3Up to 15Unlimited
UsersUp to 150Up to 1,000Unlimited
Connected systemsUp to 5Up to 15As scoped
Modules
Executive✓✓✓
Operations · Documents · KnowledgeOne of these✓✓
Voice · Vision——✓
Deployment and models
DeploymentBlackGust Cloud or private cloudCloud, private cloud or on-premisesOn-premises or air-gapped
ModelsFrontier + hybrid routingFrontier + localLocal, fine-tuned on your corpus; frontier optional
Team and support
Forward-deployed engineers1 lead, on site at key phasesLead + 2, continuouslyEmbedded team of 5+ and an architect
SupportBusiness hours24/7 for critical incidents24/7, 1-hour response, named duty engineer
Procurement-ready documentation——✓

GPU hardware and infrastructure for Sovereign are priced separately. The full side-by-side is on the pricing page.

For IT teams

Technical overview for architects and CIOs

A short brief. Full technical documentation, an architecture description and a threat model are available under NDA.

Source accessRead-only by default. Writes happen only through explicitly permitted, approved actions.
IntegrationsREST/GraphQL APIs; direct connections to PostgreSQL, Oracle and MS SQL replicas; Snowflake and BigQuery; Kafka streams; SAP, Dynamics and Salesforce connectors; file exports; IMAP email; webhooks.
AuthenticationSingle sign-on via LDAP / Active Directory, SAML or OIDC. Roles are inherited from your directory.
EncryptionData is encrypted in transit and at rest. In on-premises deployments, the client holds the keys.
DeploymentContainers on Kubernetes, or Docker Compose for small environments. Infrastructure as code.
ObservabilityMetrics, logs and tracing of agent requests. Export to your SIEM.
BackupScheduled backups and recovery to the RPO/RTO agreed in the contract.
Platform APIYour systems can call agents and receive results through the API.
Security reviewWe complete your security questionnaire and walk your team through the architecture. See security.
FAQ

Platform questions, answered

What CIOs, CISOs and business owners ask most before a demo.

Do we have to replace our ERP, CRM or data warehouse?

No. The platform connects to the systems you already run, read-only by default, and builds the organization model on top of them. Your systems remain the source of truth.

Is our data used to train AI models?

No. Frontier models are used through enterprise APIs whose terms exclude training on your data, and personal data is masked before it leaves. Sovereign models are fine-tuned on your corpus inside your perimeter and are not shared with anyone.

Can the platform run fully on-premises or air-gapped?

Yes. On-premises takes 2–4 weeks to deploy and air-gapped 4–8 weeks, depending on your infrastructure. In an air-gapped network the platform runs on local models only, with no internet access. GPU hardware is priced separately.

Which language models do you use?

Frontier models from OpenAI, Anthropic and other leading providers, and open models such as Llama, Qwen and Mistral running locally. The platform is model-agnostic, so a model can be swapped without rebuilding the solution.

How do you stop agents from making things up?

Every answer must link to a record or document. When data is insufficient, the agent says so. Each agent is tested against a reference set of real questions before release and before every update, and actions that change data need human approval.

How long does it take to connect a system?

A system with a documented API or database replica is usually connected within the pilot. Undocumented legacy databases and paper archives take longer and are scoped during the diagnostic.

Which regulations does the platform support?

We design deployments to support compliance with the GDPR and the EU AI Act, CCPA/CPRA, LGPD, PIPL, APPI, PIPA, the Australian Privacy Act and sector rules such as HIPAA-style health data requirements. Data residency is solved by the deployment option. Details on the security page.

Can our own systems call the agents?

Yes. The platform API lets your applications send a task to an agent and receive the result with its sources, under the same permissions and audit log as any user.

What does the platform cost?

It is licensed as an annual package: Foundation at $90,000, Enterprise at $250,000 and Sovereign at $490,000 per year. A diagnostic ($1,490, credited to the pilot) and a pilot come first. See pricing.

What do you need from our IT team?

Read-only access to the agreed sources, a technical contact, network access for the deployment and help setting up single sign-on. Our forward-deployed engineers do the integration work.

See the platform on your own scenarios

A demo takes about an hour. Bring your IT and security leads; we'll show the architecture, an agent configuration and the audit log.