Where the agent runs
The agent harness may run on a workstation, server, container, Kubernetes environment, private virtual machine, or client-controlled cloud service. That location does not determine where the model runs.
Enterprise AI privacy services in Portland and across Maine
Enterprise AI agent data privacy and private deployment
AI Impact Maine helps organizations understand and control where AI-agent information is processed, transmitted, retained, logged, and stored. We design local, private-cloud, hybrid, and enterprise API architectures based on the organization’s data sensitivity, model requirements, infrastructure, and operating responsibilities.
Data-flow visibility
Client-controlled identity
Least-privilege tools
Human approval
Service Overview
Privacy depends on the complete agent system—not one installation choice.
This service is for Maine organizations evaluating private AI agents, secure enterprise AI, local AI agent deployment, isolated cloud AI agents, or direct commercial APIs while handling confidential business records, internal knowledge, operational data, or other sensitive information.
A locally installed AI agent does not automatically mean all information remains local.
Information may still leave client-controlled infrastructure when an agent uses cloud-hosted AI models, external APIs, websites, search services, messaging platforms, SaaS applications, remote databases, third-party tools, or telemetry and logging services.
Where inference happens
A local model runs on client-controlled hardware. A private-cloud or direct enterprise model is processed by the selected cloud or model provider under its current service configuration and terms.
Where memory and records live
Conversation state, files, vector stores, logs, backups, tool results, and agent memory can each have different storage, retention, deletion, access, and regional-processing behavior.
“No training” is not the same as “no processing” or “zero retention.”
It does not necessarily mean the provider never receives the request, no temporary processing occurs, no security or abuse-monitoring retention occurs, every feature qualifies for zero retention, or every external tool follows the same privacy policy.
Primary Service
Private cloud AI agent deployment
AI Impact Maine can design and deploy an agent inside a client-controlled Microsoft Azure subscription, Amazon Web Services account, Google Cloud environment when appropriate, private hosted infrastructure, isolated virtual network, or hybrid on-premises and private-cloud environment.
Client-controlled cloud account + isolated VPC or VNet + privately hosted agent runtime + approved enterprise model endpoint + client-controlled storage, memory, identity, and logging
Where the private agent runtime can operate
- A private Azure virtual network
- An isolated AWS VPC
- A private Kubernetes environment
- A dedicated virtual machine
- A containerized private service
- A client-controlled cloud server
How the model can be connected
- Hosted within the client-controlled cloud environment
- Accessed through Azure-hosted enterprise model services
- Accessed through Amazon Bedrock
- Accessed through another approved private model endpoint
- Hosted locally while the agent runtime operates in private cloud
- Routed between local and private-cloud models by data sensitivity
Private endpoints and DNS
VPC or VNet isolation
No public inbound access
Restricted outbound connections
Network allowlists
Firewall and security-group rules
Role-based access
Managed or workload identities
Customer-managed keys when supported
Encrypted storage
Private memory and vector stores
Secrets management
Region and residency selection
Retention configuration
Client-controlled logging
Backup and recovery
Human approval controls
Agent and sub-agent isolation
Emergency shutdown
Documented operating ownership
A private network is a control boundary, not a blanket privacy guarantee.
A private or isolated cloud deployment reduces public exposure and gives the organization greater control over networking, identity, storage, and logging. It does not mean the cloud provider performs no processing, and it does not automatically guarantee that information never leaves the selected environment. Model endpoints, external tools, telemetry, backups, support features, and cross-region processing must all be reviewed.
Architecture Options
Four practical ways to place the agent, model, memory, and controls.
The right design depends on data sensitivity, model quality, latency, hardware, integrations, operating skills, contractual terms, and which organization is prepared to maintain each layer.
Fully local
Client-controlled agent + model + memory and storage
Agent and model run on client-controlled hardware. Prompts can remain inside the organization when external tools are disabled. This can fit highly sensitive workflows but requires appropriate hardware, model-license review, hosting, updates, backups, monitoring, and security controls.
- Hardware and capacity review
- Local model selection and licensing review
- Agent harness installation
- Storage, memory, logging, and backup design
- Tool restrictions, network isolation, and human approval
Private or isolated cloud
Client-controlled agent + private cloud account + approved model endpoint
Use a client-controlled Azure, AWS, Google Cloud, or other private environment with private networking, identity controls, encrypted storage, retention review, and centralized logging. The model may be hosted privately or provided through an approved enterprise model service.
- Azure VNet or AWS VPC architecture review
- Private networking, identity, and model selection
- Retention-setting and data-flow review
- Agent integration, monitoring, and documentation
- Backup, rollback, shutdown, and staff handoff
Direct enterprise API
Client-controlled agent + direct commercial model API
OpenAI or Anthropic receives the request to perform inference. Their current commercial documentation says business/API content is not used for model training by default, but default security or abuse-monitoring retention and feature-specific application state may still apply.
Zero Data Retention can require approval or a separate agreement. Eligibility depends on the organization, endpoint, model, feature, and contract. Stored conversations, files, assistants, background jobs, batches, and other stateful features can behave differently.
Hybrid and data-minimized
Local agent and storage + local minimization + approved cloud inference
Sensitive source records can remain local while only the minimum necessary, redacted, or tokenized content is submitted for approved inference. Local retrieval can limit prompt contents, workflows can use different models, and higher-risk actions can require human approval.
- Data classification, redaction, and prompt minimization
- Local retrieval and model routing
- Permission and external-tool boundaries
- Human approval, logging, and incident review
Architecture Comparison
Compare control, processing, retention, and operating responsibility.
| Architecture | Agent and model | Hardware and processing | Networking and exposure | Storage and retention | Operations and best fit |
|---|---|---|---|---|---|
| Fully local | Both run on client-controlled hardware. | Information can stay on company-controlled hardware if external services are disabled; the organization processes inference locally. | Can be isolated or allowlisted; public internet access is optional and should be minimized. | Organization controls memory, files, vector stores, logs, deletion, and backups. | Client or managed operator maintains hardware, model, patches, backup, and monitoring. Fits high-sensitivity bounded workflows with adequate capacity. |
| Private or isolated cloud | Agent runs in a client-controlled VPC/VNet; model is privately hosted or accessed through an approved enterprise endpoint. | Information leaves company-owned hardware for the client-controlled cloud. The cloud and model service process the request as configured. | Private endpoints, private DNS, no public inbound access, egress controls, allowlists, and identity policies are available. | Client controls agent memory and storage; model-service retention and feature behavior require verification. | Shared responsibility across client, implementation partner, and cloud provider. Fits enterprise cloud AI agents needing private networking and scalable operations. |
| Direct enterprise API | Agent runs under client control; provider-hosted model receives inference requests. | Content leaves company hardware and is processed by the contracted model provider. | Outbound encrypted API access; provider-side public service boundary and optional enterprise controls. | Default abuse-monitoring and application-state behavior varies; approved zero-retention controls are endpoint- and feature-specific. | Client operates the agent and provider operates the model. Fits teams prioritizing model capability and simpler infrastructure after contract review. |
| Hybrid local and private cloud | Agent, retrieval, and storage can remain local while approved content is routed to local or private-cloud models. | Only selected, minimized content leaves company hardware; local and cloud processors depend on routing. | Policy-based routing, private endpoints, outbound restrictions, and human approval for sensitive actions. | Sensitive source data stays local; cloud request retention still requires model- and feature-level review. | Client manages classification and local systems; cloud provider manages hosted services. Fits mixed-sensitivity workflows and phased modernization. |
On mobile, each architecture is displayed as a labeled card so every comparison field remains readable without horizontal scrolling.
Current Provider Documentation
Provider names do not remove the need to verify the exact deployment.
These are planning summaries, not contractual promises. During every engagement, AI Impact Maine verifies the current official documentation and the client’s applicable terms for the chosen account, model, deployment type, region, API, feature, and tool.
Microsoft Azure and Foundry
Microsoft states that prompts and completions submitted to models sold through Azure are not available to OpenAI or other underlying model providers, and are not used to train foundation models without permission. Microsoft hosts and processes those models in Azure, and some stateful or safety features can store data. Global and DataZone deployments also affect where processing occurs.
Amazon Bedrock
AWS documents retention modes that can keep supported requests from durable storage or provider sharing, retain data within AWS, or permit provider sharing for models that require it. The selected model’s current allowed modes, region, cross-region inference, feature behavior, and account configuration must be checked before deployment.
Direct OpenAI or Anthropic API
OpenAI says API data is not used for training unless the customer opts in, while default abuse-monitoring logs can be retained and approved controls vary by endpoint and feature. Anthropic says standard API inputs and outputs are generally deleted within 30 days, subject to documented exceptions; separate approved zero-retention arrangements apply only to eligible products and organizations.
Zero retention does not mean zero processing. “No training” does not mean no receipt, no temporary processing, no security review, or identical handling by external tools. Provider terms and product behavior can change; current official documentation and contractual terms must be reviewed for every implementation.
Data-Flow Assessment
Know where agent data actually goes
The model is only one part of the complete agent data path. A useful assessment follows information from its source through instructions, retrieval, inference, tools, storage, review, and recovery systems.
- User input: what people type, speak, paste, or submit.
- System instructions: prompts, policies, hidden context, and orchestration rules.
- Business records: retrieved documents, databases, and internal knowledge.
- Uploaded files: storage, scanning, extraction, expiration, and deletion.
- Agent memory: short-term state, profiles, long-term notes, and summaries.
- Vector database: embeddings, source chunks, metadata, access, and backups.
- Model provider: inference processing, safety controls, state, and retention.
- Tools and APIs: payloads, responses, permissions, and third-party terms.
- Browser and web: websites, downloads, cookies, forms, and remote content.
- Logs and telemetry: prompts, tool calls, errors, traces, alerts, and vendors.
- Backups: copies, regions, encryption, access, lifecycle, and recovery.
- Human review: approval screens, audit consoles, exports, and notifications.
- Sub-agents: delegated context, shared memory, credentials, and handoffs.
- Messaging platforms: email, chat, SMS, gateways, archives, and recipients.
- Business applications: CRM, finance, productivity, and operational systems.
Privacy and Security Controls
Turn the data-flow diagram into an operating boundary.
Control design connects the approved business purpose to the minimum data, identity, model, tools, memory, networking, logging, and human authority needed for the workflow.
Data inventory
Data classification
Approved use cases
Approved model list
Approved agent list
Provider and contract review
Retention-setting review
Private networking
Role-based identity
Least-privilege tools
File and folder boundaries
Credential handling
Encryption in transit and at rest
Agent-memory boundaries
Data-loss prevention
Redaction and tokenization
Human approval
Activity logging and auditing
Sub-agent governance
Emergency shutdown
Backup and recovery
Incident-response procedures
Periodic architecture review
Change and version ownership
AIM WorkAgent and AIM ConnectAgent
AIM WorkAgent can be designed around client-approved business records, model routes, tools, and human review. AIM ConnectAgent adds approved web, messaging, gateway, or internal interfaces, so each channel and destination becomes part of the data-flow assessment.
AIM AgentGuard
AIM AgentGuard supports agent inventory, permissions and tool review, monitor-only logging review, activity auditing, governance, and remediation guidance where integrations and records are available. It does not replace endpoint security, DLP, identity security, network security, backups, legal review, or incident response.
Service Offers
Start with a decision, then build one bounded architecture.
Scope is confirmed before work begins. No prices are published because effort depends on data sensitivity, systems, providers, tools, operating model, documentation needs, and the selected deployment boundary.
Enterprise AI Agent Data Privacy Assessment
- Business use-case review
- Data inventory and classification workshop
- Agent data-flow diagram
- Model and provider comparison
- Local-versus-cloud recommendation
- Retention and logging review
- Tool and connector inventory
- Identity, permission, and human-approval review
- Risk register and recommended architecture
- Implementation roadmap, written findings, and executive summary
Private Cloud AI Agent Deployment
- Cloud-account and architecture review
- Azure VNet or AWS VPC design
- Agent-runtime deployment
- Private model-endpoint connection
- Identity and permission configuration
- Private storage and memory configuration
- Tool, connector, and network-egress restrictions
- Logging, monitoring, and retention review
- Human approval controls
- Backup, rollback, security documentation, and staff handoff
Private AI Agent Architecture and Pilot
- One bounded workflow
- Selected agent harness
- Selected model architecture
- Client-controlled memory design
- Limited approved tools
- Human approval points
- Logging and test plan
- Data-flow validation
- Rollback plan
- Staff handoff and documentation
How the Engagement Works
Map first, choose architecture second, connect tools last.
Bound the use case
Identify the owner, users, business outcome, sensitive information, decisions, approved actions, and prohibited actions.
Trace the data path
Map the runtime, model, memory, storage, tools, identities, networks, logs, backups, sub-agents, and human interfaces.
Compare options
Evaluate local, private-cloud, direct API, and hybrid designs against current documentation, contracts, capabilities, and operating capacity.
Implement and verify
Configure one bounded architecture, validate flows and approvals, test failure and rollback, document ownership, and review residual risk.
Limitations and Accuracy
Technical privacy planning is not a compliance guarantee.
What AI Impact Maine provides
AI Impact Maine provides technical assessment, architecture, implementation, documentation, staff handoff, and operational support. Engagement scope depends on supported systems, available documentation, access, logs, integrations, and the client’s operating responsibilities.
AI Impact Maine does not provide legal advice or compliance certification. Qualified legal, privacy, security, procurement, and compliance professionals should review applicable requirements and contracts.
What no architecture guarantees
- No guarantee of perfect confidentiality, security, compliance, or breach prevention
- Local deployment does not automatically establish compliance or security
- Private cloud does not mean data remains on company-owned hardware
- Zero retention does not mean zero processing
- External tools and connectors have their own data-handling terms
- Provider terms, models, endpoints, features, and behavior can change
Privacy requirements depend on the client’s data, industry, contracts, jurisdictions, providers, models, endpoints, and enabled features. Current official documentation and contractual terms must be verified during every engagement.
Common Questions
Enterprise AI agent data privacy FAQ
Can an enterprise AI agent keep all data inside the organization?
It can be designed to keep approved prompts, models, memory, and storage on client-controlled infrastructure when external tools and cloud services are disabled. The complete workflow must still be reviewed because browsers, updates, telemetry, backups, messaging, and other connectors can create external data paths.
Does a locally installed agent guarantee that data stays local?
No. A local agent runtime can still send information to a cloud model, external API, website, search service, SaaS application, remote database, messaging platform, or logging service. Local runtime, local model, and local storage are separate architecture decisions.
Can an organization use OpenAI models without sending data directly to OpenAI?
Microsoft states that models sold through Azure are hosted in Microsoft's Azure environment and prompts and completions are not made available to OpenAI or other underlying model providers. Microsoft still processes the information to operate and protect the Azure service, and deployment type, enabled features, storage, abuse monitoring, and regional processing must be reviewed.
Can an organization use Claude without sending prompts directly to Anthropic?
Amazon Bedrock can provide access to supported Claude models through AWS. AWS documentation says provider access and retention depend on the selected model and configured retention mode: some modes keep requests within AWS, while certain models or features can require provider data sharing. The current model's allowed modes must be verified before deployment.
What is the difference between no training and zero data retention?
No training generally addresses whether customer content is used to improve models. Zero data retention addresses whether eligible request and response content is stored after processing. Neither phrase by itself describes every log, safety control, stateful feature, external tool, backup, or contractual exception.
Does zero data retention mean the provider never processes the prompt?
No. A hosted model provider must process the submitted content to perform inference. Zero-retention controls address storage after processing, subject to the provider's current documentation, approved controls, supported endpoints, features, contract, legal requirements, and safety exceptions.
Can different business workflows use different privacy architectures?
Yes. A hybrid design can keep higher-sensitivity records and retrieval local while routing minimized, redacted, or lower-risk content to an approved private-cloud or enterprise API endpoint. Each route needs a documented data classification, permission boundary, logging plan, and human approval model.
What does an Enterprise AI Agent Data Privacy Assessment include?
It includes a business use-case review, data inventory and classification workshop, agent data-flow diagram, model and provider comparison, retention and logging review, tool and connector inventory, identity and permission review, human-approval recommendations, risk register, recommended architecture, implementation roadmap, written findings, and executive summary.
Does AI Impact Maine provide compliance certification?
No. AI Impact Maine provides technical assessment, architecture, implementation, documentation, and operational support, not legal advice or compliance certification. The client and its qualified legal, privacy, compliance, and security advisers remain responsible for applicable requirements.
How does AIM AgentGuard support data privacy?
AIM AgentGuard can support approved environments with agent inventory, permission and tool review, logging-gap analysis, monitor-only activity review, selected session auditing, governance documentation, and remediation guidance. Coverage depends on supported integrations and available records, and it does not replace endpoint, identity, DLP, network, or incident-response controls.
What is Private Cloud AI Agent Deployment?
Private Cloud AI Agent Deployment places the agent runtime, identity, storage, memory, and logging inside a client-controlled cloud account and isolated network, then connects only approved model endpoints and tools. It improves control over networking and operations but does not mean the cloud provider performs no processing or that every enabled service keeps data inside one environment.
Start with the information, workflow, and decision that need protection.
Tell us what the agent needs to access, which models or cloud environments you are considering, and who must approve sensitive actions. Do not send confidential records through the contact form; we will establish scope before reviewing sensitive material.