00 / SECURITYTHE FULL GUIDE

Security, explained.

AI is only as safe as where your data goes. Here are the three ways to run a private assistant, what each one means for your data, and what we set up whichever you choose. We never host your data in any of them.

Vendor terms change. Checked October 2026.

01 / WHY IT MATTERSTHREE REAL RISKS

Most AI risk is about where data goes.

The model is rarely the weak point. The weak points are people, permissions, and contracts. Three show up again and again.

Risk 01

Shadow AI

There is no approved tool, so staff paste client data into personal chat accounts. Those accounts have their own terms. You cannot see them, audit them, or turn them off.

Risk 02

Oversharing

An assistant surfaces what people can already access. If a folder is open to everyone by mistake, the assistant makes that visible overnight. Fix permissions before you turn AI on.

Risk 03

Unclear vendor terms

Who keeps your prompts? For how long? Can they train on them? The answers differ by vendor and by plan. They are written down. Someone has to read them.

02 / THREE WAYSWHERE THE DATA LIVES

Three ways to run it.

The assistant, its search index, and the model can sit in different places. Pick a tab to see where the boundary falls. In every option the consultant stays outside it, and your data is never stored with us.

Where the data lives: vendor's cloud, with the consultant outside the boundary. Vendor's cloud Your cloud account Your building Your building Data lives here Assistantanswers, cites Indexsearch index Model vendor's model e.g. Bedrock local open-weight cloud, no-training passages only setup access · logged · revocable Consultant always outside

Everything inside the box sits with the vendor, under its business terms.

Side by side
TopicA. Subscription (vendor-hosted)B. Your cloudC. On-site
What it is A business plan of an existing AI product. Examples: ChatGPT Business or Enterprise, Claude Team or Enterprise, Microsoft 365 Copilot, Gemini for Workspace, or Onyx Cloud. An open-source assistant (for example Onyx) deployed in your own AWS, Azure, or Google Cloud account. Models run through a service such as AWS Bedrock. A server or NAS in your office. Open-weight models run locally, for example with Ollama or vLLM. It can run fully offline.
Where data lives With the vendor, under its business terms. In your cloud account. In your building.
Setup effort Lowest.Medium.Highest.
Ongoing cost Per-user subscription. Your cloud bill plus model usage. Hardware and power. Little or no usage fees.
Control Least.High.Highest.
Healthcare (BAA) Only on some plans. Check each vendor. You sign a BAA directly with your cloud provider. No third party involved for the model.
Best for A fast start, general productivity, small teams. Regulated data, custom connectors, room to grow. The strictest confidentiality, unreliable internet, sensitive archives.
What we do Pick the right plan. Configure admin and retention settings, SSO and MFA, connectors, and permissions. Write the AI-use policy. Train the team. Design, deploy, connect, clean data, harden, hand over. Spec the hardware, install, connect, clean data, harden, hand over.
Also possible

Hybrid

Keep the index and the assistant in your building. Send only the few relevant passages to a cloud model that runs under no-training terms. Whole files and the full archive never leave.

Option A and B, vendor by vendor

What the vendors say today

Names are examples only. We are not affiliated with any of them, and we are not recommending one here. Vendor terms change. Checked October 2026. Always confirm against the vendor's current terms before you sign.

Vendor data terms and healthcare agreements, as checked October 2026
VendorTraining on your contentHealthcare (BAA)Price (list)
Claude Team (Anthropic) No model training on your content by default.1 The Team plan is not included for HIPAA. Enterprise needs the account owner to accept a BAA when HIPAA is activated, and some features are not covered.12 See vendor's current pricing.1
ChatGPT Business and Enterprise (OpenAI) Not used for training by default on business plans.3 Only on specific healthcare and enterprise offerings. Confirm with OpenAI.3 See vendor's current pricing.
Microsoft 365 Copilot Prompts, responses, and data accessed through Microsoft Graph are not used to train foundation models.4 A BAA is offered through Microsoft's online services terms. Confirm that your licence is in scope.5 See vendor's current pricing.
Gemini in Google Workspace Customer data is not used to train models without your prior permission.6 Can support HIPAA workloads on a qualifying edition. Some features are outside Google's BAA. Confirm your edition.6 See vendor's current pricing.7
Onyx Cloud Onyx states it does not train on your data. Confirm retention and data location with Onyx.8 Not found in the public pages. Confirm with Onyx. See vendor's current pricing.8
AWS Bedrock (Option B) Your content is not used to improve base models and is not shared with model providers.9 HIPAA eligible. You accept the BAA with AWS for your own account.9 Usage-based. You pay AWS directly.9

Option C uses open-weight models run locally with tools such as Ollama or vLLM. There is no vendor to sign with for the model.

Sources
  1. claude.com/pricing
  2. privacy.claude.com, BAA for commercial customers
  3. openai.com/business-data
  4. learn.microsoft.com, Microsoft 365 Copilot privacy
  5. learn.microsoft.com, HIPAA and HITECH offering
  6. Google Workspace, generative AI privacy hub
  7. workspace.google.com/pricing
  8. onyx.app/pricing
  9. aws.amazon.com/bedrock/faqs
03 / WHICH FITS YOU?FOUR QUESTIONS

Which one fits you?

Answer four questions. You get a starting point, not a verdict.

Question 01Do you handle regulated data (health, financial, legal privilege)?
Question 02Do you have IT staff or a managed IT provider?
Question 03How fast do you need it?
Question 04Do you need custom systems connected (DMS, practice management, EHR)?

A starting point — the Discovery & Roadmap confirms it.

Your starting point

Answer all four questions to see a recommendation.

04 / OUR BASELINEWHICHEVER YOU CHOOSE

What we set up, whichever you choose.

The same baseline goes into every build. The details change with the option.

  • 01Permission inheritanceAnswers respect the permissions in the source system. We review those permissions first.
  • 02Least-privilege service accountsEach connector gets read access to only what it needs.
  • 03SSO and MFAYour people sign in with the accounts they already use, with a second factor.
  • 04EncryptionIn transit and at rest, with keys in your control where the platform allows it.
  • 05Audit logsA record of who asked what, kept in your environment.
  • 06Retention settingsYou decide how long chats and logs are kept. We set it and write it down.
  • 07No-training model termsOnly models whose terms or settings rule out training on your content.
  • 08Agent guardrailsDefenses against prompt injection and data exfiltration. Agents cannot send email or delete files without a person's approval.
  • 09Human review of outputsAnswers cite their sources. Important work gets checked by a person.
  • 10BackupsThe index and configuration can be restored.
  • 11Incident response planA short written plan: who does what, and who is told.
  • 12OffboardingWhen someone leaves, their access to the assistant ends the same day.
05 / ACCESSHOW WE ACCESS YOUR SYSTEMS

We work from outside the boundary.

Setup needs access. We keep it small, visible, and temporary.

  • 01
    Accounts you issueYou create the accounts. We never use shared or personal logins.
  • 02
    Least privilege, with MFAOnly the access the current task needs, behind a second factor.
  • 03
    Logged and time-limitedYour logs show what we did. Access expires on a date.
  • 04
    Revoked at handoverWhen the work ends, you switch our accounts off.
  • 05
    Temporary copies only by planWe do not store your data on our systems. A temporary working copy is made only if your signed project plan allows it, and it is deleted after the task ends, on the schedule in your plan.
  • 06
    Paper where it is neededAn NDA, a data processing agreement, and a BAA where the law requires one.
In plain terms

You own the accounts, the servers, and the contracts with your cloud and model providers. We design and build. We do not hold your infrastructure, and we do not store your data on our systems. When we use AI coding tools to build connectors, they never see your live data.

Read the privacy notice →

06 / OBLIGATIONSBUILT TO SUPPORT

Built to support your obligations.

We build systems designed to support your duties. We do not certify anything, and we do not give legal advice. Your own counsel and compliance staff decide what your rules require.

Healthcare

HIPAA

Where protected health information is involved, a business associate agreement is needed with each party that can access it. In Option B, you sign a BAA with your cloud provider. In Option C, no third party handles the model. We sign a BAA with you where the law requires one.

Financial

FTC Safeguards Rule

The rule expects you to oversee your service providers. We give you a written summary of how the system is secured, so you have it on file.

Legal

ABA Formal Opinion 512

The opinion covers vendor diligence, client consent, and supervision when lawyers use AI. We supply the technical facts. The firm makes those decisions.

Securities

SEC and FINRA records

Where recordkeeping rules apply, archive search can be set to respect your retention rules. Your compliance team defines what those are.

07 / NEXT STEPDISCOVERY & ROADMAP

Not sure which setup is right?

That is what the Discovery & Roadmap is for. We look at your systems and your data, then write down which option fits and what it will take. The first call is free.

Book a free intro call Privacy notice

Vendor names on this page are examples. We are not affiliated with them. Vendor terms change; checked October 2026.