AI Data & Privacy Policy

AI systems are only useful when they are connected to real data. That means the companies we work with hand us access to the systems that run their business. We think that trust should be met with a written, public standard rather than a reassurance in a meeting.

This policy applies to every LightCI engagement in which we access, process or build systems that process your data (“client data”). It supplements your master services agreement and any data processing addendum. If a signed agreement says something more specific or more protective, the agreement governs. Our general Privacy Policy covers how we handle personal information on our own website and in our own business.

Every system we build follows the same shape. Data is read, with scoped permissions, from your systems of record (for example your CRM, ERP and finance systems, support platform, document stores, data warehouse and product databases) into a semantic layer and agents that run in your environment: your cloud account, your keys, your logs. Model calls go out to the provider under enterprise terms with no training and zero retention, and outputs are returned to your systems. Inputs stay in your environment, outputs return to your systems, and logs are written to your observability stack.

LightCI engineers access these systems through named accounts issued by your identity provider, with least-privilege, time-boxed access that is revoked at close. No production data is held on our machines. We document this data map for each build before it goes live, and you keep it.

The one thing that does not appear on that map is a LightCI data store. We do not maintain copies of your production data, and we design so that we never need to.

Depending on the engagement, the systems we connect to may contain:

• Business records. Customer accounts, deals, invoices, contracts, support tickets, product usage, internal documents, financial models and reporting.
• Personal information of your people and your customers as it exists inside those records: names, contact details, roles, communications, transaction history.
• Credentials and access that you issue to us: accounts, API keys, tokens and roles.
• Telemetry from the systems we build: prompts, tool calls, outputs, latencies, errors and evaluation results, written to logging you control.
• Working sessions. Meetings with your team may be recorded or summarised with an AI notetaker where participants have been told. Recordings and notes are kept in our workspace, limited to the engagement team, and deleted on request.

We ask for the minimum needed for the job. Much of a build can be done on synthetic, sampled or masked data, and we default to that until a system is ready for production data.

For client data, you are the controller (or “business” under U.S. state law) and LightCI is a processor (or “service provider”). We process client data only on your documented instructions, for the purposes of the engagement, and not for our own purposes.

You are responsible for having the right to share the data with us, for notices to your own customers and employees, and for telling us about any special handling requirements. We are responsible for following this policy, your instructions and the security measures below, and for helping you respond to data-subject requests, audits and regulators for the systems we built.

Business contact information about your staff, such as the names and emails of the people we work with, is handled under our general Privacy Policy, where we act as controller.

• We use provider tiers whose terms exclude training on inputs and outputs. Consumer products and free tiers are not permitted for client data.
• If you commission a fine-tuned or custom model, it is trained in your environment or under your provider account, and the resulting weights, data and evaluations are yours.
• Evaluation datasets, golden answers and test cases built from your data stay in your environment and are not reused for other clients.
• What we carry between engagements is know-how: architectures, patterns and techniques. Never your records, prompts containing your data, or outputs derived from it.

We are model-agnostic in design and Claude-first in practice. Our default provider is Anthropic, accessed either through your own enterprise agreement, through the Claude Developer Platform on enterprise terms, or through your cloud provider’s hosted offering of Claude (such as Amazon Bedrock or Google Cloud Vertex AI) inside your account.

We give you the sub-processor list for your build before production, tell you before adding one, and honour your right to object. You may restrict us to particular providers, regions or deployment models, including cloud-only or on-premises.

Systems we build are deployed into infrastructure you own: your cloud accounts, your tenants, your identity provider, your secrets manager. API keys for model providers are issued under your accounts wherever possible, so usage, billing and logs are visible to you and survive our departure.

• Data at rest is encrypted with keys you control. Data in transit is encrypted end to end.
• Logs, traces and evaluation results are written to your observability stack, not ours.
• Deployment regions follow your data-residency requirements.
• Where a component must be hosted by LightCI during a build, it is isolated per client, encrypted, access-controlled, listed in the data map and decommissioned at close.

• Named accounts issued through your identity provider, never shared credentials. Multi-factor authentication is mandatory.
• Least privilege. Read-only where read-only is enough. Scoped to the systems in the statement of work.
• Time-boxed. Access is tied to the engagement and reviewed at each phase gate.
• Managed devices. Full-disk encryption, screen lock, current patches, endpoint protection.
• No local copies. Production data is not downloaded to laptops or personal drives. Work happens in your environment or on masked samples.
• AI tooling under enterprise settings. Our engineers use AI coding assistants on enterprise, no-training tiers. When working in your codebase, those tools are configured to your policy or disabled.
• Revoked at close. We hand you a checklist of every account, key and role we held, and confirm removal in writing.

Agents act. That makes their data handling a design question, not just an access question. Every agent we deploy follows these defaults, which you can tighten:

• Scoped tools. An agent can only call the systems and actions it was built for, with permissions no broader than its purpose.
• Approval gates. Consequential or irreversible actions such as sending, paying, deleting or changing customer records require a human approval step until you decide otherwise.
• Full audit trail. Every prompt, tool call, decision and output is logged in your environment and attributable to a run.
• Data minimisation in context. Agents retrieve what a task needs, not whole tables. Sensitive fields are masked or excluded from model context unless the task requires them.
• Injection and exfiltration defences. Untrusted content is treated as data, not instructions. Outbound destinations are allow-listed.
• Kill switch. You can pause or disable any agent we build without our involvement.

Some categories of data get extra rules, agreed before we touch them.

• Health information (PHI). Handled only under a Business Associate Agreement and with your compliance team involved. De-identified or synthetic data is used for building wherever possible. PHI never enters a consumer AI tool.
• Payment card data. We design so that we do not need it. Card numbers are tokenised or excluded from any system we build.
• Financial and deal information. We work with investors and their portfolio companies and treat material non-public information accordingly: engagement-team-only access, no trading on information learned in an engagement, and information barriers between clients.
• Special categories of personal data under the GDPR and equivalent laws are processed only on explicit written instructions and with a documented lawful basis on your side.
• Children’s data. We do not build systems that process children’s personal data.

We work with many companies, sometimes in the same sector and sometimes backed by the same investors. Your data, prompts, evaluations and outputs are never shared with, or reused for, another client.

• Separate workspaces, credentials and repositories per client.
• Engagement teams see only the clients they work on.
• Case studies and references are published only with your written approval of the exact wording.
• Our people are bound by confidentiality obligations that survive their time at LightCI.

During an engagement, client data lives in your environment under your retention policies. Working material we hold, such as notes, masked samples and meeting recordings, is limited to what the engagement needs.

At the end of an engagement, or earlier on request:

1. All access you granted is revoked and confirmed in writing.
2. Any working copies of client data held by LightCI are returned or securely deleted within thirty days.
3. Any LightCI-hosted component is decommissioned and its storage wiped.
4. We provide a deletion certificate on request.

Exceptions are limited to what a law or a signed agreement requires us to keep, and to systems you have asked us to continue operating under a support contract.

If we become aware of a confirmed security incident affecting your data, we notify your designated contact without undue delay and no later than seventy-two hours after confirmation, or sooner if your agreement says so. The notice includes what we know, what we are doing, and what we need from you. We cooperate fully with your investigation and any regulatory notification you must make.

Security researchers and anyone who spots a problem in a system we built can write to us at privacy@lightci.com.

At any point in an engagement you can:

• Request the current data map and sub-processor list for your build.
• Restrict model providers, regions, deployment models or data categories.
• Review the access our team holds and the logs of what it did.
• Require additional masking, approval gates or human review in any workflow.
• Ask for deletion of any working material we hold.
• Audit our compliance with this policy, on reasonable notice, yourself or through a third party.

We may update this policy from time to time. We notify active clients of material changes before they take effect, and a change never reduces the protections in a signed agreement. Questions about how your data is handled, or a request under this policy, can go to the team below.