Skip to content
Pink beads of personal detail drift between dark vessels, and one sealed amber vial holds only a few.

Agentic security

AI agents and personal data: what an agent may see, keep and delete

The short answer

AI agents and personal data can work together safely when each agent sees only the records its task needs, keeps nothing longer than the work requires, and can find and delete what it holds when someone asks. Make those rules part of the system, and log what the agent did without copying the personal details into the log.

Written byAmelia GagneCo-founder and CEO, Kief Studio
Technical review byBrian GagneCo-founder and CTO, Kief Studio

Published

Contacts screen showing eight people as cards with email, phone, stage, next step and overdue markers, beside a menu for email and deal stages.
AI agents and personal data meet on screens like this one: names, emails, phone numbers and where each deal stands, the records an agent needs clear rules for.

Where AI agents and personal data meet

The owners I work with in regulated fields, such as health practices, financial firms, law offices and the companies that serve them, usually start with the same question about agents: where does our clients' information go once an agent starts reading it? How capable the AI is comes later.

That's the right question, because an agent moves personal data in ways a person doesn't. A person reads an email and remembers the gist. An agent copies the email into its working text, may summarize it into a memory, passes pieces to tools, and leaves a trail in logs. Each of those is a place the data now lives, and each one needs a rule.

The places personal data lands when an agent works

When you map an agent's work, personal information tends to collect in five places.

  • The context window: everything the agent is reading for the current task, sent to the AI model each time it thinks.
  • Agent memory: notes the agent saves so it can recall something next week, which outlive the task that created them.
  • Tool results: the records a tool hands back, such as a customer's file or the last ten invoices.
  • Drafts and outputs: the reply, the report or the summary the agent wrote.
  • Logs: the record of what the agent did, which can quietly become a copy of everything it saw.

What counts as personal information in Massachusetts

Massachusetts uses a specific definition. Under 201 CMR 17.02 (opens in a new tab), personal information is a resident's first name and last name, or first initial and last name, combined with a Social Security number, a driver's license or state ID number, or a financial account or card number. Information lawfully drawn from public records is excluded.

That definition is narrower than what most people would call private. Health details, home addresses and a client's family situation can fall outside it and still be things your clients expect you to protect, and other laws may cover them. I'd design for the broader set and treat the Massachusetts definition as the floor.

One stream splits into five branches, each dropping pink beads into a different vessel, and an amber ring sits at the split.

What an agent may see: data minimization

The first rule for AI agents and personal data is data minimization: an agent gets the records its task needs and nothing more. It's an old principle with a new urgency, because anything an agent can read, a hidden instruction in an email can try to get it to repeat. Brian Gagne, Kief Studio's CTO, covers that attack in prompt injection when the AI can act.

The guidance lines up. The Federal Trade Commission's guide for business (opens in a new tab) puts it in two words, scale down: keep only what you need for your business. OWASP's entry on sensitive information disclosure (opens in a new tab) (LLM02) says to grant access only to data that is necessary for the specific user or process. Massachusetts asks for the same thing in 201 CMR 17.04 (opens in a new tab): restrict access to records containing personal information to those who need it for their job.

In practice, that means building the agent's access around the task. An agent that drafts appointment reminders needs a name, a time and a phone number. It doesn't need the client's full file, and it shouldn't be able to open it.

A dark metal iris stays nearly shut, letting one narrow amber beam through while the wide pink haze of records waits behind it.

What an agent may keep: context and memory

Context is temporary. It exists for one task and then it's gone from the agent's side, though by then it has already been sent to the AI model. Memory is different. A saved note can sit for months, and it gets read again every time the agent recalls it.

That makes memory the place to be strict. OWASP's agentic risk list names ASI06 Memory and Context Poisoning, where false or planted information in what an agent remembers steers its later decisions, and the OWASP GenAI Security Project's exploit round-up (opens in a new tab) for the third quarter of 2026 maps real incidents to it. Personal data in memory carries a second risk: it becomes a shadow customer database that nobody chose to build and nobody remembers to clean.

Rules for AI agents and personal data in memory

These are the rules I'd set before an agent saves anything about a person.

  • Remember the fact the work needs, like a client prefers mornings, never the record it came from.
  • Point to the system of record instead of copying from it, so there's one place to correct or delete.
  • Keep notes an agent learned apart from rules a person wrote, so a learned note can never overrule a rule.
  • Give saved notes an expiry, and review what an agent remembers the way you'd review a shared drive.

Context

Context exists for one task and then it's gone from the agent's side, though it has already been sent to the AI model.

Memory

A saved note can sit for months, and it gets read again every time the agent recalls it.

A walnut card box of saved notes, each held by a brass date clip lit amber, and one thick card stuffed with papers edged in pink.

What an agent must be able to delete: retention and deletion requests

Data retention is how long you keep something. The FTC's guide calls the last step pitch it: properly dispose of what you no longer need. Massachusetts sets a standard for that too. M.G.L. c. 93I (opens in a new tab) requires that, when they're disposed of, electronic records containing personal information be destroyed or erased so the information cannot practicably be read or reconstructed.

Agents make deletion harder because they make copies. When a client asks you to delete their information, or when your own retention period runs out, a complete deletion request has to reach every place an agent may have put it.

  • The system of record: the CRM, the practice software or the file share.
  • The agent's memory and any notes it saved about that person.
  • Drafts, summaries and reports that quoted them.
  • Logs that captured their details instead of a reference.
  • The AI company's side, under the retention terms in your contract with them.
  1. System of record. Delete it from the CRM, the practice software or the file share.
  2. Agent memory. Remove it from the agent's memory and any notes it saved about that person.
  3. Drafts and reports. Remove the drafts, summaries and reports that quoted them.
  4. Logs. Clear the logs that captured their details instead of a reference.
  5. The AI company. Cover the AI company's side under the retention terms in your contract with them.
An amber sweep of light passes over five smoked translucent panes, and the pink traces inside each one dissolve into dust.

Logging without hoarding personal data

You need a log. Massachusetts asks for reasonable monitoring of systems for unauthorized use of or access to personal information in 201 CMR 17.04, and when something goes wrong, the log is how you answer what was touched. What you don't need is a log that duplicates every record the agent read.

IBM's Think piece on agent data leaks makes the point directly: treat prompt and response logs as sensitive data (opens in a new tab), because they duplicate whatever passed through the model. The same piece reports that, in IBM's 2026 research, security incidents involving unapproved AI use rose from 20% to 43% of organizations. Unapproved tools are often the ones logging everything somewhere you can't see.

The pattern that works is to log the action and a reference, not the content. Record which agent acted, what it did, which record it touched by ID, when, and who approved it. Leave the client's details in the system that already protects them.

Slim graphite tags hang on a rail, each lit by a tiny amber point, while the pink parcels they refer to stay locked away behind.

Organizations is highest in 2026 report at 43%.

Security incidents involving shadow AI, where workers use unapproved AI, more than doubled, from 20% to 43% of organizations between IBM's 2025 and 2026 reports. Source: IBM Cost of a Data Breach Report 2026
Organizations
2025 report20
2026 report43
Security incidents involving shadow AI, where workers use unapproved AI, more than doubled, from 20% to 43% of organizations between IBM's 2025 and 2026 reports. Source: IBM Cost of a Data Breach Report 2026 (opens in a new tab)

AI agents and personal data under Massachusetts law

This isn't legal advice, and your counsel should confirm how these rules apply to your business. What I can do is describe what the rules ask and how the engineering meets them.

201 CMR 17.00: the written program

Under 201 CMR 17.03 (opens in a new tab), anyone who owns or licenses personal information about a Massachusetts resident needs a written information security program. Several of its required parts map straight onto agents.

  • Designate someone to maintain the program. Name the person who owns each agent's access.
  • Prevent terminated employees from accessing records. Retire an agent's keys the day the agent is retired.
  • Oversee service providers and require safeguards by contract. Read how your AI provider stores and keeps what it's sent.
  • Review the program at least annually, or after a material change. Adding agents that read client records can be one, so treat it as a reason to review.
  • Document responses to incidents, with a post-incident review. Your agent log is the evidence for it.

M.G.L. c. 93H: when something goes wrong

Massachusetts' breach law, M.G.L. c. 93H (opens in a new tab), requires a business that owns or licenses a resident's personal information to give notice as soon as practicable and without unreasonable delay to the attorney general, the director of consumer affairs and business regulation, and the resident. Nobody can do that well without knowing what an agent touched, which is one more reason the log has to name records, agents and approvers.

A thick bound binder lit amber on a walnut shelf beside a locked steel records cabinet with faint pink seams.

What we've seen with owners in regulated work

Kief Studio runs its own company on LTFI, and we build systems for clients who answer to regulators and to their own clients. A few patterns about AI agents and personal data hold across all of them.

  • Most of the worry about AI comes from not knowing where data goes, and a simple map of it answers most questions.
  • Minimization is easier to design in than to add later. Deciding what an agent can't see is a one-time choice; cleaning up what it already copied is a project.
  • Deletion is the honest test. If you can't say where a client's information lives today, you can't promise to delete it.
  • The people closest to clients make the best rules. The front desk knows which details matter and which shouldn't travel.

How LTFI handles AI agents and personal data

LTFI is the agentic business operating system Kief Studio makes and runs. These are the controls that apply when its AI staff work with your clients' information, and they ship today.

  • Each company runs on its own dedicated server, so your data stays on your company's own server, managed, patched and watched by us.
  • Your AI runs on keys you own, under your own contract with the AI company, so the terms that govern what it keeps are yours to read and choose.
  • Rules, permissions, credentials, deleting and money stay with people. An AI teammate can't change them.
  • Anything that sends, posts, publishes or spends waits for the person who owns it.
  • The activity log records every action by the staff and the team in plain words, newest first, with who approved it.
  • Your contacts, files and content are yours, and if you leave, we export them for you.
A single server cabinet with an amber-lit brass lock stands apart, with a cluster of racks tangled in pink cables far behind.

What to do about AI agents and personal data this month

Two lists: one for the person who owns the business, one for whoever runs its systems.

If you own the business

  1. List the AI tools that can read client or employee information, including the free ones people signed up for.
  2. Decide which client details an agent should never see, and write them down.
  3. Name who handles a deletion request, and make sure they know every place an agent might keep a copy.
  4. Read the retention terms in your contract with each AI company, or ask your IT provider to.

If you run the systems

  1. Scope each agent's data access to its task, by field where the tool allows it.
  2. Store references in agent memory instead of copies, and give saved notes an expiry.
  3. Log actions with record IDs, agent, time and approver, and keep prompt and response logs out of reach or off by default.
  4. Write a deletion runbook that covers the system of record, agent memory, drafts, logs and the provider.

Key takeaways

  • AI agents and personal data mix safely when access is scoped to the task, memory is kept thin, and deletion reaches every copy.
  • Personal data collects in context, memory, tool results, drafts and logs, and each one needs a rule.
  • Log what an agent did and which record it touched, not the record itself.
  • In Massachusetts, 201 CMR 17.00 and M.G.L. c. 93H already apply whenever your agents can reach residents' personal information as those rules define it.

Where this lines up with the standards

The frameworks and rules this page maps to, with the exact item and what it asks of you.

FrameworkItemWhat it asks
Massachusetts201 CMR 17.03 and 201 CMR 17.04 (opens in a new tab)Restrict access to need, monitor for unauthorized use, and keep a written program.
MassachusettsM.G.L. c. 93H (opens in a new tab)Give notice of a breach to the attorney general, the director of consumer affairs and business regulation, and affected residents.
MassachusettsM.G.L. c. 93I (opens in a new tab)Destroy or erase electronic records so personal information can't practicably be read or reconstructed.
FTCProtecting Personal Information: A Guide for Business (opens in a new tab)Take stock, scale down, lock it, pitch it and plan ahead.
OWASP Top 10 for LLM ApplicationsLLM02 Sensitive Information Disclosure (opens in a new tab)Grant access only to necessary data, sanitize what reaches the model, and set retention and deletion policies.
OWASP Top 10 for Agentic ApplicationsASI06 Memory and Context Poisoning (opens in a new tab)Guard what an agent remembers against planted or false information.
NIST Privacy FrameworkPrivacy risk management (opens in a new tab)Find and manage privacy risk as part of running the business, so new tools protect the people whose data they use.

Key terms

Questions

AI agents and personal data

Can AI agents access personal data safely?

Yes, when each agent sees only the records its task needs, a person approves anything that sends or shares it, and every action is logged. Scope comes first: an agent that can't open a record can't leak it.

Do AI agents remember personal data?

Some do. Many agents can save notes to memory that outlast the task. Keep those notes to the fact the work needs, point to the system of record instead of copying it, and give saved notes an expiry.

How do I delete personal data from an AI agent?

Start from a map of where the agent puts data: the system of record, its memory, drafts, logs and the AI company's side. A deletion request has to reach each one, so the map is what makes deletion possible.

Should I put customer data into AI tools?

Only into tools you've approved, under a contract whose retention terms you've read, and only the fields the task needs. Free tools connected without approval are where most of the risk sits.

What should an AI agent log?

Which agent acted, what it did, which record it touched by ID, when, and who approved it. Treat prompt and response logs as sensitive, because they copy whatever passed through the model.

Does 201 CMR 17.00 apply to AI agents?

The rule covers any business that owns or licenses Massachusetts residents' personal information, whatever software handles it. If an agent can reach that information, the systems it uses fall under the rule's access limits, monitoring and written program, and giving each agent its own ID is how the engineering meets the unique ID control.

Is it legal to use AI agents with client data in Massachusetts?

The Massachusetts rules described here don't single out AI agents, and the same data security duties apply to the information they handle. Your counsel should confirm the details for your business; the engineering side is scoped access, a written program, monitoring and a clear record of what each agent touched.

Who is responsible for AI agents and personal data in a small business?

The business is, and in Massachusetts the written security program has to name someone to maintain it. Name one person who owns each agent's access and handles deletion requests, and make sure the log records who approved what.

What is the biggest risk with AI agents and personal data?

The biggest risk is copies nobody planned. An agent that saves client details to memory or logs full conversations builds a second database that's easy to forget and hard to delete. Keep references, not copies.

What should a policy on AI agents and personal data cover?

Which data each agent may see, what it may keep and for how long, who approves sending it outside the business, how it's deleted when someone asks, and what gets logged. Keep it short enough that your staff will actually read it.

Keep reading

  • A heavy threshold gate lets one cleared path through while dozens of pink trails stop at it.

    Start here

    Agentic AI security

    Agentic AI security explained for owners: identity, least privilege, approvals and an audit trail for AI agents, from the team that runs them in Massachusetts.

    Brian Gagne

  • A column of stacked plates with pink-lit seams sealed by an amber top.

    Oversight

    Audit trails for AI agents

    An AI agent audit log proves what an agent did, who asked and who approved it. What to record, how to keep it tamper evident, and how long to keep it.

    Amelia Gagne

  • A blurred hub drowns in pink cables while a clean hub in front holds one amber line.

    Identity and access

    AI agent permissions

    AI agent permissions decide what an agent can read, change and send. How to scope its tools, catch permission drift and run an access review.

    Brian Gagne

  • Hidden pink-lit devices behind dark panels, and one wide amber archway beside them.

    Threats

    Shadow AI

    Shadow AI is the AI your team uses without approval. How to find it without spying, write an AI policy people follow, and offer a better tool.

    Amelia Gagne

  • A smoked translucent slab carries a hidden pink thread toward the agent's tools, and an amber ring stops it before it reaches them.

    Threats

    Prompt injection

    Prompt injection turns an email, a web page or a file into orders for your AI agents. Why it can't be trained away, and the controls that limit the damage.

    Brian Gagne

  • Reference

    Glossary

    An agentic AI glossary in plain words: AI agents, agentic IAM, non-human identities, prompt injection, shadow AI, MCP and more, each with a short definition.

Most MSPs are resellers. Kief Studio is the manufacturer.

A reseller sends your problem to the company that made the software. We are that company: we build LTFI and run it, so the people you talk to can change it. Book a walkthrough and see it on your own business.

Nothing plays until you press play.