
Three questions about every AI agent
If your goal is audit and accountability, you need to answer three questions about every AI agent: who does it work for, what did it do, and can you prove the record was never changed? An AI agent is software that works for an employee: it reads, decides and acts using an AI model.
- Agents often run on shared keys, so a request traces back to a key, not to a person.
- What an agent sent and what came back sits in scattered application logs, if it was kept at all.
- When someone leaves, the agents they set up can keep running with nobody answering for them.
- And at the end, the auditor still has to take your word that nothing was edited.
Regulators already see the gap: FINRA's 2026 oversight report, written for broker-dealers, adds a new section on generative AI that covers AI agents and lists outcomes that are hard to trace among their risks.¹ IBM found that 92% of organizations with an AI-related breach lacked proper AI access controls.²
Accountability means a named person behind every agent, and a record nobody has to take on trust.
How Argent gets you there
Argent Strategic Intelligence sits between your AI agents and the AI models and tools they use, and decides each request before it goes out. Every decision it makes leaves a signed record, and every agent answers to a person.
| What you need | What Argent does | Capability |
|---|---|---|
| A person behind every agent | Every agent is registered to a named, active human owner. An owner who is not an active person is refused. | Identity tethering |
| No orphaned agents | When an owner is switched off in the console or in your directory, their agents are refused on the very next request, and an admin hands them to a new owner. | Owner offboarding |
| A record of every decision | Each decision on traffic routed through Argent is a signed record linked to the one before it. A record that cannot be signed is not written. | One-click audit evidence |
| Who changed what | Admin changes made in the Argent console go on the same signed record, shown beside the requests on the Evidence page. | One-click audit evidence |
| The whole conversation | What each agent sent and what came back is kept encrypted and replayed request by request. Every read of the words goes on the signed record under the reader's name. | Session replay |
| A record before anything is blocked | Watch mode records what each control would have refused or changed while requests pass unchanged. | Watch before you enforce |
| Agents nobody registered | Proxy-log imports and managed machines show AI in use that nobody registered, tagged Discovered, so it can be given an owner. | Shadow AI discovery |
| An answer for the board | Reports compares this period's checks with the last one, including ownerless agents by department. | Give the board what it wants |
The path: from first request to examiner-ready
You do not switch accountability on in one go. The record starts with the first routed request; the controls follow when you decide.
- Route. Point your agents at the Argent gateway. Every decision is signed from the first request that goes through it.
- Give every agent an owner. Register each agent to a named person, synced from your directory if you use one. AI in use that nobody registered shows up tagged Discovered.
- Watch. Start in watch mode. Requests pass unchanged, and what each control would have done is stamped into the signed record.
- Enforce, area by area. Switch one control at a time to enforce. Each switch is signed too, so you can show which posture held at any moment.
- Review as you go. Find a conversation by agent, owner and day, replay it and flag it with a note, or assign a case to put the agent on the needs-attention queue.
- Hand over proof. Give your auditor a read-only Auditor account. They build the evidence package themselves and check it on their own machine.
A walk-through: the question six months later
An example: an insurer's claims team uses an AI agent to draft letters to policyholders. Six months later, internal audit asks about one letter: which agent wrote it, for whom, from what, and who has looked at it since.
- Find it. The auditor opens Explore, asks what the claims agents did that week, and opens the result as a replay.
- See who was acting. The request card shows the chain from company to department, person, agent and client tool, each by name.
- Follow the conversation. The session timeline lists every request in order. Opening one replays it step by step, each step with the decision Argent made and its signed row.
- Read the words, on the record. The Auditor role sees the record, never the text, so a tenant admin opens the prompt and the reply. Before the text appears, that read is written to the signed record under the admin's name.
- Flag it. The admin flags the session with a one-line note, which is signed too.
- The owner has left. The adjuster who owned the agent left in March. Their account was switched off in the directory, and the agent was refused on its next request until an admin handed it to a new owner.
- Prove it. The auditor builds the evidence package for that month and checks it offline. It prints
TENANT PACKAGE OK.
Proof points: why the record holds up
An audit trail is only worth what it can survive. These are the properties an auditor can test, not take on faith.
- Signed one by one. Each record carries its own digital signature and is linked to the record before it, so a changed or removed record shows.
- No record, no request. A record that cannot be signed is not written, and by default Argent refuses new requests rather than serve one it cannot record.
- Sealed every day. At midnight UTC each day is sealed with a signed daily root that links it to the next day.
- Checked every hour. Argent's verifier re-checks sealed records against the signing keys every hour and logs every pass, so you can show the checks kept running.
- Alerts on a break. A chain break, a missed check or a key mismatch alerts the console, a webhook or email.
- Copies are on the record. Every evidence export is listed: who took it, when, the scope, rows, size and SHA-256.
- Checked by the auditor. The check runs on the auditor's own machine, against a key from their own records, and names any missing, extra, reordered or altered record.
Checked by the auditor, not taken on your word.
What the rules already ask for
Each rule below applies to a named audience. Argent gives the security team the signed records and controls these rules ask about, for every agent that works through it.
| Rule | Who it applies to | What it asks | Where Argent helps |
|---|---|---|---|
| HIPAA Security Rule, 45 CFR 164.312(b) | HIPAA covered entities and their business associates³ | Mechanisms "that record and examine activity in information systems that contain or use electronic protected health information"³ | A signed record of every request and decision, with session replay |
| NYDFS, 23 NYCRR 500.6 | Companies regulated by the New York Department of Financial Services, such as banks and insurers⁴ | Systems that "include audit trails designed to detect and respond to Cybersecurity Events", kept "for not fewer than five years"⁴ | A signed trail of every agent decision and admin change, with alerts on a break |
| NIST SP 800-53 Rev. 5, AU-9 | US federal information systems⁵ | "Protect audit information and audit logging tools from unauthorized access, modification, and deletion"⁵ | Records signed one by one and chained, checked every hour, readable by role |
For SOC 2 and ISO 27001 audits, the Reports page maps your records to the criteria as a suggested reading guide for your auditor. For the sector view, see Argent for banking, insurance, healthcare and the public sector.
Capabilities, integrations and where to go next
Accountability is several capabilities working on one record. Each has its own page.
- Capabilities in this goal: identity tethering, owner offboarding, one-click audit evidence, session replay, watch before you enforce, shadow AI discovery and give the board what it wants, or see all capabilities.
- Owners from your directory: Microsoft Entra, Okta, Ping Identity, SailPoint, SCIM and LDAP and Active Directory.
- AI provider keys in your vault: HashiCorp Vault and OpenBao. Agents: Claude Code, Codex CLI and the rest of the integrations.
- Related goals: compliance evidence, agent identity and ownership, see all AI in your company, stop data loss, threat defense, budget and token spend, model control and routing, sovereign AI and all goals.
- Use cases: AI in regulated and air-gapped environments, AI coding agents, customer-facing AI assistants, enterprise copilots and MCP and agent-to-agent tool access.
- Industries: financial services, insurance, healthcare, public sector, legal and all industries.
- Where it runs: air-gapped on premises, private cloud or your own public cloud VPC; compare options.
Take the walkthrough or get the data sheets.