Access
Read-only.
Discovery never writes to your systems.
For CTOs and IT leaders
One governed platform, instead of a hundred personal setups.
Every team is wiring AI into your systems its own way, on personal tokens you never issued. Sonsara finds those setups and rebuilds the useful ones as company agents on one platform, with company credentials, versions, tests, approvals and a full audit trail.
82%
of pastes into AI chat tools come from unmanaged personal accounts, outside anything IT can see.
LayerX, Enterprise AI & SaaS Data Security Report, 2025
97%
of organisations whose AI models or apps were breached had no AI access controls in place.
IBM, Cost of a Data Breach, 2025
The sprawl
You're already running AI agents. You just didn't deploy them.
Browser agents, automations and assistants connected to your systems by the people who use them. Each one is an identity, a credential and an integration you'll be asked about, and none of them is in your inventory.
01 · Credentials
Tokens nobody issued.
Personal access tokens in prompts, OAuth grants to AI apps, connector keys in config files. None of them is rotated, scoped or revoked when the person moves on.
02 · Change
Changed at runtime, with no history.
A prompt is edited, a model swapped, a new tool connected. It all happens outside change control, so nobody can say what ran last Tuesday.
03 · Upkeep
Breaks quietly, lands on you.
When an API or a permission changes, a personal automation stops without telling anyone. IT finds out from the team it was quietly holding up.
Before and after
Same work. One place to run it.
What changes when a team's personal setups become company agents.
Area
Personal setups
One per person, unknown to IT
Company agents
One platform, governed by you
01Credentials
PersonalPersonal tokens pasted into prompts
CompanyCompany-held, scoped, on the agent's account
02Change
PersonalEdited in place, no history
CompanyVersioned, test-run, approved before going live
03Quality
PersonalTried once, then trusted
CompanyEvals per version; failing ones block approval
04Upkeep
PersonalBreaks silently when something changes
CompanyWatched while it runs; broken ones flagged
05Access
PersonalWhoever has the laptop
CompanyOwners, editors and users, with SSO
06Record
PersonalWhatever the tool happens to keep
CompanyEvery run, change and grant in one activity log
Discovery without a rollout
Nothing on laptops to start.
Discovery connects read-only to your identity provider and the audit logs of your work tools. One app consent or admin key per system, and first findings in days. The result is the inventory of AI tools and agents you've been asked for.
- Entra ID, Google Workspace and Okta for consents and sign-ins
- Microsoft 365, Atlassian, Slack and GitHub audit logs for usage patterns
- Optional later: existing DNS, gateway or SIEM exports, and a device check through your own MDM
Install
None to start.
No agents, proxies or extensions.
Data
Metadata only.
No prompts, files or messages.
People
Pseudonymised.
Re-identification is an audited action.
The platform
Built to be run, not just built.
Company agents are organised one parent agent per department, with sub-agents for specific jobs. People talk to the parent; you govern the whole tree, in one place, with SSO.
Connectors
3,000+ integrations.
CRM, finance, HR, support, storage and developer tools, held on the agent's account with scoped access. No keys in anyone's config files.
Versions
Tested, then approved.
A new version is test-run first and checked against its evals. The old one keeps running until an owner or editor approves the change.
Maintenance
Broken ones flagged.
When a connector disconnects, the agent is marked down and its owners are asked in Slack to reconnect. It doesn't land in your queue.
Audit-ready
Change control and logging that auditors recognise.
Agents get the same discipline as the rest of your estate: versions, test runs, named approvers and a log of every run, tool call, policy block, change and access grant. That record supports evidence for ISO 27001, ISO 42001 and SOC 2.
What Sonsara recordsISO/IEC 27001ISO/IEC 42001SOC 2
Every run and tool callWho asked, which systems and data, the outcome
A.8.15LoggingA.8.16Monitoring activities
A.6.2.8Event logsA.6.2.6Operation and monitoring
CC7.2System monitoringCC7.3Evaluating events
Every changeNew version, test run, evals, who approved it
A.8.32Change management
A.6.2.4Verification and validationA.6.2.5Deployment
CC8.1Change management
Every access grantOwners, editors and users; scoped company credentials
A.5.15Access controlA.5.18Access rights
A.9Use of AI systems
CC6.1–6.3Logical access
Every AI tool and agentFound by detection, registered in the vault
A.5.9Asset inventoryA.5.23Cloud services
A.4ResourcesA.10Third parties
CC9.2Vendor risk
Every policy blockActions blocked, tagged or held for a person
A.8.12Data leakage preventionA.5.34Privacy and PII
A.9Responsible use
CC6.7Data transmission
Run history kept 30 days · export it any timeOur mapping of control areas · your auditor decides what evidence they accept
Where people use them
No new app to roll out.
Approved agents answer in Slack, Microsoft Teams, Claude and ChatGPT. Same agent, same controls, whichever door someone uses.
- Slack: mention the department's agent in a channel or message it directly
- Microsoft Teams: chat with it like a colleague
- Claude and ChatGPT: call it from the assistant people already use
- Every run logged in one place, whichever surface it came from
FAQ
What IT leaders ask us.
Whose credentials do agents use?+
The company's. Connectors are held on the agent's account with only the access the job needs, so an agent doesn't depend on any one employee's login, and nothing breaks when someone leaves.
How do changes get to production?+
Every change makes a new version. It's test-run on real data with nothing sent or changed, checked against the agent's evals, and approved by an owner or editor. Until then the current version keeps running.
What do we have to install?+
Nothing to start. Discovery is read-only access to your identity provider and work tools' audit logs, and agents run on the platform, reached through Slack, Teams, Claude and ChatGPT.
What if a tool we use isn't connected yet?+
Most business tools are available through our 3,000+ integrations. When an agent needs a connector its owners haven't set up, they're asked to connect it.
Which models do agents run on, and where?+
Approved models only, chosen per agent, running in the data region you set, such as EU · Dublin.
Keep reading
Under the hood.
Book a demo
See the AI work already happening in your company.
Read-only access, nothing to install, first findings in days.