Security
Security and control, stated plainly.
What an oktogonAI agent can touch, what it needs a person for, what it can never do and how it meets the EU AI Act. And, just as plainly, what we do not claim.
Deployment model
The agents run inside your infrastructure.
Each solution is deployed into your own environment and works against the systems your data already lives in: your ERP, CRM, mailboxes and document storage. It is not a shared service your records are copied into. The model endpoints each agent calls, and the data terms that apply to them, are agreed with you before deployment.
Your infrastructure
Deployed into the same environment as the systems it works with.
Model endpoint
Where it runs, and its data terms, agreed with you before deployment
The control layer
Four controls, in every deployment.
Scoped permissions
Each agent reads and writes only what its workflow needs. Permissions are set at deployment and enforced at runtime.
Approval rules
You decide what executes on its own, what needs a person's approval and what always goes to an employee.
Exception handling
Anything outside the rules is stopped and handed to a person with the reason attached. It is never guessed.
Activity log
Every action an agent takes is written to a log inside your environment.
EU AI Act
Compliant with the EU AI Act, as it stands today.
The Act was last amended by the Digital Omnibus, in force since 27 July 2026. This is what applies to agents that do operational work, and how every deployment meets it.
Outside the high-risk list
Quoting, order entry, purchasing, invoice processing and customer requests are not uses the Act lists as high-risk, and nothing we build is a prohibited practice. A workflow built for one customer is classified before it is built.
Agents say they are agents
Where an agent writes to someone directly, the message says it comes from an AI agent and on whose behalf, as Article 50 has required since 2 August 2026.
People stay accountable
Approval rules send material decisions to a person, and the activity log shows what each agent did and why.
Models from Code signatories
We build on models from Anthropic and OpenAI. Both have signed the EU's General-Purpose AI Code of Practice, which covers transparency, copyright, and safety and security.
In practice
One agent's permissions, written down.
The Quote Agent, in an example configuration: the same grant shown on the homepage. Every solution page lists its own.
Can
- Read incoming requests
- Access product data
- Check customer pricing
- Prepare quotes
- Update CRM
Needs approval
- Quotes above €50,000
- Discounts above 5%
Cannot
- Change master pricing
- Delete records
Segregation of duties
Where money moves, a person decides.
In finance workflows the agent prepares, matches and posts. It never approves its own postings and never releases a payment, and a change to supplier bank details stops the document whatever its value.
See the Finance AgentCompliance
What we claim, and what we don't.
- Complies with
- EU AI Act, GDPR
- Not held
- ISO 27001, SOC 2
Compliance with the EU AI Act and the GDPR is our own assessment, not a certificate. oktogonAI does not currently hold ISO 27001 or SOC 2 certification, and this page will say so until a certificate exists.
Models and partners
Model choice per workflow.
Use the right model for the right workflow without locking operations to one provider.
- Claude models
- GPT models
- Business applications
- CRM and marketing
oktogonAI is a member of the partner programmes above. Partner status covers the platforms oktogonAI builds on and integrates with; it does not mean any of these companies endorses a specific deployment.
This website
No cookies. Nothing loaded from anyone else.
This site sets no cookies, uses no browser storage and loads nothing from other domains: fonts and images are served from our own. The one request that leaves this domain is the review form, sent to the service that delivers it to us, and only when you press send. If analytics is switched on, it is Plausible, which is cookieless and stores no personal data.
Questions
Security questions, answered plainly.
Where do the agents run?
Inside your own infrastructure, against the systems your data already lives in. The model endpoints they call, and the data terms that apply to them, are agreed with you before deployment.
Can an agent do something it was not configured to do?
No. Permissions are scoped at deployment and enforced at runtime; an action outside them is not available to the agent.
What happens when an agent is unsure?
It stops and hands the case to a person with the reason attached. The threshold (confidence, value, customer, action type) is configured with you.
Is every action logged?
Yes. Every action an agent takes is written to an activity log inside your environment.
Do your agents comply with the EU AI Act?
Yes, with the Act as amended by the Digital Omnibus, in force since 27 July 2026. The agents do operational work outside the Act's high-risk categories, and where an agent writes to someone directly it says it is an AI agent and on whose behalf, as Article 50 has required since 2 August 2026. We build on models from Anthropic and OpenAI, both signatories of the EU's General-Purpose AI Code of Practice.
Do you hold ISO 27001 or SOC 2 certification?
No. oktogonAI does not currently hold ISO 27001 or SOC 2 certification. Our agents comply with the EU AI Act and the GDPR, which is our own assessment rather than a certificate.
Bring your security questions to the review.
The review session covers integration requirements and where the agent sits in your environment. It is the right place to put your checklist on the table.
Start with one workflow
Fixed scope for the first workflow, agreed before we start. Expansion is your decision once it proves value.
How it worksStart with quotes or orders
Two primary workflows first. Procurement, finance and customer operations once the first one proves value.
See the solutions