Agentic AI · Security
A Syntasa agent wants access to your customer data. Here is its permission slip.
Written forHeads of data platform, data engineering, security and privacy
5 min read
Every team in the building wants AI working on your customer data. You have to approve it, and you get the call when something goes wrong. Here is how to say yes and keep control.
Permission slip
- Agent
- Customer Insights Agent
- Acting for
- Priya, CRM manager. Never more than Priya can do.
- May
- Read orders and customer profiles. Create audiences. Email customers, with approval above 500 recipients.
- May not
Export data·See email addresses- Expires
- When the task ends, or in 8 hours
- Every step
- Logged, with the reason why
Write the slip. Then try to break it.
Set what the agent may do. Then give it a task and watch what happens. Go on, ask it to do something it shouldn’t.
The safest agent is the one that can’t do anything you didn’t write down.
Four rules it can’t break
Declared authority
What an agent may do is written down and reviewable. Never more than the person who asked.
Expiring credentials
Access is tied to a person, a task and a time limit. It ends when the task ends, or at the time limit, whichever comes first.
Approved destinations only
Agents work inside your perimeter. They can send data only to destinations you approve, and every transfer is logged.
A chain of whys
Every action is logged with the data, the model and the rule behind it.
Your data doesn’t move. Your rules don’t change.
Many AI tools ask you to copy your customer data into their system and rebuild your permissions there. Then you have two sets of rules to keep in step, and two places an attacker can reach.
Syntasa agents run where your data already lives. Your warehouse does the work, so your access policies, encryption and audit logs all still apply. There is no second copy and no second set of rules. Each permission slip maps to a role and service account in your warehouse, so the warehouse enforces it, not the agent.
- Google BigQuery
- Snowflake
- Amazon Redshift
- Azure Synapse
- Databricks
- Trino
- Your VPC
- GovCloud
- Fully air-gapped
You choose the model. A hosted model is one of the destinations you approve, and you see exactly what it receives. To send nothing out, run an open-weight model yourself.
Built for the strictest reviews. FedRAMP High authorized and ISO/IEC 27001 certified, with support for GDPR and CCPA/CPRA compliance.
customer records unified into one view, inside the retailer’s own cloud
to production for Syntasa customers
What changes for your team
Fewer tickets
Marketing and ecommerce teams answer their own questions, inside rules you set once.
Fewer new controls to review
Agents use the access controls you already approved. Saved audiences and logs stay in your own warehouse.
AI that people can actually use
AI gets used fastest when it is the easiest thing to approve.
What you get from Syntasa
Syntasa’s agents, including the Customer Insights Agent and the Decisioning Agent, run inside your cloud and answer to your rules.
- A permission slip for every agent. Your data team sets and changes it. The person using the agent can’t widen it.
- Access tied to a person, a task and a time limit, which ends when the task ends.
- A log of every action, with the data, model and rule behind it.
Customer Intelligence is Syntasa’s platform for customer data. The Customer Insights Agent and the Decisioning Agent run on it.
Your first step: book 30 minutes with John and a Syntasa engineer. Bring your security questions and your access policies. You leave with a written list of what an agent could and couldn’t reach in your setup.
Put our agents through your security review.
Bring your hardest questions and your strictest policies. Our engineers show you what an agent can reach, what it can’t, and the log it leaves.
- On your cloud, or fully air-gapped.
- With your own access policies in place.
- With a Syntasa engineer on the call.
Questions we get
Only to destinations you approve, such as your email platform. Every transfer is logged. Agents run inside your boundary and work on your data where it lives.
Every action traces to a named person and to a slip your team approved. An agent acts for that person and can never do more than they can.
Yes. A hosted model is one of the destinations you approve, and you see exactly what it receives. To send nothing out, run an open-weight model yourself, even fully offline.
The permission slips, the roles and service accounts they map to, and the log. The same controls you already use for people.
Also in this series: what the Customer Insights Agent does with that access.
John Bizimana
Director of Enterprise GTM at Syntasa. You can reach him atjohn.bizimana@syntasa.com.