Trust and security
AI that stays inside your rules.
This page is for risk, security and compliance teams. It sets out where the agents and models we build run, which models may see which data, how our engineers reach your systems, how autonomy is granted and taken back, and the record every decision leaves.
Runs where your data has to stay.
The same controls apply in all four setups, so choose by your infrastructure and your data rules.
Managed by Addo
The fastest start. We host, update and monitor the agents and everything they run on, with your data kept separate from every other client.
Dedicated
Single-tenant infrastructure reserved for you, in the region you choose, run by our operations team.
Your cloud
Everything runs inside your own AWS, Azure or Google Cloud account. Data stays in your perimeter and you control upgrades.
On premises
The full stack on your hardware with open-weight models, maintained by our engineers.
You decide which models see which data.
Each agent may only call the models you approve for each kind of data. We build that rule into each agent, record which model ran each step and set a spending limit for each workflow.
Inside your perimeter
Open-weight models such as Llama, Mistral and Qwen, in your cloud or on your hardware, for data that cannot leave.
Your cloud's model services
Azure OpenAI, Amazon Bedrock or Google Vertex AI, under your own account and agreements.
Frontier model APIs
OpenAI, Anthropic and Google through their enterprise APIs, only for the workflows and data you approve, with personal data masked before it is sent.
Client data is never used to train models for anyone else.
Autonomy is earned, then granted.
Rules are set by action, value and risk. An agent moves up a level only when its measured accuracy on your cases passes the bar you set and your workflow owner signs off the step up. It can be moved back at any time.
Level 1
Assist
Every agent starts here, in shadow. It prepares the work and your team decides.
Level 2
Approve
Routine cases complete on their own. Exceptions, high values and low-confidence cases wait for sign-off.
Level 3
Autonomous
The whole workflow runs within policy. Your team reviews samples and everything the agent escalates.
Access runs through your single sign-on with role-based permissions, and agents inherit the permissions of the source systems they read, so an agent sees only what those systems allow.
Every decision leaves a record.
Each agent checks policy before it acts. Afterwards your auditors get a record like this one: every check and its result, the rule version that applied, and who an exception would have gone to.
An example from accounts payable: one supplier invoice, read, matched and posted.
Example decision record AP-0193
Supplier invoice to posted entry
- Systems
- Supplier portal, ERP, audit log
- Autonomy
- Autonomous within tolerance. Above it, the accounts-payable lead approves.
- Policy
- Price variance tolerance 3%, version 7, owned by Finance
- Models
- Two, both approved for finance data. Each step records which one ran.
- Checks
- Extraction confidence 0.96, above the 0.90 threshold
- Supplier active, bank details unchanged
- No duplicate of this invoice in the supplier history
- Purchase order matched
- Goods receipt matched
- Line 6 is 2.4% over the agreed price, inside the 3% tolerance. Accepted, with a note for the supplier file
- Tax code valid for the supplier's country
- General-ledger account and cost centre taken from the order
- Posting period open
- Outcome
- Posted as document 5105-22831, 1 minute 13 seconds after the invoice arrived
Engineers work inside your perimeter, wherever they sit.
On site or virtually embedded, an Addo engineer is a named user in your estate, not a contractor with a copy of your data.
Identities you issue
Every engineer works under an account your IT team creates, with the roles and data scopes you set.
Access you control
Through your VPN, virtual desktops or bastion hosts, with your multi-factor authentication, and revocable by you at any time.
Nothing copied
Data, documents and credentials stay in your systems. Nothing is downloaded to Addo machines or shared between clients.
Logged and ended
Access is logged on your side, reviewed with you, and removed when the engagement ends.
Bring your security team in early.
Send us your security questionnaire, or ask for our ISO/IEC 27001 certificate and its scope. We walk risk, security and compliance teams through these controls on your own workflow.
Start with one workflow, under your rules.
A strategist and an engineer will look at how the work runs today, which controls it needs, and what agents can return.