The same stages as any business, with evidence written for a supervisor. Before building, map each obligation that touches the process to a specific control. Then test on your own cases, run in shadow, start at drafting, and widen per capability. At each stage, keep the record a supervisor or auditor would ask for, and name the senior person accountable.
What changes when the business is regulated?
Three things. Obligations already apply to the process, and the agent inherits them: complaint deadlines, fair treatment, record-keeping, data protection, outsourcing rules. Someone senior is named as accountable, often by regulation. And you must be able to show, not just assert, that controls work.
So the deployment stages stay the same, but each one produces more evidence, and a new stage comes first: mapping the rules. The general stages are in how to deploy an AI agent in a business.
Which obligations usually apply?
| Area | Examples of where it comes from | Control | Evidence |
|---|---|---|---|
| Risk classification | EU AI Act risk categories; Annex III lists credit scoring of individuals as high-risk | Classify each agent's use before build | Classification record |
| Data protection | GDPR Articles 5, 6 and 35; equivalent national laws | Lawful basis, minimum data, impact assessment | Assessment and data access decision |
| Automated decisions | GDPR Article 22; ECOA and Regulation B adverse action reasons in the US | Human involvement where required; stated reasons | Held verdicts; reason records |
| Human oversight | EU AI Act Article 14 for high-risk systems | People able to understand, stop and override | Stop switch tests; approval records |
| Record-keeping | EU AI Act Article 12; sector record rules | Automatic, append-only record of every decision | The record and a reconstruction test |
| Model risk | SR 11-7 in the US; PRA SS1/23 in the UK; OSFI E-23 in Canada | Validation, monitoring, inventory | Test results; monitoring reports |
| Third-party and operational resilience | DORA in the EU; APRA CPS 230 in Australia | Vendor assessment, exit plans, incident routes | Due diligence file; resilience tests |
| Accountability | Senior Managers and Certification Regime in the UK | Named senior owner | Responsibilities map |
Not every row applies to every business. Which ones do depends on your sector, your market and what the agent does. Markets differ too: see how AI agents are regulated in Australia or in India for two examples.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What are the practical steps?
- Compliance and the process owner: map obligations to controls. One line per obligation, as in the table above.
- Risk: classify the use. Is any part high-risk, or a decision with significant effects on individuals?
- Data protection officer: complete the impact assessment and agree what data the agent may see.
- Process owner: describe the job, with hand-over rules that reflect the obligations, such as vulnerable customers going to a specialist.
- Experts: test on your own cases, including cases where the right answer was to refuse or escalate.
- Run in shadow, then draft for people to approve.
- Accountable executive: approve go-live on the evidence pack.
- Widen per capability on evidence, with risk's agreement, and recertify on a fixed cycle.
What does this look like in practice?
Take a mid-sized bank in Germany deploying an agent to prepare responses to business banking customers' payment queries. Compliance maps the obligations: GDPR for customer data, DORA for the ICT third-party arrangement, BaFin's supervision as the national authority, and the bank's own complaint-handling rules. The use is not high-risk under the EU AI Act, since it makes no credit decisions, and that classification is recorded with its reasoning.
The impact assessment limits the agent to the account and payment data the queries need. Testing on past queries includes payment disputes that turned into complaints, so the test proves the agent escalates them. The agent drafts at depth 2 for six months before routine status replies move to depth 3. A similar setting is described in AI agents for operations leaders in business banking in Germany.
Which controls do supervisors look for first?
- A named accountable person.
- An inventory of agents and what each may do.
- Evidence from your own cases, not vendor claims.
- A record that reconstructs any decision.
- A working stop switch and an incident route that includes agent errors.
Where data must stay inside your own network, hosting choices change; see deploying an AI agent on-premises.
Where Pryme fits
Pryme was built for this setting. Depth is earned per agent on your own cases through Attest, certificates lapse after 90 days, every action passes five bounds at the moment of action, and every verdict is filed in an append-only record. That gives each row of the obligation map something concrete to point to. More in deploying an AI agent: train, shadow, certify, widen.
Questions people ask
Should we tell our supervisor before deploying?
Check your supervisor's expectations; some expect notification for material outsourcing or new technology. Being able to explain the agent clearly helps either way.
Do we need a separate AI policy?
Many firms add an agent policy beneath their existing risk framework rather than starting a separate one.
Is a pilot exempt from the rules?
No. If the pilot touches real customers or data, the obligations apply from the first day.