Qualification is the process of making an AI agent fit for a specific job at a specific customer by changing what surrounds the model, its brief, sources, bounds and certification, rather than the model itself. An agent is qualified when it has been certified on that customer's own cases, at the depth the evidence supports.
Why does qualification matter?
It answers a question every regulated customer asks: will our data be used to train your model? Under qualification, no. We never update model weights for a customer. Everything specific to the customer lives in parts that can be read, changed, reversed and audited. The reasoning is set out in qualification, not training.
It also makes the agent explainable to a supervisor. A brief is a document someone can read. A source list can be checked. A limit can be shown. Weights cannot be inspected in any useful way, so an agent adapted by retraining asks the supervisor to trust a change nobody can see.
How does qualification work in practice?
- Brief: describe the role, the tasks, what the agent must not do and when it must hand over. Agent Studio can assemble a draft from a plain description of the job.
- Sources: choose the documents and systems the agent may rely on, each with its provenance.
- Bounds: set the customer's own limits, such as amounts, products and times.
- Certification: replay the agent against the customer's filed cases and score it at every gate.
- Depth: grant the depth the evidence supports, capped by the role's ceiling, as described in earned depth.
- Upkeep: monitor, and recertify when something changes or the certificate nears expiry.
What does that look like in an example?
A building society deploys a pre-trained mortgage servicing agent. Qualification starts with the brief: the society's own products, its rules on payment date changes, and an instruction to hand any mention of financial difficulty to the arrears team. Its sources become the society's product terms and arrears policy, not generic guidance. Its limits cap any fee waiver at a small amount. It is certified on 300 of the society's own filed cases. Payment date changes earn depth 3; conversations about arrears earn depth 2. No model was trained, and every one of those choices is on file.
How is qualification different from training?
Training changes the model's weights using data, and the change is baked in: hard to inspect, hard to reverse, and shared by anyone else using that model. Qualification leaves the model untouched and changes what surrounds it. If a source is wrong, you remove it. If the brief is unclear, you rewrite it. Each change is recorded and can be undone, and each one triggers recertification where it affects the agent's behaviour.
Qualification is also wider than certification. Certification is its final step, the proof. Qualification includes the work of making the agent worth certifying in the first place.
Questions people ask
Does qualification mean the agent never improves?
It improves through better briefs, better sources and better bounds, each tested before it takes effect. What it never does is absorb the customer's data into the model.
Who does the qualification work?
The customer's capability owner, who knows the job, with tools that draft the brief and run certification. Risk and compliance review the bounds and floors. The full method is in what it means to certify an agent.