Scenario 3. Procure an agentic call centre solution
This section examines how agencies can procure an AI-enabled capability from the market. A common example is an agentic call centre solution, where agencies assess vendor offerings to determine whether they provide sufficient governability, oversight and assurance for operation in a government context.
Key challenges
Agencies going to market for an agentic solution face a wide vendor landscape. Call centre solution presents a procurement‑led architectural challenge. Agencies are assessing agentic services that may interact directly with the public, act across multiple backend systems, and initiate actions on the agency’s behalf.
As agentic systems are an emerging capability, key aspects such as orchestration, decision pathways and action boundaries may be partially or fully hidden behind vendor-managed services. This can create a risk that procurement decisions are influenced primarily by visible features and demonstrations rather than a clear understanding of the architectural properties required to govern the system over time.
Feature-based claims can mask important differences in architectural control, oversight capability, trustworthiness and assurance. They may also fail to reveal action scope, escalation logic, oversight boundaries and granularity of control, all of which can materially affect an agency's ability to govern the service once deployed.
How AIRA supports this scenario
AIRA supports this by structuring the decision space around architectural properties rather than feature lists, enabling agencies to compare offerings on how effectively governability is engineered into the service from the start.
In summary, AIRA supports this scenario as follows:
- Compare offerings using architectural criteria: Use a structured market assessment based on AIRA domains and layers. As different vendors may expose different capabilities and control mechanisms, AIRA helps agencies compare how agency control is exercised across offerings, including where control is primarily through governance, where agencies can additionally configure behaviour, and where greater design flexibility is available.
- Assess communication capabilities: In a procurement context, communication matters because it determines how the agentic system integrates with agency backends. Agencies procuring a call centre solution need to understand whether the offering can integrate with their customer relationship management (CRM), case management and telephony systems, and under what control conditions.
- Assess offerings for sufficient traceability: Use AIRA to evaluate whether offerings provide the visibility, auditability and attribution needed to reconstruct actions across interactions, system behaviour and integrations. Where interactions are transformed, such as through real-time language translation, agencies may need to determine whether original interactions and AI-transformed content are preserved for review, audit, and oversight.
- Assess trustworthiness and assurance over high-consequence interaction: Compare how offerings handle high-consequence interactions across the trustworthiness and assurance domain. Some high-consequence interactions include legally sensitive, financially consequential or crisis-related interactions, as well as interactions with users who face accessibility barriers or have limited ability to engage with or challenge system behaviour.
- Compare offerings for how they enable human agency, oversight and control: Compare how offerings provide supervisor visibility, timely escalation, pre-action review and other meaningful intervention pathways.
- Identify the simplest offering: A practical architectural objective is to identify the simplest offering that can operate within clear, reviewable limits and be governed over time. This includes explicit action scope, meaningful intervention pathways, sufficient disclosure and assurance to support ongoing oversight, change-notification obligations and exit provisions that preserve governability as the service evolves. Where these conditions cannot be met, agencies may need to narrow action scope, require stronger disclosure and contractual control, or defer procurement rather than accept reduced governability.
Figure 3: Agency control profile for an example agentic call centre triage solution
Shows an example control profile based on publicly available documentation of a real agentic call centre solution. In practice, agencies may create a control profile for each shortlisted vendor offering. Comparing these profiles helps reveal differences in governability, control boundaries, assurance requirements and associated risks before selecting a solution.
Legend:
- (G) govern
- (C) configure and govern
- (D) design, configure and govern
Detailed scenario analysis
Functional stack
For an agency procuring an agentic solution, architecture can help understand primary control levers across vendor offerings against a common functional stack illustrated in Table 7.
| AIRA Layers | Indicative capabilities and components |
|---|---|
| Experience |
|
| Application and Orchestration |
|
| Intelligence and analytics |
|
| Data and metadata |
|
| Platform and infrastructure |
|
The layers that are most consequential for governance – application and orchestration and intelligence and analytics – are often the least visible to the buyer. Articulating the stack helps agencies identify which levers they hold directly, where they depend on vendor disclosure and how limited visibility into those components affects their regulatory, compliance and T&A posture. This analysis could inform both the disclosure requirements and the contractual control points.
Communication
Agentic call centre systems introduce a more complex communication surface than static AI tools or simple chat interfaces. Within a single interaction, the agent may query multiple backend systems, retrieve policy content, update customer records, trigger downstream workflows and transfer to a human agent. Each of these actions can involve different authorisation requirements and audit implications. Each integration point represents a trust boundary and a potential failure mode that AIRA’s communication stack helps make visible.
Indicative communication capabilities that can be used to compare vendor offerings include:
- interaction orchestration across channels, systems and services
- integration with agency systems (CRM, case management, policy systems)
- granularity of control for data and actions across trust boundaries
- handoff and escalation communication between agent and human operators
- interoperability using open standards for agent communication, tool invocation and observability.
Trustworthiness and assurance
T&A analysis structures how agencies articulate architectural requirements for disclosure and compare responses consistently.
Table 8 shows T&A domains applied to an agentic call centre context. The domains frame the architectural properties agencies may need to understand and evaluate to support this assessment. The table is illustrative rather than exhaustive.
| AIRA Group | Key architectural drivers | What architecture can help assess |
|---|---|---|
| Observability | The system operates in real time across multiple systems and interactions. | How execution steps, tool calls, action outcomes, escalation events and error conditions are instrumented and exposed; whether the agency has independent monitoring access or relies on vendor-presented views |
| Traceability | The system may act on behalf of citizens and decisions may need to be reconstructed. | How each agent action, transformation and decision point is logged with sufficient fidelity to reconstruct what happened, on whose behalf and why; how actions taken on a citizen's behalf are attributed and retained |
| Explainability and interpretability | Decisions may need to be understood by staff, reviewed later or explained to affected individuals. | How the system provides a human-readable basis for each triage decision or escalation, sufficient for a receiving human agent to understand what the agent did and why |
| Validity and reliability | Triage decisions and escalation logic must behave predictably and consistently over time. | Testing evidence for escalation logic reliability and exception handling coverage, intent classification performance on agency specific cases, knowledge base currency controls |
| Safety | The system interacts directly with citizens, including vulnerable or high risk situations. | Scope-enforcement design as the primary agentic safety control; mandatory escalation for crisis callers, vulnerability disclosures, complaints and legally sensitive interactions; abstention and graceful degradation behaviours |
| Fairness | The system serves diverse user groups with varying needs, contexts and access channels. | Fairness testing across language diversity, accessibility needs and digital-literacy variation, given a service population spanning the full public; whether triage outcomes disadvantage particular cohorts |
| Sustainability | The system depends on evolving models, integrations and vendor-managed components. | Model-update cadence and the likely impact of model changes on accuracy, safety and fairness; change-notification obligations for model, action-scope and infrastructure changes; dependency and exit provisions; cost forecasting and controls |
| Security and resilience | The system integrates with multiple backend systems and operates across trust boundaries. | Whether IRAP and security evidence cover the actual components and integration points in scope; residual risks from out-of-scope integrations |
| Privacy | Interactions may involve Indigenous data, PII or sensitive information handled across systems. | Data minimisation controls (the agent retrieves only what the interaction requires); handling, retention and disposal of personal information; data flows across integration points and jurisdictions |
Human agency, oversight and control
In this scenario, architecture focuses on how different offerings enable HAOC in agentic systems. Because the capability is procured rather than built, the architectural task is comparative: understanding how effectively each offering supports meaningful human control over system actions in practice.
| Indicative capability | What agencies may need to understand and compare across offerings |
|---|---|
| Action scope and limits | How each offering defines and enforces what the agent can read, write, trigger or escalate and where those constraints are applied across systems. |
| High-consequence interaction handling | How each offering identifies and handles interactions with legal, financial or safety implications, including whether it supports pre action human review where applicable. Agencies may also need to assess how ambiguous or weak signals (such as potential vulnerability) are handled, including whether these result in escalation to human operators. |
| Escalation and handoff | How escalation thresholds are defined, how consistently they are applied and how interactions are transferred to human operators, including what context is preserved, |
| Intervention and override | What mechanisms exist for human operators to pause, override or terminate system actions and how quickly and effectively those interventions can be applied. |
| Accountability and attribution | How system actions, agent decisions and human interventions are recorded and attributed and whether responsibility can be clearly reconstructed and reviewed. |
| Monitoring and supervisory visibility | What level of visibility is provided into system components, behaviour, including whether agencies can independently monitor interactions for each customer cohort |
| Accessibility and user agency | Whether interactions are understandable and usable across different user groups and how limitations in accessibility may affect the ability of users to engage with or challenge system behaviour. |
When approaching the market, architecture can help agencies clarify with vendors which controls are built into the system at design time, which are configurable to agency context, which run-time responsibilities remain with human agents and supervisors and what skills those roles require to operate the service safely and effectively.
Across these views, AIRA points towards procuring the simplest service that can operate within clear, reviewable limits: an explicit, scope-enforced action surface; escalation and handoff engineered in; audit access and evaluation capability sufficient for the agency to govern the service over time; and fault and incident responsibilities between vendor and agency defined in advance. Where those conditions cannot be met, the response is to narrow action scope, require stronger disclosure and contractual control, or defer – to not accept reduced governability for surface-level capability.