The Problem
Deterministic flows are predictable, but they're also brittle and expensive. Every question a customer might ask, every branch and edge case, has to be anticipated, designed, built, and maintained by hand. Complex journeys — the ones that touch several systems and depend on the customer's specific situation — explode into sprawling flow diagrams that take weeks to build and break the moment a backend changes.
The AI Use Case
An agentic flow gives the AI a goal and a set of capabilities — APIs it can call, knowledge bases it can search — and lets it reason out how to resolve the query, calling what it needs, when it needs it. Instead of hard-coding "if this, then that" for every path, you describe the tools and the guardrails and the agent orchestrates the rest.
- Pulls live data from APIs — order status, account details, eligibility — to answer with the customer's actual situation.
- Draws on knowledge bases — for the policy, product, and how-to context around the task.
- Chains steps dynamically — handling multi-part requests without a pre-built branch for each.
- Saves build effort — far less deterministic flow design to create and maintain.
This is the leading edge of the shift from deterministic to agentic self-service, and increasingly what platform capabilities like Amazon Connect's agentic tooling are built to enable.
The Value
- Resolves genuinely complex queries that were impractical to script.
- Much less build and maintenance — describe tools and goals, not every branch.
- Adapts to the individual — responds to the customer's real data, not a generic path.
- Faster to change — add a capability rather than rewire a flow.
The Risks — Read This Part
Agentic flows trade control for flexibility, and that trade brings real risks that have to be designed for from day one:
Latency
Every reasoning step and API call adds time. An agent that thinks, retrieves, calls two APIs, and reasons again can leave the customer sitting in silence — deadly on a voice call. You need aggressive latency budgets, parallel calls where possible, streaming responses, and filler that keeps the caller informed.
Trust — will it do the right thing?
Handing an AI the ability to act (not just answer) raises the stakes. Will it call the right API with the right parameters? Will it refuse to do something harmful or out of scope? Guardrails, allow-lists of permitted actions, confirmation steps before anything consequential, and human handoff on low confidence are non-negotiable.
Predictability & testing
A deterministic flow does the same thing every time; an agent may not. Testing shifts from "walk every branch" to evaluating behaviour across many scenarios, with ongoing monitoring in production.
Don't go agentic everywhere at once. Use deterministic flows where the path is fixed and the cost of a wrong step is high (payments, identity, anything regulated). Reserve agentic flows for the complex, variable journeys where their flexibility genuinely pays off — and put firm guardrails around what they're allowed to do.
Because you're giving AI the power to act, this is exactly the kind of capability that warrants a proper AI risk assessment and clear success criteria before it goes live.
Explore More AI Use Cases
Part of a growing series of practical AI use cases for contact centres. Browse the full set on The AI Use Cases hub, and see the wider picture in 15 ways AI can transform your contact centre. Got one you'd like covered? Get in touch.