AI customer support that combines deterministic and natural-language workflows lets the AI agent handle the conversation in plain language while fixed code or rules run the steps that must happen the same way every time, such as identity checks, eligibility, refund amounts and the order of write actions. Salesforce Agentforce, Lorikeet, Zendesk AI agents, Fin (now part of Salesforce), Cognigy, Ada, Decagon and Sierra all document a version of this on their own sites. They differ mainly in where the fixed logic lives: a script language, code tools, scripted dialogues, if/else blocks or sandboxed code.

Key takeaways

• Make a step deterministic when different people would reach the same answer from the same inputs: eligibility, thresholds, calculations, required ordering. Leave intent, clarifying questions and tone to the model.

• Rigid scripts alone fail customers. The CFPB says chatbot "doom loops" are often caused when an issue falls outside the bot's limited capabilities.

• Free-form agents alone fail audits. OWASP lists excessive autonomy as a root cause of LLM apps taking damaging actions.

• Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing costs, unclear value or inadequate risk controls. A hybrid design is how you show the controls.

Which AI support platforms combine deterministic and natural-language workflows?

Lorikeet writes conversations as natural-language skills and runs payments and identity checks as code the agent can invoke but never alter. Salesforce says Agentforce uses hybrid reasoning, combining LLM reasoning with deterministic, rules-based execution in the same engine. Zendesk pairs goal-oriented generative procedures with scripted dialogues. Fin's Procedures are written in natural language, with if/else logic and code where rules must hold. Ada computes refunds and eligibility in sandboxed Python while its reasoning engine runs the conversation. All eight are compared below.

How we chose and checked these

We included platforms whose own pages or documentation describe both a natural-language way to define agent behaviour and a deterministic mechanism (code, script, rules or scripted flows) inside the same agent. Every vendor fact was checked on the vendor's public pages on 30 October 2026 and is linked where it appears. We did not use analyst estimates, review scores or private conversations. The order reflects how explicitly each vendor documents the two modes working together. It is not a quality score.

Lorikeet publishes this guide and is one of the tools listed. We have tried to describe every product, including our own, using only what each vendor publishes, and we link to the source for each claim.

What does combining deterministic and natural-language workflows mean?

It means one AI agent uses two kinds of instruction. A natural-language workflow is a procedure written in prose that the model follows and adapts as the conversation moves. A deterministic workflow is code, a rule set or a scripted flow that gives the same result from the same inputs, every time. Anthropic's engineering guide describes the same split: workflows orchestrate models and tools "through predefined code paths," while agents "dynamically direct their own processes and tool usage."

A refund request shows why you want both. The customer may open with two questions, change their mind, or give the order number halfway through. A natural-language procedure handles that. But whether the order is eligible, how much to refund and the rule that identity is verified before the refund call are not matters of judgement. Those run as code, and the model only receives the result to explain.

When should a step be deterministic?

Make a step deterministic when getting it wrong costs money, breaks a rule or cannot be undone. Typical candidates:

• Identity and authorisation checks before any write action.

• Eligibility, limits and thresholds, such as refund windows or dispute values.

• Calculations: refund amounts, fees, dates.

• Disclosures that must be given in fixed wording.

• Several system calls that must always run together and in order.

Keep the rest in natural language: working out intent, asking for missing details, handling more than one request at once, and voice conversations where callers interrupt and correct themselves.

Comparison table

1. Salesforce Agentforce: Salesforce teams that want the split written into one script

Salesforce's Agentforce guide is the most explicit public description of the hybrid model we found. Hybrid reasoning is "Agentforce's approach to combining probabilistic LLM-based reasoning with deterministic, rules-based execution in the same engine." Builders draft rules in plain language in Agentforce Builder and express reliable behaviour in Agentforce Script, which Salesforce says gives "full deterministic control over agent behavior." With Script, actions can have deterministic transitions that always run after an action completes, and session variables carry facts between subagents so the agent does not ask twice. A Testing Center validates behaviour before deployment, and Agentforce Observability shows where to refine logic once agents are live.

Pricing: Flex Credits or Conversations, or per-user licensing.

Trade-off: it assumes Salesforce is your system of record and that you have admins comfortable with a script language. Outside Salesforce, it is a platform build.

2. Lorikeet: regulated support where strict steps sit inside natural conversations

Lorikeet's how-it-works page sums up the design as "the interface is permissive, the operation is strict." In the product docs, a concierge prompt sets tone, routing and business rules, and skills hold each procedure in plain language. Any decision that "different people would answer the same way given the inputs" moves into a code tool, which composes smaller tools in a fixed order, for example checking order status and then issuing a refund. Triage workflows run deterministic pre-checks when a conversation starts, and wrap-up workflows run fixed steps on escalation, auto-close or after a call. The docs also cover natural-language workflows that call node-based structured subworkflows when steps must run in a set order, such as verifying identity before a write.

GiveCard uses this shape: each recipient intent triggers its own workflow, which already knows whether verification is needed and runs that program's verification steps through GiveCard's API. Carmoola moved from rigid, form-driven automation resolving about 30% of inbound questions to 60% resolved end to end.

Pricing: Start $2,100 a month, Scale $5,100 a month, paid annually, plus per-resolution rates; escalations are not charged. 30-day free trial.

Limits: deterministic logic in the middle of a conversation lives in code tools, which are JavaScript, and the docs now label the node-based workflow editor as the older model. Teams that want a visual flowchart for every step will find Cognigy closer to that.

3. Zendesk AI agents: Zendesk teams choosing per use case

Zendesk's help center describes goal-oriented generative procedures and scripted dialogues as two ways to control replies, with authorized actions and API integrations alongside. A generative procedure is a flexible flow that uses agentic AI, and it can insert actions, API integrations and links to other procedures or dialogues, so a flexible conversation can hand one step to a scripted flow. Zendesk's AI agents page says you describe workflows in natural language and agents reason through multi-intent requests.

Pricing: AI agents are included in every Suite and Support plan, billed per automated resolution.

On voice, Zendesk says AI agents handle complex call flows from start to finish, and automated QA evaluates every interaction. It is the natural first test for a team already on Zendesk Suite that wants to decide, use case by use case, how much freedom the agent gets.

Trade-off: the help center notes generative procedures do not currently support rich-text formatting such as buttons or carousels, and the full setup lives inside Zendesk.

4. Fin: support teams that want to write both modes in one document

Fin's Procedures are written in natural language, and Fin says it "combines agentic behavior with deterministic control." Inside a Procedure you add if/else logic and use code to enforce rules like verifying eligibility, calculating dates or updating records, and connect systems through Data Connectors or MCP. Fin reasons at every turn, so a customer who interrupts or switches topic moves to the right step. Simulations test full conversations and can be rerun after every change. Procedures work on Intercom or with Zendesk, Salesforce or Freshdesk. Fin can draft a Procedure from an outline of your process, and Fin says Procedures are faster to set up and easier to maintain than its earlier Tasks approach.

Pricing: $0.99 per outcome.

Trade-off: the deterministic parts live inside each Procedure, so shared rules used across many Procedures need discipline to keep consistent.

5. Cognigy: contact centres that want a visual flow builder with generative nodes

Cognigy's documentation says the platform combines agentic AI with deterministic, rule-based automation, and that its intent-driven NLU is for "building deterministic or rule-based processes." Its AI Agent Studio, the Flow editor, is a low or no-code visual interface for business users, conversation designers and developers, with 100+ channel and system connectors and a native voice gateway. It can sit in front of existing contact centre platforms. Agents can answer from ingested knowledge sources such as Confluence using retrieval, and Cognigy's xApps add small web apps inside a conversation for transaction-ready steps. It suits large contact centres with conversation designers who already think in flows.

Trade-off: it is a broad contact centre platform, so expect a larger build than a support-only tool, and pricing is not published on the pages we checked.

6. Ada: brands that want calculations out of the model

Ada's platform page describes Playbooks as structured workflows that follow multi-step SOPs with real-time data, while the reasoning engine decides the next step. The deterministic part is explicit: "refunds, entitlements and eligibility come from your sandboxed Python, not from the model." Ada Computer then carries out each step in your systems, in a decide, act, check loop that runs until the task is done. MCP tools stay off until enabled, and agents deploy across voice, messaging and email. It suits high-volume brands that want the model choosing what to do next but never computing a refund itself.

Trade-off: pricing is not published, and sandboxed Python needs engineering ownership.

7. Decagon: enterprises that want procedures reviewed like code

Decagon's AOPs "combine the flexibility of natural language with the reliability of code." CX teams write procedures like SOPs, engineers version them with Git-based tracking, and Decagon explains the agent's reasoning at any point in a conversation, with runtime and latency numbers. Its Duet assistant can draft AOPs from uploaded conversation transcripts, AOPs pull live data and trigger workflows across systems, and guardrails for brand voice, escalations and hallucinations apply across chat, voice and email. It suits enterprises that want CX teams writing procedures and engineers reviewing them.

Trade-off: the public AOP page does not show how the code side is authored, and pricing is not published.

8. Sierra: brands that want journeys with controlled system access

Sierra's Agent Studio builds agents from composable building blocks and step-by-step workflows, with traces of every decision and tool call. Its trust page says access to systems of record is "deterministic and controlled," and that LLMs are wrapped in supervisory layers. Agent Studio lists 40+ pre-built integrations, no-code building, simulations and a regression test suite, and Sierra covers voice, chat, email and WhatsApp with outcome-based pricing. It suits large consumer brands that want one vendor across many channels.

Trade-off: outcome-based rates are not published.

How do you test a hybrid workflow before launch?

Test the two layers separately, then together. Unit-test the deterministic parts with fixed inputs and expected outputs. Then run simulated conversations, including interruptions, multi-intent openers and adversarial prompts, and check that the right deterministic step fired at the right moment. Save those simulations and rerun them after every prompt, model or procedure change. NIST's AI Risk Management Framework is a useful structure for recording what you tested and who approved it.

What still needs a human?

Deterministic code makes steps consistent; it does not make them right. People still own the rules themselves, compliance sign-off on wording and thresholds, exceptions the rules did not anticipate, and any case where the customer disputes the outcome. OWASP recommends human approval for high-impact actions before they are taken.

How to choose

1. List your ten most common ticket types and mark each step as judgement or rule.

2. Check where the vendor puts rules: script, code, scripted flow or if/else blocks, and who on your team can maintain that.

3. Ask whether shared rules can be reused across procedures, so a policy change is made once.

4. Confirm you can simulate full conversations and rerun them after every change.

Further reading: deterministic versus natural-language workflows, agentic AI versus workflow chatbots, regression testing when models change and AI agents that take backend actions. To build a skill with code tools on your own tickets, start a 30-day free trial.