AI and the Sequencer
The Sequencer can be used to send structured requests to AI services, receive AI responses, and use those responses inside controlled Autologyx workflows.
This makes Autologyx Catalyst a model-agnostic orchestrator for AI-augmented processing. Catalyst remains responsible for the workflow, data model, permissions, routing, task creation, and audit trail, while AI services can be used for the parts of a process where language understanding, classification, extraction, summarisation, or similar analysis adds value.
Use this section when you want Catalyst to start the AI interaction as part of a workflow.
How this pattern works
In this pattern, Autologyx starts and controls the AI interaction.
A typical Sequence might:
- collect structured data from an Object record,
- prepare a prompt, file, text extract, or structured payload,
- send an outbound request to an AI service,
- receive a response,
- store the result on a record,
- create, assign, or complete a Task,
- route the workflow based on the response.
This is different from an AI assistant or agent initiating a request into Autologyx. With the Sequencer pattern, Catalyst is the originator. The workflow decides when AI is used, what context is sent, and what should happen when the response comes back.
When to use AppliedAI and the Sequencer
Use AI and the Sequencer where AI should support a defined workflow step.
This pattern is useful where you need:
- predictable process control,
- repeatable AI requests,
- structured inputs and outputs,
- workflow routing based on AI analysis,
- human review at defined stages,
- integration with records, tasks, documents, and notifications,
- auditability around how AI is used.
Common examples include:
- extracting key terms from contracts or documents,
- classifying inbound requests,
- summarising long documents or message chains,
- triaging work into the right process path,
- identifying risk indicators,
- preparing draft outputs for human review,
- converting unstructured content into structured record data,
- running playbook-style prompt functions against a document or record.
Autologyx as the model-agnostic orchestrator
AppliedAI in the Sequencer is designed to be model agnostic.
This means the workflow can be designed around the business process rather than around one AI provider. Depending on the customer environment and available integrations, Catalyst can send structured requests to different AI platforms, hosted models, private endpoints, or AI gateways.
The same process design can support different model choices where appropriate. For example, one workflow step might use a model that is well suited to summarisation, while another step might use a different model or endpoint for classification, extraction, or reasoning over a longer context.
This helps customers choose the most appropriate model and provider for the task, while keeping the surrounding workflow stable.
Native AI connectivity
Catalyst can support AI interaction through dedicated actors and integrations for commonly used AI platforms.
These may include:
- OpenAI,
- Anthropic,
- Azure OpenAI and Azure OpenAI,
- AWS Bedrock,
- privately hosted models or customer-specific AI gateways.
Dedicated actors are useful where the platform provides a native configuration experience for a provider. They can make it easier to configure prompts, model settings, authentication, request inputs, and response handling without building a full custom API request from scratch.
Provider availability and supported models depend on the customer environment, licensing, configuration, and enabled integrations.
Using the API Call actor with AI services
Where there is no dedicated actor for a specific AI service, Catalyst can use API-based integration patterns.
The API Call actor can be used to initiate outbound REST requests to an AI endpoint, gateway, or privately hosted model. This gives configuration teams flexibility where a provider is not covered by a native actor or where a customer has a specific AI service architecture.
An API Call actor pattern might be used to:
- send record data to an AI REST endpoint,
- send extracted document text for analysis,
- include structured instructions or prompt variables,
- call a private model endpoint,
- call a customer-controlled AI gateway,
- receive a JSON response,
- store returned values on the source record,
- pass returned data to later actors in the Sequence.
For more information, see Using the API Call actor with AI services.
AI tooling for processing content
AppliedAI workflows often need more than a single prompt and response.
Depending on the process, the Sequencer may also use supporting tools and actors for:
- vectorisation,
- text processing,
- document conversion,
- file preparation,
- record lookup,
- relationship lookup,
- looping over repeated items,
- storing intermediate values,
- updating records with structured outputs,
- triggering downstream tasks or communications.
These supporting steps help turn unstructured content into structured information that can be used by the rest of the workflow.
For example, a workflow might convert a document into usable text, split the analysis into several prompt functions, call an AI service for each function, and store the results in structured fields on a record.
Playbook-style prompt functions
Some AI workflows benefit from separating prompt logic from the Sequence design itself.
One pattern is to use an Object Class to store reusable playbook functions. Each playbook record can define a specific extraction, classification, summary, or review instruction.
A Sequence can then use those playbook records to run a set of repeatable AI checks. For example, a contract review process might use playbook functions to extract:
- parties,
- term,
- renewal provisions,
- termination rights,
- governing law,
- liability caps,
- data protection obligations,
- unusual or missing clauses.
This pattern is especially useful where the same document or record needs to be tested against several review questions. It can also make prompt management easier because prompt functions are held as structured configuration data rather than hidden inside a single long Sequence step.
Where multiple playbook functions need to run against the same source content, the Sequence may use a Loop Actor pattern to repeat the AI request for each function and collect the outputs.
What happens to AI responses
AI responses are most useful when they are brought back into the workflow as structured information.
Depending on the process design, a response can be used to:
- populate record fields,
- create or update tasks,
- add a summary to a record,
- generate a draft output,
- trigger a review step,
- route the record down a different path,
- prepare content for a message or document,
- support a human decision.
Where an AI response will affect workflow routing, task creation, data changes, or user-facing outputs, the Sequence should be designed carefully. Customers should consider whether the result should be reviewed by a user before it is used to drive important actions.
Human review and controlled automation
AI and the Sequencer is not an all-or-nothing automation pattern.
Some workflows may use AI only to prepare a draft summary or recommendation. Others may use AI to classify or extract data, but still require a user to review the result before the workflow continues. In lower-risk or highly repeatable use cases, some AI-supported steps may be allowed to continue automatically.
The right design depends on:
- the risk of the process,
- the quality and consistency of the source data,
- the confidence required for the output,
- the consequences of an incorrect result,
- the permissions and governance model,
- the customer's internal AI usage policies.
An important reminder about AI-generated output
Important reminder
Large language models are not deterministic in the same way as conventional workflow logic.
Even where a prompt, model, workflow, and integration have been carefully designed, an AI model may occasionally produce an inconsistent, unexpected, incomplete, or inaccurate response. This can include hallucinated content.
There are techniques that can reduce this risk, such as carefully designed prompts, structured outputs, validation steps, comparison against source data, deterministic checks in the Sequencer, and secondary model review. However, Autologyx cannot absolutely guarantee that a model will never produce an incorrect or unexpected result.
Where an AI response produces a specific structured value, the Sequencer can often apply deterministic logic to check or gate what happens next. For example, the workflow may test whether a returned value matches an expected option before continuing.
This is more difficult where the model produces output that requires semantic evaluation, such as a narrative summary, legal analysis, interpretation, or recommendation. In those cases, customers should strongly consider a human-in-the-loop review before the output is used for final decisions, external communications, record updates, or other important workflow actions.
How this differs from MCP services
AI and the Sequencer and MCP services are complementary, but they are not the same pattern.
With AI and the Sequencer:
- Catalyst starts the request,
- the workflow decides when AI is used,
- the Sequence controls the context and next step,
- AI responses are processed inside the workflow.
With MCP services:
- an approved AI assistant or agent starts the request,
- the assistant uses defined tools to work with Autologyx,
- the agent may read records, work with tasks, retrieve knowledge, or perform controlled actions depending on the enabled service,
- tool access is governed by the MCP service, permissions, and customer configuration.
Use the Sequencer pattern when the workflow should control the AI interaction. Use MCP services when an AI assistant or agent needs to work with Autologyx through defined tools.
For more information, see MCP services.
Things to remember
- The Sequencer pattern is best where Catalyst should initiate and control the AI interaction.
- AI services should receive only the information needed for the task.
- Native AI actors can simplify common provider integrations.
- The API Call actor can support AI endpoints that are not covered by a native actor.
- AI responses should be treated as workflow inputs, not automatically trusted decisions.
- Use human review where the output affects important decisions, record data, tasks, or communications.
- Model choice, prompt design, permissions, auditability, and customer AI policies should be considered before using AppliedAI in production workflows.