Workflow AI Action Nodes

Overview

AI (Artificial Intelligence) is the simulation of human intelligence processes by machines, especially computer systems, to perform tasks such as learning, reasoning, problem-solving, and decision-making. There are literally thousands of ways in which AI services can be added to business processes (digitised in Softools) to deliver steps changes in performance improvement.

This article will run through what the AI capabilities are in the Softools Workflow AI Nodes and how to implement them:

• Choosing the correct AI Node
• Typical AI Workflow Structure
• AI Workflow Nodes:
  • Send Azure OpenAI Classification Node
  • Send Azure OpenAI Prompt Node / Send Managed AI Prompt
  • Send OpenAI Web Search
• AI Deployment & Governance
• Tips & Best Practices
• Azure OpenAI Privacy
• Inspirations for More AI Use Cases

Choosing the correct AI Node

Softools offers a variety of AI Nodes that perform different operations. Some are managed and others self-managed nodes.

Self-Managed Classification Node

This allows you to use AI to categorise data into a pre-defined list of Options in a Select List. You supply the AI with Field data that gives context to the Record and then the AI will classify the Record.

Examples:

• Creating a SWOT analysis based on information written about a Company "Classify this statement about my Company into Strengths, Weaknesses Opportunities or Threats" 
• Categorise Support Tickets based on description "Read the ticket below and assign it to one of the following categories ..."
• Identifying the sentiment in customer feedback "Classify the feedback as positive, negative or neutral"

Managed Prompt Node

This utilises AI capability to of prompt engineering which is the process of designing, optimising, and refining inputs (prompts) given to AI models, such as language models, to achieve desired outputs or behaviour effectively and efficiently. You define:

• SYSTEM: What the role of the AI is such as a Risk Consultancy identifying risks for new projects.
• USER: The item to ask the system such as What are the Risks of Building a second Warehouse.
• ASSISTANT: The expected format of the output such as the top 10 associated Risks.

For the managed prompt it uses a Softools managed Azure OpenAI deployment. The model is updated for you as newer versions become available; it currently targets GPT 5 Chat. This does mean that the responses will be based on the models pre-training cut-off so latest data or niche facts may be missing.

Good For: Summarising, innovation, strategising, marketing content or providing well known facts

Examples:

• Creating Marketing content: "Generate 10 catchy blog titles and outlines about the use of AI in manufacturing plants"
• Producing a concise summary of an Inspection "Give me headline information from the following Long Text Fields"
• Give a starting point coaching plans "What are the 5 best training initiatives for this employee based on their performance review?"

Self-Managed Prompt Node

This node configuration in Softools will be the same as the Managed Prompt node but the flexibility here is that you are in control of how the AI operates. 

You can choose the AI model and you can also use your own database of content that processing will be done on. The advantage of this is that you can use specific internal databases of knowledge to provide a response from or by calling the data in a Softools App when the Prompt runs, get insights about the data you have in your Softools Applications. 

Good For: Same as the managed Prompt but with the flexibility of choosing the model. Working across own data libraries, it can provide better internal specific insight. It can also call Softools data into the model to give insights into Softools App data.

Examples:

• Generate a response to a Client Support ticket based on an Internal Knowledge Centre as the data source "How do I update my email address?"
• Procedural and regulation checks "Does my latest Marketing Campaign comply with our GDPR Policies?"
• Summarise a trend on a set of Softools records "What is the improvement in Compliance based on the past 5 GMP Audits including the most common failures?"
• Identify similar Records "What other projects do we have planned that would be good to consider running in parallel to this one?"

Managed Web Search Node

This allows configuration of an input that then performs a web search to return a response with websites that it has used cited in the items returned. The benefit of a web search is that it can fetch live data and so has access to more current facts, news and trends. It worth noting that as it is a web search it could be disrupted by noise and inconsistent quality. It can also be more costly as it needs to issue a search before compiling and feeding into a model.

Good For: It is good for when you need recent information, niche facts, references and URLs and up to date information

Examples:

• Perform competitor analysis with up-to date information on pricing models and features "What are the best Project management tools in 2025 for price and functionality?"
• Finding company information from the internet given a company name "Give me a description of who Softools are and what they offer as a company"
•  Identify the best materials to use for product development "Find recent (2025) market data about global sustainable packaging trends and summarize top 3 insights"

When to use Prompt or WebSearch

Knowing when to use the prompt node and when to use the web search is based on the following

Use a prompt node only when:

• You’re dealing with static or well-known information (e.g., explaining a concept, generating creative text)

• Up-to-date accuracy is not critical

• You want simplicity and speed

Use web search tool when:

• You need the latest facts, recent events, references to specific web sources

• You want the model to include citations or external links

• You are building something like a knowledge-assistant, news summariser, or retrieval-augmented workflow

Suppose you’re building a workflow: “User asks: What are the latest regulatory changes in data privacy in the UK?”

• Using prompt node only: you ask the model “What are the latest changes…” and the model answers based on its knowledge (which may be outdated).

• Using web search: the workflow triggers a search for “UK data privacy regulation changes 2025”, feeds results to the model, then the model synthesises an answer with citations and up-to-date info.

• Cost implication: For the web search you have to issue a search query, process results, feed into model, maybe handle tool integration. This will likely and more cost or overhead of maintaining the AI flow.

The current costs for Open AI can be found here in the OpenAI pricing documentation.

Typical AI Workflow Structure

AI Nodes are one action in a Workflow. An AI workflow will need a minimum of one trigger and two actions.

Trigger: What will cause the AI processing to run

AI Action: What AI Action will happen to get a response

AI Response Action: What will we do with the response

Ideally we would also add in tracking for the AI Nodes which can be done following the set-up in our AI Token Tracking Article.

To demonstrate this, lets look at an example where we are running Inspections and for each 

Send Azure OpenAI Classification Node

This node allows you to utilise the power of Azure OpenAI for text classification tasks directly within your workflows. This can be used for:

• Automating decision-making based on text inputs.
• Classifying text into predefined categories.
• Enabling smarter workflows using AI-driven insights.

Destination URL: This is the endpoint where your prompt data will be sent. Make sure to use the proper URL of your deployed Azure OpenAI model and include the api-version query parameter.

• Example URL:

  https://api.openai.azure.com/v1/classifications

API Version: Specify the version of the Azure OpenAI API you are targeting. This ensures compatibility between your workflow and the OpenAI model.

• Example:?api-version=2023-06-01

API Key: Enter your Azure OpenAI API key, which is required for authorisation. This ensures that only authenticated requests are processed by the model.

Max Tokens: Set the maximum length of the response in tokens. This defines how much data the AI can generate or classify for a given input.

Temperature: Adjust how "creative" or "focused" the AI responses should be, adjust this setting based on the nature of your classification task:

• Low (0.0–0.4): Focused and deterministic responses.
• High (0.5–1.0): Creative and diverse responses.

Response Variable Name:

You can define a variable name to store the response. This makes it easier to use the results in subsequent workflow steps. This can follow into another node like an email or update field value 

• Example: Var.ClassificationResponse

Classifications: Choose from a list of your select lists for the AI model to classify the text into. The model will choose the best-matched category.

Context Fields: Provide text fields that contain the input data you want to classify. You can include up to 5 fields, enabling flexibility for diverse text classification tasks.

Practical Use Cases

• Sorting customer feedback into categories like "Positive," "Negative," and "Neutral."
• Automating ticket routing based on content (e.g., technical vs. billing issues).
• Classifying emails or text inputs for workflow triggers.

Send Azure OpenAI Prompt Node / Send Managed AI Prompt

These nodes allow you to use the capabilities of Azure OpenAI's chat completion API to generate AI-powered responses directly within your workflows. It is versatile and can be used for tasks such as:

• DATA QUALITY: Address issues of data completeness and quality in legacy systems / Excel
• DOC DATA EXTRACTION: Extract summary data from a linked / uploaded Document
• DATA KEY WORDS & CLUSTERING: Identify common labels for data records to accelerate search and retrieval
• COMPARATIVE ANALYSIS / MATCHING: Identify database records similar to the current record & situation
• DATA CREATION / IDEAS SUGGESTIONS: Generate solutions for common business issues (PromptAI/GenAI)
• SUMMARIES: Create summary reports for Executives based on multiple records

Each use case requires carefully crafted prompts to maximize the relevance, accuracy, and efficiency of the AI model’s responses.

Azure OpenAI Prompt Node vs. Managed AI Prompt Node

The Azure OpenAI Prompt Node allows you to configure the AI node to your own managed AI services and requires an API Key for authorisation with Azure OpenAI. The Managed AI Prompt uses a Softools managed Azure OpenAI deployment. The model is updated for you as newer versions become available; it currently targets GPT 5 Chat.

Azure OpenAI Prompt Node Properties

If you are using your own services to run the Azure OpenAI Prompt Node then you will need to complete the following properties on the workflow node. If you are using the Managed AI Prompt Node then this infrastructure details are handled and inbuilt by Softools so not needed.

Destination URL: This is the endpoint where your prompt data will be sent. Ensure the URL corresponds to your deployed Azure OpenAI model and includes the api-version query parameter.

• Example URL: https://api.openai.azure.com/v1/completions

API Version: Specify the version of the Azure OpenAI API being targeted. This ensures the model responds appropriately.

• Example: ?api-version=2023-06-01

API Key: Enter your Azure OpenAI API key. This is required for authentication to ensure only authorised requests are processed.

Temperature: Adjust the randomness and creativity of the AI's output:

• Low (0.0–0.4): More focused and predictable responses.
• High (0.5–1.0): Creative and varied responses.

Azure OpenAI & Managed AI Prompt Node Properties

The remaining properties will be in both the Managed and Self-Managed versions of the AI Prompt node.

Response Variable Name: Define a variable name to store the AI's response for use in subsequent workflow steps, such as email generation or updating field values.

• Example: Var.PromptResponse

Max Tokens: Set the maximum number of tokens for the AI’s response. This limits the length of the output.

Message Structure: Messages are used to structure the conversation for the Azure OpenAI chat completion API. Each message includes a role and content:

• Roles:
  • system: Defines the AI assistant’s behavior.
  • user: Provides input from the user. This is usually dynamically referenced via expression to a text Field as the user inputs different information each time to get a response.
  • assistant: Represents the AI’s responses.
• Example Message:
  • System: "You are a risk advisor."
  • User: "I'm starting a project creating an application to track our projects"
  • Assistant: "Lay out the risks with a score of 1-5 for severity and likelihood"

Practical Use Cases

• Generating natural language summaries from text fields.
• Providing dynamic and tailored responses to customer queries.
• Drafting email content or report text based on workflow data.

Send OpenAI Web Search

This node is a powerful way of gaining AI insights into Business Process by searching the Web and returning a summary which also includes the websites that have been cited for the information. This node is simpler to configure than the Azure Open AI Prompts and it also does not require an API Key so you are able to harness the power of AI without the need for managed services to run the AI

This is a very versatile function which means the use cases are limitless but lets look at a few.

• Finding Innovative Solutions to items raised in an Issue Log
• Performing Competitor Analysis for a Product Procurement process
• Defining Business Strategy based on Performance in a KPI Scorecard App

Configuration takes 2 steps. The first performs the Web Search and pulls back a summary. This is then stored as a variable in the workflow so it then needs a second action to push that summary value into a Field in the Record.

Step 1: Perform the Web Search

The Web Search needs the Search Query that we want to use. The idea here is to try and be specific with your prompt by adding as much context from the Record as we can. This then gets stored in a variable.

In our example above we are using the use case where we have an Asset Register and each Asset has quarterly inspections. The inspector will write his recommendations in his report but we want a quick way to summarise this and gain insight from a web search to make these into confirmed actionable items.

Response Variable Name: The Response Variable needs a reference which in this case is 'Response' we can then call this later on using [Var.Response] to use the value returned from the search into a new node.

Search Query As Expression: The search query is what we are asking the Web. The more specific we can be the better quality response we will get. Using our example, we could have searched for [InspectionRecommendations] 

This is very limited in accuracy so we can ask the format we want the response to be in such as the 5 best actions to take

'Give me the 5 best actions based on: ' + [InspectionRecommendations]

This is better but if we really want to refine our search we can add some context such as the location of our company and company size. This will mean the response is far better targeted to what we want out of the search.

'Give me the 5 best actions based on: ' + [InspectionRecommendations] + '. Factor in that the company location is ' + [Location] + ' and company size is ' + string([NumberOfEmployees]) + ' employees.'

Notes: Notes do not have any effect on the functionality of the node but they are useful to remind yourself why this workflow node has been used or list any limitations or planned future improvements. It is particularly of value when collaborating on the build with other App Builders.

Once configures, click on OK to add and confirm edits to the node.

Step 2: Push the Stored Variable into a Field

We now choose what we want to do with our new found AI insight. The simplest choice is to place the value back into the Record in a new Field.

Continuing our use case, we have performed a web search based on out Inspectors Recommendations and now have a variable that is storing the response.

We can reference this variable into a Field Update by using the expression [Var.Response] with the variable name in place of 'Response'.

For more on creating records and updating field values see these workflow nodes here.

Here is a sample response for a simple use case of what to do given the recommendation to replace the office coffee machine. You can see that where appropriate websites references have been cited.

AI Deployment & Governance

Pre-deployment

Before deploying your AI service on a production environment it is best to give it a test run and to ask yourself the following questions.

• Does it actually run / work?
• Does it give the quality of results you need?
• Does it hallucinate – make up values?
• Does it cost too much to run?

Post-deployment

Because AI services are changing so rapidly we recommend that you schedule a review of all AI services every quarter.

• New AI services are available that offer better results
• The cost of the service may have changed making some AI services too expensive or others now cheap enough to deliver a clear return on investment

Tips & Best Practices

Here are some general guidelines:

• Add multiple messages to the make the required response clearer
• Prompt Engineering is a combination of art and science - accept that every time you run the AI workflow it may create a different answer!
• Use trial and error to adapt your messages until you get consistent results in the format you need.
• Specify exactly what you want: e.g., add a User Message to tell the AI service to add the URL for any sources it provides, or ask it to provide the results in plain text or perhaps in Japanese.
• Be explicit with the expected response e.g. System - "Only return the percentage value, without % symbol"
• Users can access the full json response, also useful in SendWebhook node e.g. jsonvalue([Meta.ResponseContent], 'choices[0].message.content', 'text')

Azure OpenAI Privacy

See the microsoft support article for assurance on data privacy when using AI prompts through Microsoft Azure OpenAI services

https://learn.microsoft.com/en-us/legal/cognitive-services/openai/data-privacy?tabs=azure-portal

The key points are that

Your prompts (inputs) and completions (outputs), your embeddings, and your training data:

• are NOT available to other customers.
• are NOT available to OpenAI.
• are NOT used to improve OpenAI models.
• are NOT used to train, retrain, or improve Azure OpenAI Service foundation models.
• are NOT used to improve any Microsoft or 3rd party products or services without your permission or instruction.

The Azure OpenAI Service is operated by Microsoft as an Azure service; Microsoft hosts the OpenAI models in Microsoft's Azure environment and the Service does NOT interact with the OpenAI API, which is what Softools uses to provide AI services through Microsoft Azure

Inspirations for More AI Use Cases

1. Content Generation

· Writing articles, blog posts, or essays.

· Creating marketing copy or product descriptions.

· Generating creative content like poetry, stories, or scripts.

2. Text Summarisation

· Summarising long documents, articles, or reports.

· Extracting key points from meetings or research papers.

3. Data Analysis and Querying

· Generating insights from structured and unstructured data.

· Creating prompts to simplify complex data queries.

4. Language Translation

· Providing accurate and context-aware translations.

· Preserving tone, cultural nuances, and technical language.

5. Learning and Tutoring

· Designing educational content and personalized quizzes.

· Explaining complex topics in simpler terms.

6. Problem Solving and Ideation

· Brainstorming ideas for projects or businesses.

· Providing solutions to technical or strategic problems.

7. Sentiment Analysis

· Analysing and categorizing sentiment in text, such as customer feedback or social media posts.

8. Personalised Recommendations

· Generating tailored recommendations for books, movies, or travel plans.

· Creating prompts for e-commerce product suggestions.

9. Testing AI Models

· Evaluating and improving the model's behaviour with edge cases.

· Designing prompts to test specific functionalities or biases.

10. Legal and Compliance Support

· Drafting contracts, agreements, or legal clauses.

· Verifying compliance with specific regulations or policies.

11. Decision Support

· Generating pros and cons for decision-making.

· Analysing scenarios and providing risk assessments.

12. Customer Support

· Crafting templates for frequently asked questions.

· Escalating and solving complex customer issues via AI-powered responses.