AI in workflows
In Beezr, AI is a step in a workflow. It reads a document, a web page or your own records, answers in exactly the shape your process needs, and hands the result to the next step, where your checks, approvals and permissions still apply.
You are here: AI in workflows.
Choose the provider and the model, and use your own API key.
A file field, a web address, a list from this workflow, or nothing at all.
What to do with it, as plain text or as a Smart Script that can use your records.
On the workflow canvas you press + on a line and pick a step. AI agent sits in the Data group with the JSON parser, the CSV parser and the custom function. Put it where the process needs it: after a trigger, after a Load data step, before an approval. The decisions, loops and approvals around it work as they always do.
Open an AI step and its panel has numbered sections. You choose a model, say what it reads, write the instructions, define the shape of the answer and name the result. The callouts below use the panel’s own numbers. Everything after that is an ordinary workflow.
order_linesList of objects×product_idNumber×quantityNumber×unit_priceNumber×The model is told to answer in exactly this shape.
✓ Unique in this workflow. Use letters, numbers and _.
Provider, model and your API key.
A content field, such as an attached PDF, a web address, a list from this workflow, or nothing.
What to look for and what to do with it. Plain text, or a Smart Script that can load your own records first.
Text, yes or no, a number, or fields, objects and lists. The model is told to answer in exactly this shape.
The answer gets a name. Later steps use it, and a list in it can be looped over.
Each AI step has its own provider, model and API key: OpenAI, Anthropic, Google or OpenRouter, which takes any model id you type. You bring the key and Beezr stores it securely. When a better model comes along, change the provider and model on the step.
Pick OpenAI, Anthropic, Google or OpenRouter, then the model. OpenRouter takes any model id you type.
The step uses your own account with the provider.
The saved key stays hidden. Leave the field empty to keep it, or type to replace it.
Instructions can be plain text or a Smart Script, with an Insert data button and a list of functions. So a step can load your own records before it asks the model anything. Here a Load data step puts the product list into the workflow, and the instructions read it, so the model matches each line of an order to a product you actually sell.
1(function () {
2 var products =
3 FunctionManager.getListFromContext("productList");
4 var names = [];
5 for (var i = 0; i < products.size(); i++) {
6 var f = products.get(i).getFieldValueMap();
7 names.push(f.get("name").getCurrentValue());
8 }
9 return "Match each order line to one of: "
10 + names.join(", ");
11})();An AI step returns an answer. What happens next is whatever you drew: put an approval task after it and the workflow waits until a person approves or rejects. It is the same approval task used everywhere else in Beezr, so the answer meets your checks before it reaches anyone.
Illustration of a Beezr workflow.
Pick one to see what the step reads, what it answers with and what happens next. These are ideas to start from, built with the same step. The values in the illustrations are examples.
Your checks, approvals and permissions still apply.
Illustrations of the AI agent step. Values are examples.
The AI step reads workflow inputs and hands its answer to the steps you choose.
AI is a step on the canvas, next to triggers, decisions, loops and approvals.
Workflows and approvalsThe step can read a file field on a record, or a list loaded from your records. A later step writes the answer back to a record.
Data and interfacesA step can read a web address directly. For images and scanned documents, the Integration Hub has document-reading connectors such as Amazon Textract, Google Cloud Vision and Mistral OCR.
IntegrationAn approval task after the AI step is how a person signs off. Permission profiles keep deciding who can see the records that result.
Security and governanceBeezr’s easy to drag and drop tools are built to customize and automate the processes in few clicks. Beezr’s SaaS model is recommended for anyone looking for cost effective, efficient, and productive workflows that can be amazingly quick to implement in real time.
Rekha SetpalHead of in5 ProgramOpenAI, Anthropic, Google and OpenRouter. You choose the provider and model on each AI step, and with OpenRouter you can type any model id it offers.
Yes. Each AI step uses a key from the provider you chose. The saved key stays hidden. Leave the field empty to keep the saved key, or type to replace it.
A file field on a record (PDF, Word, Excel, PowerPoint, text or HTML), a web address, a list from the same workflow, or nothing, in which case it works from the instructions alone.
Text, a yes or no, a number, or structured data made of fields, objects and lists. The model is told to answer in exactly the shape you define, and the answer is saved under a name that later steps use. A list in the answer can be looped over.
The AI step returns an answer. The next step is one you drew: a record update, an email, or an approval task that waits for a person to approve or reject.
Yes. The step sends its instructions and what it reads to the provider you chose, using your own account and key. Choose a provider whose terms fit the data you give it.
Yes. Instructions can be plain text or a Smart Script, with an Insert data button and a list of functions. A Load data step can put records into the workflow first, and the instructions can read them.
Reading a document, a page or a list and returning something structured: order lines from a purchase order, terms from a contract, a score for a resume, a category for a request. The eight examples on this page show the shape. Each one runs inside a workflow, with your own checks around it.
Show us the process you are trying to improve. We will tell you whether Beezr looks like a fit and how we would approach it.
