IBM watsonx.ai

Bring IBM watsonx.ai foundation models into your workflows. Agents run text generation, chat, and question answering against Granite and other hosted models, create embeddings, tokenize text, and list available foundation models.

6 actions OAuth available

IBM watsonx.ai website ↗ Platform Documentation ↗ Capability data verified 2026-07-25

A supplier email with an attached document lands in the operations mailbox

Ask a Question classifies the message as an order change, a complaint, or an invoice query

Create Embeddings converts the document text into vectors for retrieval against the indexed policy set

Chat drafts a grounded response from the retrieved passages, with any tool calls the model requests captured rather than executed

The agent inspects the finish reason, token usage, and every requested tool call

The operations lead approves the drafted reply and each requested record change before either one happens

The approved reply sends and the approved updates post to the case record

Token counts and the outcome log to the run report for cost tracking

What This Integration Enables

IBM watsonx.ai is what enterprise AI looks like when procurement, legal, and audit all get a vote: governed foundation models including IBM's Granite family, hosted in a region you choose, billed to a project you name, behind IBM Cloud IAM rather than a bearer key pasted into a script. For the operations teams FlowRunner serves, that pedigree is the point. The model is not a toy in a browser tab; it is infrastructure that has to answer the same questions as every other system of record. FlowRunner agents put that infrastructure to work inside governed workflows: classification, summarization, and rewriting with Generate Text, multi-turn and multimodal reasoning with tool calling through Chat, single-shot answers with Ask a Question, and retrieval pipelines built on Create Embeddings. - The live model catalogue drives model selection, so new releases appear without a connector update - Tokenize Text prices a prompt against a model's context window before a single billable token is spent - Embeddings return in input order, ready to zip back onto source records for search and RAG - Tool calls requested by the model route through [human-in-the-loop](/concepts/human-in-the-loop) approval when the tool changes something real

Without FlowRunner

AI adoption stalls at the review committee The models the team wants to use cannot answer where the data lives or who governs the output

Model calls are scattered and unaccountable Prompts run from personal scripts and notebooks with no record of what was asked or what it cost

Agentic tool calls run on trust When a model requests an action, whatever wired it up executes it, and nobody reviews the request

With FlowRunner

Inference runs inside the governance story Granite and other hosted models run in a chosen region, billed to a named project, logged per call

Every call carries its receipts Token counts, finish reasons, and model IDs land in the workflow log where audit can find them

Tool calls become approval requests The model proposes, the workflow routes the proposal, and a person disposes when the action has teeth

Use Case Scenarios

A policy answer desk grounded in your own documents

The team indexes operating procedures by chunking each document, sizing chunks with Tokenize Text, and embedding them with Create Embeddings into [Pinecone](/integrations/pinecone). When a question arrives from the field, the agent embeds the query, retrieves the nearest passages, and passes them to Chat with a system message that requires citations. The answer posts to [Slack](/integrations/slack) with its source passages attached. Questions the retrieval cannot ground get escalated to a person instead of answered anyway.

Inbox triage that respects the escalation path

Shared-mailbox traffic in [Outlook](/integrations/outlook) runs through Ask a Question for classification and a one-paragraph summary. Routine order confirmations file automatically. Anything classified as a complaint or a contract question lands in the queue with its summary, and the drafted reply from Chat waits for the account owner's approval before sending. The mailbox stops being a rotation duty, and the sensitive messages still get human eyes every time.

A record backlog cleaned by a governed model

Operations inherits a spreadsheet of eight thousand free-text vendor records. The agent pulls rows from [Google Sheets](/integrations/google-sheets), normalizes and categorizes each with Generate Text using greedy decoding for repeatability, and writes results back with a confidence flag. High-confidence rows update in place; the ambiguous remainder queues for a human pass. The cleanup that was never going to happen finishes in days, with the judgment calls still made by people.

Human-in-Loop Highlight

Chat is where this connector stops being a text generator and starts requesting actions: the model can return tool calls, structured requests to run a function with arguments it chose. Executing those automatically is how an inference step quietly becomes an unsupervised actor in your systems. So in FlowRunner, a tool call against anything consequential is a proposal, not a command. When the triage agent's model requests a credit adjustment tool with an amount it inferred from an angry email, the flow holds: "Model requested issue_credit for 340 against account 8827, citing the attached thread. Execute?" The operations lead reads the thread, approves or rejects, and the decision is logged next to the token counts and the trace ID. The model's reasoning runs at machine speed; its reach into systems that move money or touch customers is exactly as long as a person allows.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via email

Human decides

Agent resumes with decision

Agent Capabilities

6 actions

Text Generation

Generate Text Runs a single-prompt completion against any watsonx.ai foundation model. The prompt is sent verbatim, so it carries its own instruction formatting and few-shot examples. Supports greedy or sampling decoding, token limits, stop sequences, repetition penalty, and a per-request time limit, returning the generated text with stop reason and billed token counts.

Chat

Ask a Question Convenience wrapper over the chat endpoint for the single-question case: a question and an optional system instruction in, the assistant's answer as plain text out, no messages array to assemble.
Chat Sends a full multi-turn conversation to a chat-capable model, with system, user, assistant, and tool roles, multimodal content parts for capable models, tool and function calling, JSON-only responses, sampling controls, and stop sequences. Returns the reply, any requested tool calls, the finish reason, and token usage.

Embeddings

Create Embeddings Converts one or more texts into dense vectors with a watsonx.ai embedding model for semantic search, clustering, RAG, or similarity scoring. Vectors return in input order with their dimensionality and billed token count; over-length inputs are rejected unless a truncation limit is set.

Models

List Foundation Models Lists the foundation models available for inference in the configured region, with model IDs, providers, supported functions, context and output limits, and billing tiers. The discovery step behind model selection, readable without a project.
Tokenize Text Splits text into the tokens a specific model would see and reports the count, using that model's own tokenizer. The step that checks a prompt against a context window and estimates cost before anything is billed.

Frequently Asked Questions

What can FlowRunner do with IBM watsonx.ai?

FlowRunner agents can run Generate Text, Chat, and Ask a Question in IBM watsonx.ai, plus 3 more actions.

Does connecting IBM watsonx.ai to FlowRunner require OAuth?

Yes. IBM watsonx.ai connects to FlowRunner with OAuth 2.0, so agents authenticate without handling raw credentials.

Can IBM watsonx.ai trigger a FlowRunner workflow automatically?

IBM watsonx.ai doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.

Start building with IBM watsonx.ai

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us