Most lists of workflow automation examples describe automations instead of showing them. You read that employee onboarding is a good candidate, you agree, and you are no closer to having one. Each of the ten examples below is a real graph on a canvas you can scroll through, and you can copy any of them into an editor and press run. This post says what each one does, which model or service does the work, and what you must add before it becomes an automation rather than a demo.
What counts as a workflow automation example
A workflow is a set of steps with the data dependencies drawn between them. Automation is what happens when something other than a person starts it. Conflating the two is why so many examples of workflow automation read as inspiring and help nobody.
An example of a workflow is the shape: this input feeds that step, which feeds two more that run at the same time. An example of an automation is the shape plus a trigger and a destination: the same graph, started by a schedule or an incoming request, writing its result where a person will find it. Every template below is the first. The section near the end covers turning it into the second, the step these lists skip.
Ten workflow automation examples
These ten templates ship with Dafthunk. Each is small on purpose, three to five nodes, because its job is to be read in one glance and modified, not deployed as-is. Every canvas below is the live template, not a screenshot.
| Example | What it does | Input | Engine |
|---|---|---|---|
| Web Screenshot | Capture a full page as an image | A URL | Cloudflare Browser Rendering |
| Text Summarization | Compress long text to a summary | Pasted text | bart-large-cnn |
| Sentiment Analysis | Score text as positive or negative | Pasted text | distilbert-sst-2 |
| Text Translation | Translate between languages | Text, source, target | m2m100-1.2b |
| Speech to Text | Transcribe audio, with word timings | Recorded audio | Whisper |
| Text to Speech | Read text aloud as audio | Pasted text | MeloTTS |
| Image Generation | Draw an image from a prompt | A prompt | SDXL-Lightning |
| Image Description | Describe a picture in words | A canvas drawing | uform-gen2-qwen |
| Image Processing | Run an image through a filter chain | Webcam capture | Photon (no model) |
| AI Calculator | Let a model call a tool for arithmetic | A word problem | Llama 3.3 70B |
1. Capture a web page as an image
The Web Screenshot template is three nodes: a text input holding a URL, a screenshot node running on Cloudflare Browser Rendering, and an image output. The browser node drives a real headless Chrome, so pages that need JavaScript to render come out right. That is the reason to use a browser instead of fetching HTML.
On its own this is a demo. It earns its keep when the URL stops being typed by hand: point it at a list of competitor pages on a daily schedule, store each capture, compare against yesterday, and the same shape becomes change monitoring. The screenshot node is one of ten browser nodes, so the same starting point extends to page text, links, Markdown or structured JSON instead of pixels.
2. Summarize long text
Text Summarization is a text input, a summarizer pinned to bart-large-cnn with a max_length setting, and a text output. BART summarizes; it is not a general chat model. That makes it fast and cheap, and it means the model extracts and compresses instead of reasoning.
For meeting notes and article digests that trade is right. When the job needs judgement about what matters, swap the middle node for an Anthropic, OpenAI or Gemini node and keep the rest of the graph.
3. Score sentiment on incoming text
Sentiment Analysis is a text input, a distilbert-sst-2 classifier, and a JSON output. It returns labels with confidence scores instead of prose, and that detail makes it composable: a number can drive a branch, a paragraph cannot.
This is the classifier half of a triage automation. Add a conditional fork that reads the score, and negative feedback routes to a human while everything else gets filed. We covered the failure mode in building effective agents: a small classifier routes a vague input confidently to the wrong branch. Gate on the confidence score, not the label alone.
4. Translate between languages
Text Translation is the first example here with more than one input. Three text inputs feed one node: the text, a source language code, and a target language code. The model, m2m100-1.2b, translates directly between language pairs instead of pivoting through English.
The single translation is the dull version. The useful one is the fan-out: one source string, six target-language nodes running at once, because the runtime runs independent branches in parallel without being told to. You draw the shape, and the shape is the parallelism.
5. Transcribe audio
Speech to Text is an audio recorder input, Whisper, and a text output. The transcriber emits more than the transcript: a word count, per-word timings, and a WebVTT track.
So subtitles and timestamp search are wiring, not new capability. The outputs already exist; the template leaves them unconnected.
6. Turn text into speech
Text to Speech is a text input, MeloTTS with a language setting, and an audio output. It inverts the previous example, and the two compose: transcribe, translate, re-speak.
7. Generate an image from a prompt
Image Generation is a prompt input into SDXL-Lightning, then an image output. The node carries negative prompt, width, height, step count and guidance as settings, not buried in a config file.
8. Describe an image in words
Image Description is a canvas you draw on, a vision model, and a text description out. The canvas input keeps the template self-contained in a browser, and it is the first thing to replace.
9. Chain image filters
Image Processing is the one example with no model in it. A webcam capture runs through three Photon nodes in series, invert, contrast, pixelize, and comes out the other end as a pop-art frame.
It earns its place because nothing intelligent happens. A workflow engine is a general dependency graph, and treating every step as an AI call is how you pay model prices to resize an image.
10. Give a model a tool
In AI Calculator, a word problem goes into a Llama 3.3 70B node with a calculator wired to it as a tool. The model decides when to call it, and the node returns the answer and the list of tool calls it made. That second output is the one that matters when something goes wrong: it shows the working, not the conclusion.
The calculator is an ordinary node from the same catalog as every other node. The model can call it because nodes already carry descriptions and schemas.
From example to automation
All ten templates use a manual trigger. You press run. That suits a starting point, and it is why none of them is an automation yet. Three things turn one into the other.
- A trigger. A schedule for anything periodic, an HTTP request for anything another system starts, an inbound email for anything a person starts by writing to an address.
- A destination. A demo ends at an output node you look at. An automation ends somewhere durable: a database row, a file in storage, a message in a channel, a reply.
- A decision about failure. What should happen when the model returns nonsense or the page fails to load? Silent failure in a scheduled workflow is worse than no workflow, because you stop checking.
The first two take a few minutes. The third decides whether the automation survives a month of real inputs.
Simple workflow examples versus business process examples
Searches for workflow examples split into two intents that want different answers. One wants the technical shape, which this page shows. The other wants business process automation examples: approval chains, employee onboarding, purchase orders, workflows that run on human sign-offs with software keeping score.
Dafthunk serves the first. Its steps are computation, not human approval gates. If you need multi-stage approval routing with delegation rules and an audit trail of who signed what, a dedicated BPM tool will serve you better. The overlap is partial: the classification, extraction and notification steps inside a business process are what these examples do well.
What these examples do not cover
An honest list should say where it is thin. Two of the highest-demand categories on comparable platforms are missing here. Dafthunk has no AI video generation, and no Google Workspace surface beyond Calendar and Gmail, so the Sheets-and-Drive automations that fill template galleries elsewhere are out of reach for these ten today.
What it covers well: browser automation and scraping, the highest-traffic template category on comparable platforms, where ten browser nodes already ship; research agents; image generation and processing; and chat surfaces including Discord, Telegram and WhatsApp. The ten templates above are a slice of that, small enough to read rather than exhaustive.
Where to go next
Every example links to its own page in the workflow templates gallery, and the nodes reference lists the building blocks each one uses. Core concepts covers how triggers, executions and resources fit together, the material behind the trigger section above. If you are still choosing a platform, the ten best low-code workflow automation tools compares the field by hosting and license instead of feature count.
Dafthunk is open source on GitHub. Every template in this post lives in apps/api/src/templates, so a pull request is the fastest way to add the example you wanted and did not find.