The assumption that AI implementation requires data scientists is outdated. Data scientists build machine learning models from scratch. Most business AI today uses pre-built models from OpenAI, Anthropic, or Google, accessed via API. That is software engineering work, not data science.
What you need to build most practical AI workflows: a developer with API integration experience and clear thinking about what you are trying to accomplish.
The components of a typical business AI workflow
An AI workflow has three layers:
Data layer. Where does the information the AI needs to process come from? This could be a database, a spreadsheet, a CRM, incoming emails, a web scraper, or user input through an interface.
AI layer. Which AI model processes the data? What prompt or instruction tells the model what to do with it? What validation checks the output before it is used?
Action layer. What happens with the AI's output? Does it get written to a database? Sent as an email? Posted to Slack? Used to update a record in your CRM? Presented to a human for review?
Building an AI workflow means connecting these three layers. For simple workflows, this can be done with no-code tools. For complex workflows with business logic, API integration, and error handling, it requires a developer.
Building AI workflows without code: the no-code options
Zapier AI: Zapier added AI actions that let you include AI processing steps in your Zaps. If you receive a customer inquiry via email, you can pass it to an AI model that classifies the type, then route it to the right folder or notify the right person.
Make (formerly Integromat): More powerful than Zapier for complex logic. Supports HTTP requests to AI APIs, which gives you more flexibility than pre-built AI modules. Good for moderate complexity without writing code.
n8n: Open-source and self-hostable. More technical than Zapier or Make, but gives full control over the workflow and data. Good for teams with a technical person who is not a full developer.
Voiceflow and Botpress: No-code tools specifically for building conversational AI applications. If your use case is a chatbot, FAQ assistant, or customer support bot, these are the fastest path.
For workflows that involve standard APIs and do not require custom business logic, no-code tools can get you 80% of the way there in days rather than weeks.
When you need a developer
The cases where no-code tools reach their limits:
Complex business logic. If the AI's output needs to trigger different actions based on specific conditions in your business data, no-code tools become unwieldy. A developer can write this logic cleanly.
Database integrations. If your AI workflow needs to read from and write to a production database, a developer handles this safely. No-code tools are generally not suitable for production database operations.
Error handling and reliability. Production AI workflows need to handle API failures, unexpected outputs, and edge cases gracefully. This requires code.
Security and data privacy. If the workflow processes sensitive customer data, you need control over where that data goes and how it is handled. A developer builds this correctly.
Scale. No-code tools are expensive at high volume and have performance limits. Custom code handles scale better.
Building your first AI workflow: a step-by-step approach
Step 1: Pick the smallest repeatable workflow. What process do you do multiple times per week that involves reading text and producing text? A candidate: processing incoming job applications. Incoming application emails are summarised, scored against a rubric, and routed based on the score.
Step 2: Define the input and output precisely. Input: email with an attached CV. Output: a structured summary with a score from 1-10 on four criteria, and a routing decision (forward to hiring manager / file in not suitable).
Step 3: Write the prompt. The prompt is the instruction you give the AI. For the application screening workflow: "Review this job application for a [role description] role. Score the candidate on these four criteria [criteria with descriptions]. Assign a score from 1-10 for each. Based on the average score, recommend either 'forward to hiring manager' or 'file as not suitable.' Format the output as JSON."
Step 4: Test the prompt on real examples. Run the prompt on ten real past applications where you know what the right answer is. Check the output against your expectations. Adjust the criteria and prompt until accuracy is acceptable.
Step 5: Build the automation. Connect the email trigger to the AI call to the output action using your chosen tool. Test end-to-end with live data.
Step 6: Add a human review step. Before this workflow runs fully automatically, have a human check every output for two weeks. Log any errors. Fix them.
Getting help building AI workflows
Our AI automation team builds custom AI workflows for businesses. We typically start with a discovery conversation about your highest-value use case, build an initial workflow, and then support you in scaling it.
Use the AI feasibility checker to see which of your processes are good AI automation candidates. Then get in touch to discuss a build.