AI business workflow automation in 2026 means connecting language models to your existing systems so they can complete multi-step tasks without human intervention. You’re routing customer inquiries through Claude Sonnet 4-6, triggering data pulls from your CRM, generating reports with GPT-4o, and sending summaries to Slack. The core shift from 2024 is that models now handle context windows up to 2 million tokens (Google Gemini 3.5 Pro) and route tasks across specialized models automatically. You build workflows by defining triggers, selecting which model handles each step based on cost and capability, then connecting outputs to your tools via API or platforms like Make, Zapier, or n8n. Most businesses start with three workflows: customer support triage, document processing, and report generation. This article walks through exactly how to build each one.

Why This Matters for Your Business Right Now

The pricing structure changed. Claude Haiku 4-5 costs $0.25 per million input tokens. GPT-4o runs $2.50 per million. Gemini 2.5 Flash sits at $0.075 per million. If you’re still routing every task through the most expensive model, you’re burning budget on work that cheaper models handle fine.

Your competitors automate customer support triage, invoice processing, and weekly reporting. They respond in 90 seconds instead of 90 minutes. Cursor Bugbot now completes code reviews in 90 seconds and finds 10% more bugs at 22% lower cost than manual review. That efficiency gap compounds.

The second reason is context retention. Models with 2 million token windows (Gemini 3.5 Pro launching this month) can ingest your entire product catalog, customer history, and support documentation in a single prompt. You stop losing context between steps. A customer asks about a refund, the model sees their purchase history, warranty status, and previous tickets without you building complex retrieval systems.

Third, multi-model routing became standard. Perplexity Computer routes tasks across 20+ models depending on what the job needs. You don’t pick one model and force it to do everything. You route financial analysis to o4-mini for accuracy, creative briefs to Claude Opus 4-8, and quick summaries to Mistral Large 2. The platform decides based on your rules.

How to Build Your First Three Workflows

Start with customer support triage. Most businesses get 50-200 support emails daily. Half are simple questions your documentation already answers.

Step 1: Connect your email to an automation platform. Use Make, Zapier, or n8n. All three offer native Gmail and Outlook integrations. Create a new scenario that triggers when an email arrives in your support inbox.

Step 2: Route the email content to Claude Sonnet 4-6. Use the Anthropic API integration. Your prompt: “Read this customer email. Classify it as: refund request, technical issue, billing question, or general inquiry. If it’s a general inquiry you can answer from this knowledge base [paste your FAQ], draft a response. If not, flag it for human review and summarize the issue in one sentence.”

Set temperature to 0.3. You want consistent classification, not creative interpretation.

Step 3: Add conditional logic. If Claude returns a drafted response, send it to a human for approval in Slack. Include an approve/edit/reject button. If approved, the automation sends the email. If Claude flags it for review, create a ticket in your support system with the one-sentence summary.

Cost: roughly $0.15 per 100 emails processed. Claude Sonnet 4-6 runs $3 per million input tokens, $15 per million output. Average email is 500 tokens input, 200 tokens output.

Step 4: Monitor accuracy for two weeks. Track how many auto-responses get edited before sending. If it’s above 30%, your knowledge base needs work or your classification prompt is too broad. Tighten the categories or add more examples to your prompt.

The second workflow is document processing. Invoices, contracts, RFPs, anything that arrives as a PDF and needs data extracted.

Step 1: Set up a watched folder. Dropbox, Google Drive, or an S3 bucket. When a file appears, trigger the workflow.

Step 2: Send the document to GPT-4o with vision. Use the API to pass the PDF. Your prompt: “Extract: vendor name, invoice number, total amount, due date, line items with quantities and prices. Return as JSON.”

Temperature 0.2. You need structured output, not interpretation.

Step 3: Validate the extraction. Check that required fields aren’t empty and the total matches the sum of line items. If validation fails, flag for human review. If it passes, write the JSON to your accounting system via API or create a row in your ERP.

Step 4: Archive the original document with the extracted data as metadata. Tag it with vendor name and invoice number so your team can search later.

Cost: $0.05-0.10 per document depending on page count. GPT-4o vision handles multi-page PDFs well but costs more per token than text-only models.

The third workflow is weekly reporting. You want a summary of sales, support tickets, and product usage delivered to your leadership team every Monday at 8am.

Step 1: Pull data from your sources. Use API calls to your CRM (HubSpot, Salesforce), support system (Zendesk, Intercom), and analytics platform (Mixpanel, Amplitude). Most platforms offer webhook or scheduled export options.

Step 2: Combine the data and send to Gemini 2.5 Pro. Your prompt: “Here’s our data from last week: [paste JSON]. Write an executive summary covering: top 3 wins, top 3 concerns, recommended actions for this week. Keep it under 300 words. Use bullet points.”

Temperature 0.7. You want some synthesis and prioritization, not just regurgitation.

Step 3: Format the output. Take the model’s response and drop it into a Google Doc or Notion page. Add charts using your automation platform’s native integrations or a tool like QuickChart.

Step 4: Distribute. Email the link to your leadership team or post it in a dedicated Slack channel.

Cost: $0.20-0.40 per report depending on data volume. Gemini 2.5 Pro costs $1.25 per million input tokens. A typical report uses 50k-100k tokens of input data.

Choosing the Right Model for Each Step

Don’t default to the most expensive model. Match capability to task.

Use Claude Haiku 4-5 ($0.25 per million input tokens) for: classification, simple extraction, yes/no decisions, formatting text.

Use Claude Sonnet 4-6 ($3 per million input tokens) for: customer-facing responses, summarization, anything that needs nuance but not deep reasoning.

Use GPT-4o ($2.50 per million input tokens) for: document processing with vision, structured data extraction, tasks where you need reliable JSON output.

Use Gemini 2.5 Pro ($1.25 per million input tokens) for: long-context tasks where you’re passing in 100k+ tokens, analysis that requires connecting information across many documents.

Use o4-mini for: financial calculations, anything where a math error costs you money, compliance checks.

Use Claude Opus 4-8 ($15 per million input tokens) for: complex reasoning, strategic planning, anything where you’d normally spend 2 hours thinking through options.

The rule: start with the cheapest model that can do the job. Test it on 50 examples. If accuracy is below 90%, move up one tier. Don’t skip straight to the most expensive model because you’re worried about quality.

Common Mistakes and How to Avoid Them

Mistake 1: Not setting token limits. Models will keep generating until they hit their maximum context window. If you don’t set a max_tokens parameter, a simple summary request can cost you $5 because the model wrote 20,000 tokens.

Fix: Set max_tokens to 2x what you actually need. If you want a 200-word summary, set max_tokens to 500. You’ll never hit it, but you won’t get runaway generation.

Mistake 2: Using the same prompt for every model. Claude responds better to conversational prompts. GPT-4o wants structured instructions. Gemini handles long context better when you explicitly tell it to reference specific sections.

Fix: Test your prompt with each model you plan to use. Adjust the structure based on what works. This takes 30 minutes upfront and saves you hours of debugging later.

Mistake 3: Not logging inputs and outputs. When a workflow fails, you need to see exactly what the model received and what it returned.

Fix: Add a logging step to every workflow. Write the input prompt, model response, and timestamp to a Google Sheet or database. When something breaks, you can trace it.

Mistake 4: Automating before you standardize. If your team handles support tickets five different ways, automating it will just scale the chaos.

Fix: Document your current process. Get your team to agree on one way to do it. Then automate that one way. You can optimize later.

Mistake 5: Not building human review into the first version. You don’t know what edge cases exist until you run the workflow in production.

Fix: Every automated workflow should have a human checkpoint for the first 100 runs. After that, you’ll see patterns in what fails and can either fix the prompt or add validation rules.

Mistake 6: Ignoring rate limits. Most AI APIs limit you to 10,000 requests per minute. If you’re processing 500 documents at once, you’ll hit the limit and your workflow will fail.

Fix: Add rate limiting to your automation. Process documents in batches of 50 with a 10-second delay between batches. It’s slower but it won’t break.

Free download: The AI Operating Layer We put together a practical guide covering this and more. Download it here.

What to Build Next

Once your first three workflows run reliably, add complexity. Connect multiple models in sequence. Use Claude Sonnet 4-6 to draft a customer response, then pass it to GPT-4o to check for policy compliance, then send it for human approval.

Build feedback loops. When a human edits an AI-generated response, log the original and the edited version. Every month, review the edits and update your prompts to incorporate what your team consistently changes.

Add conditional routing based on confidence scores. Most APIs return a confidence level with their output. If confidence is below 0.8, route to human review. If it’s above 0.8, auto-approve.

Experiment with local models for sensitive data. If you’re processing financial documents or customer PII, running Mistral Large 2 locally keeps data in your infrastructure. It’s slower and requires more setup, but it eliminates API calls for regulated data.

Test multi-model routing. Perplexity Computer routes tasks across 20+ models. You can build a simpler version: send the same prompt to three models, compare outputs, and use the one with the highest confidence or the most detailed response.

The goal isn’t to automate everything. It’s to automate the repetitive work that doesn’t need human judgment so your team can focus on the decisions that do.

For a structured walkthrough of building this into your operations, book a 60-min Omni Audit — https://calendly.com/sam-mckay/discovery-call?utm_source=edna-landing&utm_medium=blog&utm_campaign=product-keywords