Lesson 3 of 7 · 9 min

AI steps and HTTP calls

Add AI steps and HTTP calls to a workflow.

Read the lesson

Goal: Use AI steps to read and sum up text. Connect outside tools safely. Keep money and safety decisions with people.

You need: Lessons 1 and 2 of this course. Core, Growth or Franchise.

Why this matters

Service businesses run on messy text. Tech notes, customer messages, call transcripts and photos with captions. Somebody in the office reads all of it and types the important parts into the right place. AI steps can do that reading in seconds. That saves hours a week. It also adds a new risk: an AI that guesses wrong about money or safety. This lesson shows where AI steps help and where a person must stay in charge.

The AI steps

Add step has an Emma actions group. Use each step for one clear job.

  • AI Note or Summary turns long text into a short summary. Example: sum up a tech's notes for the office.
  • AI Structured Extraction pulls set fields out of text. Example: find the equipment brand and age in a message.
  • AI Generate Content drafts text. Example: a first draft of a follow-up email.
  • AI Tag Conversation adds a tag to a conversation. Example: tag "billing question" or "frustrated."
  • AI Recommend Service suggests a service from what the customer said.
Fig 1

The AI steps in Add step

Emma actions

AI Note or SummaryLong text to a short summary
AI Structured ExtractionSet fields pulled out of text
AI Generate ContentA first draft of text
AI Tag ConversationA tag on a conversation
AI Recommend ServiceA service from what the customer said
Use each AI step for one clear job. Give each a fixed output format.

Give each AI step a fixed output format. "Write 3 lines: Problem, Work done, Next step." A fixed format is easy to check and easy to read. For roofing inspection notes, use 4 lines: Roof type, Area, Damage and Access.

Fig 2Example numbers

Inspector notes before and after the AI step

Raw inspector notes
  • arch shingle approx 28 sq per insp
  • hail hits N and W slopes, 3 vents cracked
  • dog in yard, side gate
  • ladder ok
Fixed 4-line summary
  • Roof type: architectural shingle
  • Area: 28 squares, as written
  • Damage: hail on north and west slopes, 3 vents cracked
  • Access: dog in yard, use the side gate
A fixed format is easy to check. Numbers are copied as written, never guessed.

Writing good AI instructions

An AI step is only as good as its instructions. Write them like a job card for a new office hire.

  • Say what to read. "Read the deal notes."
  • Say what to write. "Write 3 labelled lines: Problem, Work done, Next step."
  • Set a limit. "Under 80 words."
  • Say what not to do. "Do not guess numbers. Copy numbers only as written."
  • Say what to do when unsure. "If a line is missing, write: Not stated."

The last two lines matter most. Without them, the AI fills gaps with guesses that look like facts. "Not stated" tells the person to check. A guess hides the gap.

Test with messy samples, not clean ones. Real tech notes have typos, short forms and missing parts. If the summary is wrong on 1 of 3 samples, fix the instructions and test again.

The rule: AI reads, people decide

AI steps are good at reading, sorting and summing up text. They are not the right tool for decisions about money or safety.

So an AI step can sum up notes, tag urgent messages, draft replies or suggest a service. A person sets material quantities and prices, decides on a safety emergency, gives any discount and writes the final quote.

Build the AI step to prepare the decision. Then create a task for the person who makes it.

Fig 3

AI reads, people decide

AI step can
  • Sum up inspection notes
  • Tag a message as urgent
  • Draft a reply
  • Suggest a service
A person decides
  • Material quantities and prices
  • Whether it is a safety emergency
  • Whether to give a discount
  • The final quote
Build AI steps to prepare decisions, not make them.

Connecting outside tools

There are four ways to connect a workflow to other systems.

  1. Integrations in Settings → Connections & developer tools. Ready-made links include Zoho CRM, HubSpot, Pipedrive, Zapier, Make.com, Calendly and Acuity Scheduling. Jobber, ServiceM8, JobNimbus and ServiceTitan are in beta.
  2. Zapier or Make.com. Use these when a ready-made link does not exist. They connect thousands of apps.
  3. HTTP Request step. It sends data from a workflow to a web address that the other system gives you.
  4. Run connected MCP tool step. It runs a tool from an AI tool connection set up in AI tool connections.

Start with the simplest option that works. A ready-made integration needs no upkeep from you. HTTP Request needs the other system's address and format. If either changes, your step breaks.

Fig 4

Choosing a connection

  1. A client tool needs workflow data
  2. Is there a ready-made link in Integrations?YesUse it. It needs no upkeep from youNoTry the next option
  3. Do Zapier or Make.com connect the tool?YesUse Zapier or Make.comNoTry the next option
  4. Does the tool give a web address and format?YesUse an HTTP Request stepNoUse Run connected MCP tool, if an AI tool connection exists
  5. Send the least data you canA job number and a summary, not the full record
Start with the simplest option that works. Each step down needs more upkeep.

Send the least data you can. Send a job number and a summary, not the full customer record. Every extra field is something that can leak.

The app has no QuickBooks integration today. Use Zapier or Make.com if a client's accounting tool needs job data.

By business size

An owner-operator writes short notes and reads them themselves. AI summaries add little. Start with tags instead. A fleet of 8 to 20 techs gains the most. Many notes, read by an office that did not do the work. A multi-location client gains from one format across all locations. Every manager reads the same 3 lines.

Keep customer data safe

AI steps and outside tools both move customer data. Treat that data with care. Only connect tools the owner has agreed to. Write the list in your Operator notes. Use one connection per purpose, so you can switch one off without breaking the others. When a client leaves, or stops using a tool, remove its connection the same week. Check Audit logs if you need to see who changed a connection. Never paste a password or key into an AI step's instructions. Keys belong in the system that uses them, nowhere else.

Credits for AI steps

Plan for 1 credit for the run and 1 credit for each AI step, at the AI message rate. Add the cost of any message the run sends. An SMS segment is 2. An email is 0.5. The exact charge for each AI step is best checked in Billing → Usage history after the first week.

Build it in ServiceAgent

This build sums up roofing inspection notes when a deal is won.

  1. Open Workflows → Create → Build from scratch. Name it "Won job: notes summary."
  2. Set Trigger Type to Event Trigger, event Deal Status Changed.
  3. Add step → Condition: deal status Equals Won.
  4. Add step → AI Note or Summary. Point it at the deal's notes.
  5. In the step's instructions, set a fixed format: "Roof type. Area in squares, as written by the inspector. Damage found. Access notes. Under 80 words."
  6. Add the line: "Do not estimate quantities or prices. Copy numbers only as written."
  7. Add step → Create CRM Task. Title "Order materials." Owner the production manager. Put the summary in Description.
  8. On the No path of the condition, add a safe default: Update Contact with a tag such as deal-updated.
  9. Click Test with 3 real-looking sample notes. Compare each summary with the notes by hand.
  10. Click Publish. Check it works: win one test deal and read the task.
Fig 5

A won roofing job, from notes to order

  1. OfficeMarks the deal Won
  2. ServiceAgentDeal Status Changed; Condition: Won
  3. ServiceAgentAI Note or Summary writes 4 lines
  4. ServiceAgentCreate CRM Task: Order materials
  5. Production managerReads the summary, sets quantities and prices
The AI writes the summary. The production manager sets quantities and places the order.

Worked example: Summit Roofing

Summit Roofing has 15 people on Growth. It sends about 40 quotes a month and wins about 10 jobs (Collect lesson 3). The production manager used to read each inspector's notes and retype them. That took about 15 minutes a job. 10 × 15 = 150 minutes, or 2.5 hours a month.

Now the AI step writes a fixed summary. The manager reads it and places the order. That takes about 4 minutes. 10 × 4 = 40 minutes. The saving is 150 − 40 = 110 minutes, about 1.8 hours a month.

Fig 6Example numbers

Summit Roofing time saved

Before: 10 jobs × 15 minutes150 minutes
−After: 10 jobs × 4 minutes40 minutes
Saved each month110 minutes, about 1.8 hours
The manager still reads every summary. Retyping is what goes away.

Credits: each run is 1 credit plus 1 for the AI step. That is 2 credits a job. 10 × 2 = 20 credits a month. Growth includes 6,000 credits, so this uses about 0.3%.

Fig 7Example numbers

What the AI step costs Summit

2Credits a job: 1 run + 1 AI step
20Credits a month for 10 jobs
About 0.3%Share of Growth's 6,000 credits
Real time saved for a tiny credit cost. Check the real charge in Usage history.

The owner first asked for the AI to set quantities and prices too. You said no. A wrong quantity on a roof order costs hundreds of dollars and a lost day. The person keeps that decision.

Common mistakes

  • Mistake: Letting an AI step set prices or quantities. Fix: AI sums up. A person decides money and safety.
  • Mistake: Open prompts such as "Summarize this." Fix: Give a fixed format and a word limit.
  • Mistake: Sending the full customer record through HTTP Request. Fix: Send the least data the other system needs.
  • Mistake: Building HTTP Request when Zapier would work. Fix: Use the simplest option that needs no upkeep.
  • Mistake: Testing with one sample. Fix: Test at least 3 different notes and check each by hand.

Do this now

Write the instructions for one AI step with a fixed format and a "do not decide" line. Build it in your practice account and test it with 3 sample texts. Deliverable: the instructions plus a table of 3 inputs and 3 outputs, marked right or wrong. Time: 30 minutes.

Practice the conversation

  1. Owner Can the AI read the inspector's notes and order the shingles?
  2. Operator It can read the notes and write a clean summary for your manager.
  3. Owner Why not let it order?
  4. Operator A wrong quantity costs you money and a day. Your manager should decide that.
  5. Owner So what do I save?
  6. Operator About 2 hours a month of retyping notes, and fewer typing errors.
  7. Owner What does it cost?
  8. Operator About 2 credits a job. Around 20 credits a month for you.
  9. Owner And it sends my customer info to the supplier?
  10. Operator No. It only creates a task for your manager.

Quick check

  1. A roofing client wants an AI step to read notes and set material quantities and prices. What do you build?
    • An AI step that sets quantities and prices, with a weekly spot check
    • An AI step that sums up notes in a fixed format, and a person sets quantities and prices
    • An `HTTP Request` that sends the full notes and customer record to the supplier
    • No AI step at all
  2. A client wins 60 jobs a month. Each run uses 1 credit plus 1 for an AI step. How many credits a month?
    • 60
    • 90
    • 120
    • 180
  3. A client's accounting tool has no ready-made integration. What should you try first?
    • Zapier or Make.com
    • An `HTTP Request` with the full customer record
    • Move to Franchise for the API
    • Retype the data by hand each week

ServiceAgent UniversityLearn to run a service business on autopilot. Free courses, Blueprints and certification for Service Operators.

Learn

Get certified

Community