AI Automation Scope of Work Template (Copy, Fill, Send)

A copy-paste scope of work for an AI agent project: goal, accounts, rules, tests, monitoring, handover and milestones, plus how to read bids and test tasks.

By Carlos Familia, CRM & AI automation expert. Updated October 15, 2026. 7 min.

The life insurance agency's GoHighLevel account was empty when I opened it. By handover, this California agency ran 46 live workflows, a bilingual phone intake agent and a website chat agent. Several things that broke along the way were questions a one-page scope would have asked up front.

Which channel does the chat agent answer on? It sat on auto-pilot and answered nobody for a day, because no channel had been assigned to it. Which websites use which chat widget? A second site ran a different widget that was not in the routing. Whose prepaid balance sits behind the phone numbers? One voice account ran dry, and all 8 numbers on it went dead.

A scope of work for an AI agent project names the goal with a number, the accounts and their owner, what the agent may and may not do, how it will be tested, how failures reach you, what you get at handover and what is out of scope. Paste the template below, fill the brackets and send the same text to every bidder. Compare how they answer it before you compare prices.

What does the template look like?

AI AGENT PROJECT: SCOPE OF WORK

1. GOAL AND BASELINE
Business: [type, state or country]
Problem in one sentence, with a number: [example: we miss about a third of after-hours calls]
Baseline measured before any build: [response time, unanswered leads, missed calls, period measured]
Success measure after launch: [metric, how it is counted, who counts it]

2. WHAT GETS BUILT
Agents: [phone / website chat / SMS / Instagram / WhatsApp]
Languages: [English, Spanish]
Each agent can: [answer questions, capture requests, book, transfer]
Each agent must never: [quote prices, give medical or legal advice, other limits]
Tools it may use: [calendar, CRM, messaging], with read or write access listed per tool
Every sentence a customer sees or hears is approved by: [name]

3. ACCOUNTS AND OWNERSHIP
Created in the client's name: [voice platform, phone numbers, automation tool, CRM, model account, booking system]
Contractor works as a user. Admin stays with the client.
One balance per business. Low-balance alerts go to: [name, phone, email]

4. CHANNELS AND ROUTING
For each channel: who answers, from which number or page, and the fallback.
Every website and widget in use is listed: [urls]
Every booking link is opened in a private window and checked.

5. RULES AND DISCLOSURE
Caller learns it is an AI: [wording approved by client counsel]
Recording notice: [wording approved by client counsel]
Outbound calls: [consent on file, hours by the lead's time zone, kill switch]
Handoff: [who receives it, how, and with what context]

6. TESTS
At least [24] test questions, run by text and by phone, in every language.
Real-phone check for transfers and texts.
Tests use test contacts and a test calendar, never live customers.
A test report is delivered, with every failure and its fix.

7. MONITORING
Alerts for: no activity, failed calls or messages, low balance, a stopped workflow.
Weekly count: platform records compared with records processed.

8. USAGE COSTS
Paid by the client directly to each platform.
Contractor supplies an estimate per [1,000 conversations or 1,000 minutes] before launch.

9. HANDOVER
Prompt and settings export, tool list, knowledge base, test report, alert list,
one-page runbook. Contractor access removed on [date].

10. OUT OF SCOPE
Anything not listed above, including: [new channels, new languages, ad management, copywriting]

11. SUPPORT WINDOW
[30] days of fixes after launch. Changes beyond this scope are quoted separately.

12. MILESTONES AND ACCEPTANCE
Each milestone is accepted when its deliverable passes its stated check.

Why is each line in there?

Most of these lines trace back to that insurance build.

• Line 2, approved wording. One meeting type said the customer would place the call, when the agent was supposed to. Checking the client roadmap's copy against the live workflow definitions produced 47 fixes. Someone has to read every sentence a customer sees.
• Line 3, balances. The 8 dead numbers above all sat on one prepaid account.
• Line 4, routing and booking links. The chat agent's missing channel and the second site's widget were both routing gaps. Calendars I created through the API also had no slug at first, so the share link opened another company's calendar, and a system "Work Hours" schedule overrode a calendar's own hours. Opening every link in a private window finds both.
• Line 5, rules. The agency's compliance review is why the template has two separate disclosure brackets, one for the AI notice and one for the recording notice. In California, section 632 of the Penal Code makes recording a confidential call without everyone's consent an offense, with a fine of up to $2,500 for each violation. My California builds assume every party must agree, and the exact wording for your state is your counsel's call. The outbound hours and consent rules are in my AI appointment setter guide.
• Line 6, tests. The workflows passed 10 of 10 execution tests, and the Spanish-language phone check came back as a full pass, on a call of 143 seconds. A line that names languages makes sure the second language gets tested at all.

What if your project is small?

Not every project needs twelve sections. For one chat agent on one website, keep lines 1, 3, 6, 8 and 9: the goal with a number, the accounts in your name, the tests, the usage estimate and the handover. Add line 5 the moment the agent records calls, texts people or dials out, because that is where the legal wording lives.

The reverse also holds. If the project touches a phone line, a calendar and a CRM, as the insurance build did, every section earns its place. The lines you leave out are the ones that break.

How do you read the proposals?

Put the replies side by side. Good ones restate your goal with your number, ask who owns each account and what your booking system allows, describe tests that fail on purpose, give a usage estimate and list the handover files.

To compare prices, convert each bid into hours at the bidder's stated rate. A bid that implies a few hours for a multi-channel agent has left something out. Published rate ranges, and who publishes them, are in AI automation agency vs freelancer.

What is the paid test task?

Before the full project, pay for one small, real flow. Ask for:

1. Answers to three questions pulled from your real FAQ.
2. One booking request written to a test calendar, with the tool-call log.
3. One failure made on purpose, and the alert that fires.
4. A one-page test report.

Time-box it to a day or two. How the developer reports the failure in item 3 tells you more than the portfolio. More on screening is in my AI automation freelancer red flags.

What does the milestone plan look like?

Release payment per milestone, on acceptance. If the contractor refuses to be paid that way, ask why. The AI voice agent developer questions, the AI chatbot developer scorecard and the guide to AI agent maintenance fill in the rest, and the insurance side of this build is in insurance agent CRM. To fill in the brackets together, book a call.

About the author

Carlos Familia, CRM & AI automation expert. GoHighLevel, n8n, AI agents, servers. Full-stack developer.

I build CRM systems in GoHighLevel, n8n automations on my own servers and AI voice and chat agents for service businesses. Full-stack developer: I write the workflow, the API integration, the database and the server it runs on, and I test every build with real journeys before handoff.