AI Agent Blueprint
Clarity and focus; stones believed to store and amplify energy.
A map of what an AI agent can run in your business, and in what order.
Start with the work, not the tool
An AI agent is software that takes an input, follows instructions, uses tools you give it and produces an output, often without a person in the loop. The question is never "what can AI do". It is "which of my recurring tasks are clear enough to hand over, and in what order is it safe to do that".
This blueprint takes you from a list of your own tasks to a rollout plan, one agent at a time.
Step 1: list and score your tasks
Write down every task that happens at least weekly. Score each from 1 to 5 on the four tests. Hand over the highest totals first.
- Frequency: how often it happens. Daily scores 5.
- Clear rules: could you write the steps on one page? Yes scores 5.
- Low cost of a mistake: if it goes wrong, is it easy to catch and fix? Yes scores 5.
- Data in reach: is everything the task needs already in a system the agent can read? Yes scores 5.
| Task | Frequency | Clear rules | Low risk | Data in reach | Total |
|---|---|---|---|---|---|
Step 2: the agent map
These are the jobs agents handle well in a marketing and sales business. Each has a human checkpoint; keep it until the agent has earned your trust.
| Agent | Input | What it does | Output | Human checkpoint |
|---|---|---|---|---|
| Lead qualifier | Form or application answers | Scores fit against your rules, looks up the company | Score, short summary, route | Review borderline scores |
| Inbox and DM triage | New messages | Sorts by intent: sales, support, spam, partner | Labels and draft replies | Approve sales replies |
| Follow-up writer | Call notes or deal stage | Drafts the next message in your voice | Draft email or text | Read before sending |
| Call prep | Booked call, contact record | Gathers what is known about the person and company | One-page brief before the call | None needed |
| Content repurposer | Long video, podcast or article | Pulls key points into posts, emails, short scripts | Draft pieces for each channel | Edit and approve |
| Support answerer | Customer question | Answers from your help docs, escalates anything else | Reply or handoff | Audit a sample weekly |
| Reporting | Ad, email and sales data | Pulls numbers, compares to last week, flags changes | Weekly summary | Read and act |
| Onboarding guide | New customer | Sends steps, checks completion, nudges when stuck | Progress record | Step in when flagged |
Step 3: roll out in three phases
Every agent goes through the same three phases. Do not skip ahead, even when it looks ready.
Draft only
The agent reads and writes drafts. A person does every action. Run it until the drafts need almost no edits.
Act with approval
The agent prepares the action and a person approves it with one click. Track how often you change or reject what it proposes.
Act alone, within limits
The agent acts on its own inside written limits, and hands anything outside them to a person. Review a sample of its work on a fixed schedule.
Suggested order for most businesses: reporting and call prep first (they only read), then triage and follow-up drafts, then lead qualification, then customer-facing support and onboarding.
Step 4: write the agent spec
One page per agent. This is what you give the builder, or paste into the instructions yourself.
Role: you are the [name] for [business]. Your job is to [single job]… Inputs you receive: […] Tools you may use: [read the CRM, search the help docs, draft an email…] Rules you always follow: [tone, what never to promise, required fields…] Hand to a person when: [refund requests, legal questions, anger, anything not covered…] Output format: […] A good result looks like: [one real example]…Guardrails checklist
- The agent can only reach the systems it needs for its one job.
- It cannot send money, issue refunds or change prices.
- Anything customer-facing has a clear handoff to a person.
- Every action it takes is logged where you can read it.
- You know what each run costs and have a monthly limit set.
- Customer data is handled under the same privacy rules as your team.
- Someone owns the agent and reviews its work on a schedule.
Your first ninety days
| Window | Agent | Phase | Done when |
|---|---|---|---|
| Days 1 to 30 | Draft only | ||
| Days 31 to 60 | Act with approval | ||
| Days 61 to 90 | Act alone, within limits |
Want this built for you instead of by you?
Return to the vault