# The 4Ds of Delegating to AI: Owner's Cheat Sheet

Bruno Ashford, Agentic AI for Business Owners, Module 1

The 4Ds come from Anthropic's AI Fluency framework (Delegation, Description, Discernment, Diligence). This sheet translates them into the language of running a business. Print it, keep it next to your keyboard, and run through it before every new task you hand to an agent.

---

## 1. Delegation: should this go to AI at all?

Decide what the agent does, what you do, and what stays human no matter what.

Ask yourself:

- Is this task repetitive, rules-based and digital? (Good sign.)
- If it goes wrong, who notices, and how much does it cost? (Sets the autonomy level.)
- Does it need my personal relationship, taste or final judgment? (Keep that part.)
- What level of autonomy fits: Suggest, Act with approval, or Act alone?

Split the task, do not just accept or reject it:

| Part of the task | Who does it |
| --- | --- |
| Gather the data, draft, format, check against rules | Agent |
| Decide, approve, sign, handle the relationship | You |

Never delegate: hiring and firing decisions, final pricing on large deals, legal or medical judgment, apologising to an upset client, anything you would be embarrassed to explain was done by a machine.

---

## 2. Description: brief it like a new hire

An agent is only as good as the brief. Use this five-part brief every time.

1. Goal: what finished looks like, in one sentence.
2. Context: who it is for, why it matters, where the files are.
3. Steps or rules: how you would do it, your standards, what to avoid.
4. Output: format, length, file name, where to save it.
5. Check: what the agent should verify and report before it finishes.

Copy and fill in:

```
Goal: [what done looks like]
Context: [who it is for, why, which files or folder]
Rules: [my standards, tone, things to never do]
Output: [format, length, file name, where to save]
Before you finish: list what you read, what you changed, and anything you were unsure about.
```

Weak brief: "Write a follow-up email for the Johnson quote."

Strong brief: "Goal: a follow-up email for the Johnson roof quote sent 5 days ago. Context: the quote is in /quotes/johnson.pdf; they asked about financing on the call. Rules: friendly, under 120 words, no discount, mention our 10-year workmanship warranty. Output: plain text email with a subject line. Before you finish: tell me which details you took from the quote."

---

## 3. Discernment: judge the output like a manager

Read what comes back with a manager's eye, not a fan's eye.

Check:

- Facts: are the numbers, names and dates right? Spot check at least three.
- Sources: did it use the files you pointed to, or did it fill gaps with guesses?
- Fit: does it sound like your business and follow your rules?
- Process: did it do the steps you asked for, in the order you asked?
- Gaps: what did it say it was unsure about?

Useful prompts:

```
Show me where each number in this report came from.
What assumptions did you make that I did not give you?
What would a sceptical client object to in this draft?
```

Red flags: very confident tone with no sources, round numbers that do not match your records, anything that "sounds right" but you cannot trace.

---

## 4. Diligence: own the result

You are accountable for what goes out under your name, whoever drafted it.

- Review before anything reaches a client, a bank or a regulator.
- Keep sensitive data (client personal data, passwords, bank details) out until your security rules are in place (Module 11).
- Be transparent where it matters: if a client would care that AI drafted something, tell them.
- Log what worked. When a brief produces great output twice, save it; it becomes a skill in Module 5.
- Review the autonomy level every month. Earn trust step by step, like with a new employee.

---

## The one-minute pre-flight check

Before you hit Enter on a new task:

- [ ] Delegation: I know which part the agent does and which part I keep.
- [ ] Autonomy: I chose Suggest, Act with approval or Act alone on purpose.
- [ ] Description: my brief has goal, context, rules, output and check.
- [ ] Discernment: I know how I will verify the result.
- [ ] Diligence: no sensitive data is in scope, and I will review before anything goes out.

---

## The autonomy ladder at a glance

| Level | What the agent does | Use it when | Example |
| --- | --- | --- | --- |
| Suggest | Drafts or recommends, you do the action | Mistakes reach clients or money | Draft a reply to an unhappy client |
| Act with approval | Does the work, asks before important changes | Mistakes are internal and reviewable | Build the weekly sales report |
| Act alone | Does the work end to end, you review later | Mistakes are cheap and easy to undo | Rename and file last month's receipts |

Move a task up one level only after it has worked well several times in a row at the level below.