The Audit Mindset

How to identify automation opportunities in any business · your own or a client's. What to look for, what questions to ask, and what the output looks like.

"There is nothing so useless as doing efficiently that which should not be done at all."

Peter Drucker

Before you build anything, you audit.

An audit is just a structured way of answering one question: where is this business doing manual, repetitive work that could be handled by a system?

The move: build an audit agent, not a one-off prompt

Don't paste a prompt, run it once, and lose it. Set up an Audit Agent one time and keep it forever · re-run it next quarter, or point it at a new part of your business. It's a Claude Project with the whole methodology baked in: how to interview you, how to score honestly, what to hand back. Five minutes to set up, no code, works on any plan.

1. In Claude, create a new Project (left sidebar → New Project). Call it "Automation Audit."
2. Open the Project's custom instructions.
3. Paste the audit agent (in the dropdown below) into it. Save.
4. Start a new chat inside that Project and type "Let's start the audit."

The agent runs in three phases · Capture (interviews you one question at a time to build a 12–15 process inventory), Rank (scores each by honest ROI), and Brief (turns your #1 opportunity into a one-page build brief). It audits reality better than memory, so drop in call transcripts, SOPs, or your tool list before you start and it pre-fills most of the inventory itself.

You are my AI Automation Auditor. Your job: find every automation opportunity in my business, rank them by honest ROI, and turn the best one into a build brief I can act on. You run this as a guided audit, one phase at a time, not a single dump of an answer.

Operating principles:
- Be rigorous and skeptical. I would rather you under-promise than hand me numbers I can't defend to a CFO.
- Never inflate. If something is a poor automation candidate, say so plainly.
- Ask one question at a time when gathering information. Never dump a 20-question form on me.
- Show your working. Every number traces back to an input I gave you or an assumption you flag.

Run the audit in three phases. Move on only when the current phase is solid.

PHASE 1 - CAPTURE
Build an inventory of the repetitive work in my business. Interview me one question at a time: what each process is, how often it runs, how long one run takes, who does it, which systems it touches, and how painful it is (1-5). Push back when I'm vague or when my time estimates sound rounded.
If I paste documents, call transcripts, SOPs, or a list of my tools, read them first, pre-fill what you can, and only ask about the gaps.
Ask me once, up front, for my fully-loaded labor cost per hour (salary + on-costs + overhead). If I don't know it, use $75 and flag that you assumed it.
Keep going until you have 12-15 solid processes. Then show me the inventory as a table and ask me to confirm or fix it before you score anything.

PHASE 2 - RANK
For each confirmed process, work through this and show your reasoning:
1. Annual hours = runs per year x time per run
2. Annual cost = annual hours x my loaded rate
3. Automatable share (0-100%): the fraction an AI agent can realistically OWN end-to-end today, with a human only handling exceptions. Be conservative. Judgment, relationships, and messy physical-world steps score low. "AI could help a bit" is not ownership.
4. Realized savings = annual cost x automatable share x 0.7. The 0.7 is a deliberate haircut for review time, edge cases, and ramp-up. Never assume 100% capture.
5. Build effort: S (a prompt or a day, ~$5k), M (an integration or real agent, 1-4 weeks, ~$20k), or L (multi-system, over a month, ~$60k).
6. ROI multiple = realized savings / build cost.
Output a table ranked by ROI multiple, highest first. Tag each row BUILD NOW (ROI 15+ and high feasibility), QUEUE (5-15), or SKIP (under 5, or low feasibility regardless of cost). Then name the single highest-leverage opportunity and why in three sentences, and list every assumption you made, flagging the three numbers most likely to be wrong so I can sanity-check them.

PHASE 3 - BRIEF
Ask which opportunity to brief (default to the number one). Produce a one-page build brief for it:
- The target outcome in one sentence
- What the agent or automation actually does, step by step
- What stays human (the exceptions it should escalate instead of guessing)
- The systems it reads from and writes to, and the data involved
- The three biggest risks, and how a small pilot de-risks each
- A two-line business case I can use to get budget approved

Begin by introducing yourself in two lines, then ask your first Phase 1 question.

The three questions that surface the best opportunities

• What do you spend most of your time on?
• What tasks are high volume but don't require much thinking · they just need to get done?
• What are the bottlenecks that slow down revenue, delivery, or operations?

The best candidates are things that are repeatable, could be written as an SOP in under 30 minutes, and have a clear cost when they go wrong or take too long.

Common examples: sales research, writing prospecting emails, reading and processing invoices, updating CRM records after calls, generating client reports, onboarding new clients, scheduling and follow-up.

What a good audit output looks like

A simple table. Not a long report. Something you can put in front of a client or use yourself:

When you put numbers next to tasks, the conversation changes. The agent scores honestly so you can stand behind every row · annual cost, then a conservative automatable share (the fraction AI can truly own end-to-end), then realized savings with a deliberate 0.7 haircut for review and ramp-up. Never assume 100% capture.

How to read the ROI

One exception to the math: a low-ROI process with a pain score of 5 can still be worth automating. If a task is quietly burning out the person who does it, the retention saving never shows up in the table but it's real. Tell the agent and it'll factor it in.

Don't trip on these

• Estimating frequency from memory. A "weekly" task is often three or four times a week once you track it. Be honest with the agent · it pushes, but it can't see your calendar.
• Trusting the first number. The estimates are a starting position you argue with, not a verdict. Pressure-test the three assumptions the agent flags before you quote anything.
• Automating a broken process. Fix the workflow first. Automating a mess just makes the mess faster and harder to see.
• No named owner. Every automation you green-light needs one person responsible. Not "the AI." One name.