You audit AI opportunity in a business in 4 stages: a discovery call run from a structured question bank, pain-point analysis in Claude, a scored findings report, and a live readout with a 90-day roadmap. The whole engagement takes 8-12 working hours across 2 weeks, your tool stack costs under $30/mo, and the going rate is $1,500-$5,000 per audit.
This is the playbook — the question themes, the prompt sequence, the report structure, and the pricing.
Stage 1: The discovery call question bank
The audit lives or dies on the discovery call. You're not asking "do you want AI?" — you're mapping where the business bleeds money. Build your question bank around 6 themes:
- Revenue mechanics. Where do customers come from, what's an average deal worth, what percentage of leads ever get a response?
- The owner's calendar. What does the owner personally do every week that someone (or something) else could? This is where the expensive problems hide.
- Follow-up gaps. What happens to a lead that calls after 5pm? Who chases unpaid invoices? Silence here is billable opportunity.
- The content bottleneck. Most service businesses say some version of "the biggest bottleneck right now is content creation and execution." Ask what they'd publish if production were free.
- Systems and data. What's documented vs. tribal knowledge? Which software do they pay for and actually use?
- Team appetite. Who already uses ChatGPT on the sly? Who'll resist? Adoption risk belongs in the report.
Record everything. Fathom transcribes and summarizes unlimited calls free; if your clients hate visible meeting bots, Granola captures audio device-side with no bot in the room — free on its Basic plan.
Stage 2: Pain-point analysis with Claude
This is the part people email us about most: "what prompts actually run the audit?" Work the transcript through a fixed sequence in a Claude Project — one client per Project, so context persists:
- Extract and classify. "From this transcript and intake form, list every operational pain point. Classify each by function (sales, fulfillment, admin, content) and by cost type (lost revenue, owner time, error risk)."
- Map to solutions. "For each pain point, name the current AI or automation approach that addresses it, the tools involved, and realistic setup effort for a non-technical operator."
- Quantify. "Estimate the monthly dollar value of fixing each, using the revenue figures from the call. Flag every estimate's confidence level." Check Claude's math against the client's real numbers — you sign the report, not the model.
- Rank. "Rank all opportunities by value-to-effort. Return the top 5 as a 90-day roadmap."
Save the sequence as a reusable Skill in Claude and your second audit runs in half the time. Claude Pro is $20/mo — the entire analysis layer.
The 5 findings that show up in almost every audit
Run enough of these and the same patterns surface. Knowing them in advance makes you faster on the call and sharper in the report:
- After-hours leads die. Calls after 5pm hit voicemail and never get a callback. Fix: missed-call text-back or an AI receptionist — and the monthly value calculates directly from their average deal size.
- The owner is the proposal and content bottleneck. Everything customer-facing waits on one person. Fix: a Claude Project loaded with their voice, past proposals, and offers.
- Nothing is documented. The business runs on tribal knowledge, which means AI has nothing to run on. Fix: a documentation sprint — often the unglamorous first roadmap item that unlocks everything else.
- They pay for software nobody opens. There's almost always $200-$500/mo in unused subscriptions. Finding it can cover your fee on the spot.
- Follow-up stops after one attempt. Quotes go out; nobody chases. Fix: automated follow-up sequences tied to their CRM.
Build these into your question bank and probe for them directly.
Stage 3: The report
10-20 pages, in this order: business snapshot, findings by function, scores, and the ranked roadmap with cost and payback per project. Keep every recommendation concrete — name the tool, the price, and the number it moves.
The structure check: every finding page answers 3 questions — what's broken, what it costs per month, and what fixing it costs. A page without a number on it isn't done.
Build the deliverable in Gamma: paste the outline, apply the client's brand colors, and ship it as a trackable web link so you can see when the owner actually opens it. The Plus plan is about $9/mo billed annually.
If you want the friendlier, scored-template version of this same deliverable — the one buyers search for by name — that's the AI readiness assessment.
Stage 4: The readout
Never email the report cold. Book a 45-minute readout, walk the owner through the top 3 findings, and let them ask "so what would it cost to fix this?" That question is your implementation pipeline — automation builds, voice agents, an AI operating system install — whether you fulfill it or partner it out.
Members of AI Operator Academy run this exact workflow for clients; the $999/yr community is where the question banks and report templates get traded and stress-tested.
What to charge
$1,500 flat for your first few audits, then $2,500-$5,000 once you have sample reports — scale price to the client's revenue, not your hours. Half upfront, balance at readout. The audit is also the front end of the whole AI consulting model, so price it to start relationships, not to retire on.
FAQ
What prompts and skills do I need to run an AI audit?
The 4-step sequence above — extract, map, quantify, rank — is the core. Encode it once as a Claude Skill with your formatting rules and scoring rubric, and the analysis becomes a repeatable asset instead of an improvisation. That repeatability is what lets you quote a flat fee with a straight face.
How long does an AI audit take start to finish?
Two weeks on the calendar, 8-12 hours of actual work: intake review (1 hour), discovery call (1.5), Claude analysis (2-3), report build (3-4), readout (1). Your first will take double that. By the third, the templates do the heavy lifting.
Is it safe to put client data into Claude?
Handle it like a professional: use a paid Claude plan, keep one Project per client, strip obvious identifiers where they aren't needed, and tell the client exactly what tooling you use — put it in your agreement. For clients in regulated spaces, scope the audit to process descriptions rather than raw customer records; you're auditing workflows, not reading their files. If a client requires stricter guarantees, that's an enterprise-tier conversation, and most $500K-$10M businesses don't.
Do I need access to the client's systems to audit them?
No — and it's cleaner if you don't take it. The audit runs on the discovery call, the intake form, and screenshots or exports the client provides. You're auditing workflows and economics, not logging into their CRM. If a roadmap project later requires system access, that's scoped (and priced) in the implementation engagement, with permissions in writing.
What if the client's needs are beyond my skill range?
Expected — and fine. The audit only obligates you to diagnose, not to build. When a roadmap item is out of your range (custom integrations, a complex voice agent), white-label it or refer it for 20-50% of the project. "I sell the audit, someone else builds it" is a real model many operators run deliberately. Plenty of consultants run their whole client workflow this way.