Who this article is for

You're a business owner or executive who understands: AI isn't just hype, it's a growth tool. But you don't know where to start. Hiring a consultant is expensive. Implementing "something" is risky. According to MIT, 95% of AI projects fail to create measurable value precisely because they start without diagnostics.

This article is a step-by-step framework I use with clients. You can conduct an initial AI audit of your business independently in 2 weeks.

Framework overview: 4 phases in 14 days

• Days 1–3: Process inventory
• Days 4–7: Automation potential assessment
• Days 8–11: ROI calculation and prioritization
• Days 12–14: Roadmap and quick wins

Phase 1: Process inventory (days 1–3)

Before looking for AI opportunities, you need to understand what's actually happening in the business. According to McKinsey, most companies don't have a complete process map — knowledge lives in people's heads.

Day 1: Map all business functions

List all core business functions in a table:

• Sales: lead processing, qualification, follow-up, CRM
• Marketing: content, email, advertising, analytics
• Support: tickets, chat, phone, FAQ
• Operations: procurement, warehouse, logistics, delivery
• Finance: invoices, payments, reconciliation, reports
• HR: hiring, onboarding, KPIs, training

Day 2: Interview department heads

Conduct a 30-minute interview with each department head. Ask 5 questions:

1. Which tasks consume the most time for your team?
2. Which tasks repeat daily/weekly?
3. Where do errors happen most often and what does each error cost?
4. What data do you use for decision-making? In what format?
5. What would you automate first if you could?

Day 3: Build a process registry

Consolidate results into a single table. For each process, record:

• Name (e.g., "Responding to incoming customer inquiry")
• Department (support)
• Frequency (200 times/day)
• Time per execution (15 min)
• Who performs it (3 operators)
• Data type (structured / unstructured)
• Current tools (CRM, Excel, email)

You should end up with 15–30 processes. This is your "territory map."

Phase 2: Automation potential assessment (days 4–7)

Not every process is suitable for AI. Now you need to filter candidates.

AI suitability matrix

Score each process on 4 criteria (1–5 points):

1. Repetitiveness

• 5 — fully standardized, identical every time
• 3 — variations exist, but the core flow is clear
• 1 — unique every time, requires creativity

2. Data Volume

• 5 — hundreds/thousands of operations per day, digital data
• 3 — dozens of operations, some data is digital
• 1 — single operations, data on paper or in people's heads

3. Decision Rules

• 5 — clear "if X, then Y" rules
• 3 — rules exist but with nuances and exceptions
• 1 — requires intuition, experience, empathy

4. Error Cost

• 5 — errors are expensive (lost client, penalty, defect)
• 3 — moderate consequences
• 1 — errors are easily fixable

Total score = sum of four ratings (4–20)

Interpreting results

• 16–20 points: excellent AI candidate — automation will deliver quick impact
• 11–15 points: good candidate — worth a detailed ROI assessment
• 4–10 points: weak candidate — better to start with other processes

According to AutomationTactics, if the answer to "how do you make this decision?" is "it depends" — the process isn't ready for automation. Standardize first, then AI.

AI solution types by task category

• Classification & routing: sorting tickets, leads, documents → LLM or classical ML
• Content generation: customer responses, reports, descriptions → LLM (GPT, Claude)
• Data extraction: parsing documents, invoices, resumes → OCR + NLP
• Forecasting: demand, churn, pricing → classical ML (XGBoost, Prophet)
• Optimization: routes, schedules, inventory → operations research + ML

Phase 3: ROI calculation and prioritization (days 8–11)

Now we count the money. This is the most important phase — this is where 90% of "wishful thinking" gets filtered out. For a deeper dive into ROI calculations and industry benchmarks, see "Where AI Delivers Maximum ROI".

ROI formula for each process

Monthly savings = Volume × Time × Hourly cost × Automation percentage

Example: processing incoming requests

• Volume: 200 requests/day × 22 working days = 4,400/month
• Time: 15 min = 0.25 hours
• Operator hourly cost: $15
• Realistic automation percentage: 50%
• Savings: 4,400 × 0.25 × $15 × 0.5 = $8,250/month

Cost estimation

Typical AI project costs (per 2026 benchmarks):

• Simple bot / agent: $5K–$50K development + $3K–$13K/month maintenance
• Medium ML project: $50K–$150K + 17–30% annually
• Ongoing costs: multiply base token calculation by ×1.7

Prioritization: Impact × Effort matrix

Plot processes on two axes:

• Impact (Y-axis): monthly savings in dollars
• Effort (X-axis): implementation complexity (data, integrations, approvals)

Priorities:

1. High Impact + Low Effort → do first (quick wins)
2. High Impact + High Effort → plan for the quarter
3. Low Impact + Low Effort → if resources allow
4. Low Impact + High Effort → don't do

According to Auxis, always discount expected savings by 30% — for edge cases, maintenance, and the learning curve.

Phase 4: Roadmap and quick wins (days 12–14)

Day 12: Select 1–2 quick wins

A quick win is a project that:

• Pays for itself in 1–3 months
• Doesn't require complex integrations
• Delivers visible results to the team
• Creates internal AI advocates

Typical quick wins:

• AI FAQ bot for support (50%+ of queries are standard)
• Auto-classification of incoming requests
• AI-generated product descriptions or email campaigns
• Auto-populating documents from CRM

Day 13: Build a 6-month roadmap

Month 1–2: Quick win #1 — MVP and pilot on a limited sample

Month 3–4: Quick win #2 + scaling the first project

Month 5–6: Begin work on "high Impact + high Effort" project

Each stage must have:

• A specific success metric (response time, conversion, operation cost)
• Budget with +30% buffer
• An owner (without top-management sponsorship, success chances are 11%)

Day 14: Verify data readiness

For each selected project, answer:

1. Does the data exist? (in CRM, logs, ticketing system)
2. Is the data accessible? (API, export, or manual preparation needed)
3. Is the data clean? (duplicates, gaps, different formats)
4. Is there enough data? (ML typically needs 1,000+ examples)

If the answer to any question is "no" — it's not a stop, but it's an additional stage (and budget) in the project.

Checklist: audit deliverables

After 2 weeks you should have:

• Map of 15–30 key processes with metrics
• AI suitability rating for each process (4–20 points)
• ROI calculation for top 5 processes
• Impact × Effort matrix
• 1–2 quick wins with a concrete plan
• 6-month roadmap
• Data readiness assessment

Template with example

I've prepared a Google Sheets template with a filled-in customer service audit example — all 4 phases on separate tabs: process registry, AI suitability matrix, ROI calculation, and roadmap.

Open AI Audit Template in Google Sheets →

When to bring in a consultant

This framework covers 80% of the work. But there are situations where external expertise is critical:

• Complex integration — multiple legacy systems without APIs
• Regulated industry — finance, healthcare, government (compliance requirements)
• Scale — 50+ processes, multiple business units
• No technical team — need someone to translate business problems into technical solutions

In these cases, a 30-minute strategy session can save months of trial and error.