How I'd run an AI audit for small business operations

When a business owner asks "should we use AI?", the answer is almost never yes or no. It's "let's find out." That's what an operations audit is for: a structured way to look at how a business runs, identify where time and money are being wasted, and figure out which problems (if any) are worth solving with AI.

The specifics change for every business, but the process stays the same. Here's exactly how I'd run it.

Step 1: Map the workflows

Before touching any data or discussing any technology, I need to understand how the business actually works day to day. Not how the org chart says it works. How it actually works.

This means spending a half day with the team, usually the owner plus 2-3 people who do the hands-on operational work. Walk through a typical week:

• What are the first tasks on Monday morning?
• Where does data come from? Where does it go?
• Which tasks feel repetitive or tedious?
• Where do things get stuck or slow down?
• What breaks most often?

The key is listening for the bottlenecks that people have stopped noticing because they've worked around them for so long. The classic example: "Oh, Sarah spends every Thursday afternoon manually updating the stock spreadsheet from three different supplier portals." That's 4 hours a week that nobody questions because Sarah's always done it.

Map every workflow onto a simple diagram. Nothing fancy: boxes and arrows showing inputs, processes, and outputs. The goal is a clear picture of where time goes.

Step 2: Quantify the pain

Gut feelings aren't enough. "It takes ages" needs to become "it takes 12 hours per week." For every bottleneck identified, measure three things:

Time cost: How many person-hours per week does this task consume? Ask people to track it for a week if they don't already know. They're usually surprised.

Error rate: How often does this process produce mistakes? Mistyped data, missed follow-ups, incorrect calculations. Even a 2% error rate can be expensive if the errors are costly to fix.

Opportunity cost: What would these people be doing if they weren't stuck on this task? This is the number that usually gets the business owner's attention. If your best salesperson spends 30% of their time on admin, that's not an admin problem. It's a revenue problem.

Compile all of this into a simple table: task, time per week, error rate, person responsible, estimated annual cost. The annual cost is usually eye-opening. Four hours a week at £20/hour doesn't sound like much until you see it written as £4,160 a year.

Step 3: Classify each bottleneck

This is where the audit diverges from a general operations review into an AI-specific assessment. Each bottleneck goes into one of four categories.

Automate with simple tools. If the task follows clear rules and involves structured data, it doesn't need AI. A Zapier workflow, a database query, or a simple script will do. In most businesses, about 40% of bottlenecks fall here. Fixing these first gives the business quick wins and builds confidence.

Automate with AI. If the task involves pattern recognition, unstructured data (PDFs, emails, images), classification, or prediction, AI is the right tool. Maybe 20-30% of bottlenecks land here.

Assist with AI. If the task requires human judgement but has a time-consuming preparation phase, AI can reduce the prep work. Drafting reports, summarising documents, pre-screening applications. The human still makes the decision, but faster.

Leave alone. Some tasks aren't worth automating at all. If it takes one person 30 minutes a week and it's not error-prone, the cost of building and maintaining an automation tool outweighs the benefit.

Step 4: Check the data

For every task classified as "automate with AI" or "assist with AI", the data needs to be assessed. The most common reason AI projects fail is that the data isn't ready.

Check four things:

Does the data exist? Sometimes the information needed for an AI solution is trapped in people's heads or scattered across email threads. You can't train a model on institutional knowledge.

Is it digital and accessible? Data in paper files, locked PDFs, or legacy systems with no API is expensive to extract. Not impossible, but it adds weeks to the project.

Is there enough of it? A basic classifier usually needs a few hundred labelled examples. A predictive model might need a few thousand data points. If the business has been operating for years, there's usually enough. If it started last month, there might not be.

Is it consistent? Inconsistent labelling, duplicate records, and format changes over time all create noise. Looking at a sample of the data early spots these issues before they become expensive.

If the data isn't ready, the right recommendation is a data infrastructure project before any AI work. Less exciting than building a model, but it's the foundation everything else depends on.

Step 5: Build the roadmap

The audit output is a document (usually 5-10 pages, never longer) that includes:

1. Workflow map with bottlenecks highlighted
2. Quantified cost table for each bottleneck
3. Classification of each item (simple automation, AI automation, AI assist, leave alone)
4. Data readiness assessment for AI-classified items
5. Recommended projects in priority order (quick wins first)
6. ROI projection for each recommended project
7. Honest assessment of what won't work and why

Sequence the projects so the business gets value fast. The first project should always be a quick win: a simple automation that takes 1-2 weeks to build and delivers obvious results. This builds trust and momentum for the harder projects.

For AI projects specifically, break the ROI into:

• One-time build cost
• Monthly running cost (API calls, hosting)
• Monthly maintenance estimate (my time for monitoring and updates)
• Expected time savings in hours per week
• Payback period in months

Be conservative with projections. Nothing kills trust faster than promising 80% time savings and delivering 40%. Better to estimate 40% and deliver 60%.

What a real audit looks like

To make this concrete, here's what an audit might find for a hypothetical 15-person recruitment agency:

• 12 hours/week spent manually formatting CVs into the agency's template (simple automation, not AI)
• 8 hours/week screening incoming applications against job requirements (AI classification)
• 6 hours/week writing initial candidate outreach emails (AI assist)
• 4 hours/week updating their CRM from email threads (simple automation)
• 3 hours/week on admin tasks too small to automate

Total identified: 33 hours per week, roughly £34,000 per year in staff time.

The recommended roadmap:

1. CV formatting tool (simple template automation, 2 weeks, ~£3,000)
2. Application screening classifier (AI, 3 weeks, ~£6,000)
3. CRM sync automation (Zapier + email parsing, 1 week, ~£1,500)

Total investment: ~£10,500. Expected annual savings: around £26,000. Payback: under 6 months.

The email outreach assistant goes on the backlog for Phase 2, and the small admin tasks stay manual. Quick wins first, complex AI second.

Want an honest assessment?

If you're wondering where AI fits (or doesn't) in your business, an audit is the best way to find out. I'll spend time understanding your operations, look at your data, and give you a clear, prioritised roadmap. If the answer is "you don't need AI right now", I'll tell you that too.

Book a free initial chat and I'll let you know if an audit makes sense for your situation.