AI Audits for SMEs: Choose the Workflow Before the Model

Many small and medium-sized businesses start an AI project with the wrong question: which model should we use?

It is an understandable question, but often a premature one. Before comparing models, prices, and benchmarks, it is better to find out whether the chosen process is clear, measurable, and reversible enough to improve with AI.

An AI audit is not a slide deck full of promises. It is an operational picture: where the data lives, which decisions are genuinely repetitive, which mistakes cost the most, and where a person is still needed.

The unit of analysis is the workflow, not the model

A model is one component. The value, and the risk, come from the system around it.

For example, “use AI to answer emails” is too vague to be a project. An auditable workflow is more specific:

1. receive a request through a defined channel;
2. extract customer, urgency, and category;
3. retrieve information from authorised sources;
4. draft a reply with verifiable references;
5. request approval when the reply changes prices, timing, or terms;
6. record the outcome and any correction.

The difference may look small, but it changes everything. The second workflow has inputs, permissions, metrics, and stop conditions that can be defined.

Five questions for a practical AI audit

### 1. What is the measurable goal?

“Become more efficient” is not enough. The goal might be reducing triage time, removing manual handoffs, or increasing the percentage of correctly classified requests.

Without a baseline, the project can become a permanent demo: interesting to show, difficult to evaluate.

### 2. Which data can the system read?

Separate necessary data from data that merely happens to be available. An agent classifying a request probably does not need to read every company document.

The audit should record the source, owner, quality, update frequency, and retention rule for every data source used by the workflow.

### 3. Which actions can it take?

Reading, suggesting, editing, and sending are different permissions. A well-designed system starts with the minimum level and increases autonomy only when suitable controls exist.

For an SME, it is often sensible to start with an agent that prepares an output and a person who approves it. That is not an automation failure. It is an explicit operational boundary.

### 4. How do we verify the result?

Verification should not happen only at the end. Add checkpoints for the source used, data changed, rule applied, confidence, and reason for stopping.

For document-based workflows, a minimum test can include normal, incomplete, and ambiguous cases. If the system works only on the ideal case, it is not ready for real work.

### 5. How do we undo the action?

Every action with external consequences needs a rollback or correction strategy. This includes CRMs, catalogues, orders, email, and shared documents.

A reversible automation is easier to approve and improve. Reversibility does not remove risk, but it prevents one mistake from becoming an untraceable incident.

A simple score for choosing where to start

You do not need a sophisticated mathematical model to compare workflows. Score each one from one to five for:

• how repetitive the work is;
• how clear the objective is;
• data quality and accessibility;
• cost of an error;
• ease of verification;
• reversibility of the action;
• availability of a human owner.

A highly repetitive process with messy data and irreversible consequences is not necessarily the best first candidate. A smaller, well-documented, easily checked process may create value sooner.

That is why a serious AI audit often reduces the original technical ambition. Lowering the ambition can increase the chance of producing something people can actually use.

Security and governance are not an appendix

The audit should also cover prompt injection, excessive access, sensitive data, unverified outputs, and dependence on a single provider.

The NIST AI Risk Management Framework is a useful reference for structuring risk identification and management. OWASP also maintains guidance on risks in large language model applications. These are not checklists to attach at the end; they help teams ask the right questions during design.

The practical question for an SME is: who notices the mistake, how long do they have to correct it, and what trace remains?

From report to first experiment

An audit should not end as a forgotten PDF. The best outcome is a small experiment with clear boundaries:

• one process;
• a controlled data set;
• minimum permissions;
• a group of anonymised real cases;
• a human checkpoint;
• one before-and-after metric;
• a review window.

After the test, decide whether to extend, correct, or stop the workflow. Stopping is also a useful result: it prevents an interesting technology from becoming a permanent operating cost.

In summary

An AI audit for an SME is not meant to prove that AI can do everything. It is meant to identify where it can do something useful, under which limits, and with whose responsibility.

The process, data, permissions, and verification come before the model. Only then does it make sense to choose the technology.

If you want to assess a concrete workflow in your business, contact Studio Futuro.