AI Audit for SMEs

An AI audit for an SME should concentrate limited time on the uses that can create meaningful harm, loss or disruption. Proportionality means choosing a defensible scope and evidence plan for the business. It does not mean declaring a small company low risk, accepting supplier claims without examination or replacing operating evidence with a generic policy pack.

A small business may use a sophisticated AI service through a simple subscription. The provider controls much of the technology, while the business controls what staff upload, who can access it and how outputs are used. A useful audit examines that real division of responsibility. It should leave the owner with understandable findings and decisions, not a collection of enterprise documents that no one has capacity to maintain.

Start with actual use, including informal use

Ask staff which AI-enabled tools they use, what they enter and what happens to the output. Include features embedded in ordinary software, browser tools and personal subscriptions used for work. The purpose is to understand the workflow, not to punish candid disclosure. If staff fear that every answer will lead to blame, the inventory may omit the very uses that need attention.

Record a named owner, business purpose, input data, output recipients, permitted actions, supplier and review step. A short maintained register is more useful than a sophisticated inventory that becomes stale. Distinguish drafting from sending, recommending from approving and summarising from changing a record. Those transitions determine which consequences the audit needs to examine.

NIST describes its voluntary AI RMF as scalable to organisations of different sizes. That supports a tailored approach, not a claim that adopting a framework certifies the business. Use a proportionate method to decide what evidence is needed for the actual use and consequences.

Example: a small horticultural supplier with two AI uses

Consider a hypothetical supplier with twenty-two employees. It uses one AI tool to draft product descriptions from approved catalogue records and another to summarise internal staff meetings. Product descriptions are reviewed before publication. Meeting summaries help assign follow-up tasks and are shared with a limited staff group. Neither tool is authorised to provide plant-treatment safety advice, make employment decisions or publish without review.

The two uses should not receive the same checklist simply because both generate text. The product tool raises questions about unsupported or outdated claims reaching customers. The meeting tool raises questions about access, attribution, omissions and the handling of staff information. The business may have the same subscription supplier for both, but the data, users and consequences differ.

Original SME inventory example: two uses, different audit questions

Use	Relevant evidence	Decision the audit supports
Catalogue description drafting	Approved source version, selected drafts, reviewer amendments and publication record	Can the business keep this workflow within its factual source and review boundary?
Staff-meeting summaries	Authorised source notes, sharing settings, selected summaries and correction records	Are access and attribution controls adequate for the intended staff use?
Unapproved personal AI account	Observed business use and available account settings, without intrusive unrelated inspection	Should the business restrict the use or move it to an approved environment?

The third row represents a possible discovery, not an assertion about the hypothetical staff. The inventory should have room for unknowns and unapproved uses. An incomplete answer is a reason to investigate or restrict a workflow, not a reason to mark it low risk because the business lacks documentation.

Choose scope by consequence and controllability

Prioritise uses that expose sensitive information, communicate externally, affect people or trigger changes that are hard to reverse. Then consider how much the business can observe and control. A read-only drafting tool with a genuine review step differs from a tool that sends messages automatically. A small number of users can still create a material issue if the data or decision is consequential.

Use a plain-language prioritisation record: the potential consequence, affected people or process, current control, evidence gap and proposed audit question. Avoid a complex score that produces a reassuring average while hiding one severe issue. If the business has limited funds, it can examine the most important workflow first and explicitly leave others outside the conclusion.

Scope exclusions should be operationally useful. For example, the supplier can prohibit automatic publishing while the draft-review process is examined. That restriction changes the permitted workflow and should be enforced and communicated. It is not enough to write “out of scope” in the report while staff continue the excluded use as normal.

A minimum evidence pack that a small team can maintain

For each selected use, prepare a short description of the task and permitted actions, the supplier and application version where available, key settings, approved source material, a list of relevant events and known problems. Add records showing who reviewed outputs and what changed before use. Keep sensitive data masked where possible and agree handling before sending evidence to an assessor.

Evidence should reflect what actually happened. Do not backdate approvals or create meeting records that imply a control operated when it did not. A new procedure can be valuable, but it supports a statement about the new process, not a historical conclusion. Audit preparation explains how to organise existing material without manufacturing a favourable past.

The evidence pack need not contain every conversation. Collect enough to answer the agreed questions and explain the selection. For the meeting tool, the business may use authorised, masked examples to test attribution and sharing controls while limiting unnecessary exposure of staff information. For the catalogue tool, retained source versions and final published text may be more important than a large archive of intermediate prompts.

Practical test one: source support for product descriptions

In an illustrative test, the team selects forty product descriptions from a defined review period, deliberately including older catalogue items and bilingual variants. Twenty-eight are supported by the current approved records, seven use outdated source information and three contain claims not supported by the supplied records. Two remain unresolved because reviewers disagree on the meaning of a product term and need a subject-matter decision. Keep those two separate from both supported descriptions and confirmed exceptions. These invented results describe the selected cases; the targeted selection does not estimate the error rate across the entire catalogue.

The audit should distinguish an unsupported draft from an unsupported published claim. If a reviewer corrected the draft before publication, the review control may have contained the problem. If the old wording reached the public catalogue, the finding concerns the actual source-to-publication path. The record should connect source version, generated draft, amendment and final text rather than treating a screenshot of the draft as the whole outcome.

Management can then choose a proportionate response. It might refresh the approved source collection, block publication until source checks are complete and retest affected categories. The audit does not need to redesign the whole e-commerce platform to make this decision. It does need enough evidence to show which boundary failed and what would demonstrate correction.

Practical test two: attribution in staff summaries

For the staff-summary tool, use authorised examples with known speaker or task attribution and test whether the summary preserves uncertainty. A sentence such as “we should discuss whether Lina can lead this task” should not silently become an approved assignment to Lina. The business consequence is different from a product-description error: staff may act on a responsibility that was never agreed.

Examine who can view the source and summary, whether access changes propagate and how a correction is made visible to recipients. A deleted source does not automatically establish deletion of every derived summary, and a sharing setting in one workspace does not establish the setting in another. Ask the supplier for relevant evidence and document what the business can and cannot verify.

The outcome should be a bounded decision about the intended use. If reliable attribution cannot be established, the business might require meeting owners to confirm tasks before circulation and label unresolved items explicitly. That is a proposed control to test, not a guarantee that human review always works. The audit should later examine records of the review if an operating conclusion is needed.

A small-business decision card

Illustrative decision card after a focused SME assessment

Field	Catalogue workflow	Staff-summary workflow
Decision now	Continue reviewed drafting; hold affected publication paths until correction	Limit circulation to authorised staff and require task-owner confirmation
Accountable owner	Catalogue manager	Meeting owner with the responsible business manager
Evidence required for closure	Updated source mapping and retest of affected descriptions and publication checks	Corrected attribution examples, sharing checks and evidence of the agreed review
Remaining limit	No conclusion on unexamined catalogue categories or later source changes	No conclusion on unrelated personal accounts or unobserved supplier internals

The card is an original management aid, not a certification. It gives the owner a decision, a responsible person and the evidence needed to revisit the restriction. It should not replace the supporting findings, but it can keep the outcome usable for a small team that does not have a full-time assurance department.

Supplier responsibility does not remove customer responsibility

A software supplier may operate the model and infrastructure, while the SME determines what data is uploaded, who has access and how results are used. Ask for evidence relevant to the service and configuration you buy. A general security certificate or model card may help, but does not automatically cover your staff sharing settings or factual publication process.

Use a concise request rather than demanding every proprietary document. Ask about the specific data use, retention, access, version changes and control evidence needed for the audit. Where the supplier cannot provide evidence, identify the resulting uncertainty and decide whether restrictions or another service are necessary. Supplier evidence quality helps assess responses without confusing a marketing statement with proof.

A small customer may have little negotiation power, but it still has choices about permitted data and actions. The audit can identify where a limited use remains supportable and where a higher-consequence use cannot be justified with the available evidence. A lack of bargaining power does not make an unanswered question disappear.

Saudi applicability should be checked, not guessed from company size

Do not assume that being an SME removes all obligations, or that every requirement applying to a large regulated entity applies identically to a small supplier. The relevant analysis depends on the activity, entity and processing. Record who determines applicability and which requirements are included in the audit criteria. Where legal interpretation is needed, obtain appropriately qualified advice for the actual situation.

For staff information, the PDPL audit guide explains the evidence questions around processing and responsibilities. A technical audit should not claim comprehensive compliance merely because a tool is hosted in a particular region or a supplier advertises a compliance feature. The business needs a clear account of its own use and the evidence supporting the relevant controls.

Keep ownership simple without eliminating challenge

A small business may combine roles, but should still identify who operates the tool, who approves important use changes and who challenges the evidence. The owner who selected the software can contribute context without being the only person deciding whether the assessment is adequate. External specialist help can be focused on the judgement the business cannot competently make itself.

If the same provider built the workflow and offers the audit, ask how self-review is addressed and how the report will describe the role. Independence should be assessed specifically, not assumed from an external invoice. The buyer also needs competence appropriate to its tasks and languages; a generic AI demonstration is not evidence that the assessor can evaluate Arabic meaning or the business's source material.

Control cost through scope and readiness

Keep the first engagement bounded around a decision the owner can act on. Prepare authentic evidence, nominate available reviewers and avoid asking for a universal “AI compliant” badge. A narrow but well-supported assessment can be more useful than a broad report whose conclusion rests on documents alone. Exclusions and residual uncertainty must remain visible.

Compare AI audit pricing by included evidence work, language coverage, reporting and retesting. Do not buy only a low headline price without knowing what it omits. Equally, do not assume an enterprise-sized package is necessary because the technology is complex. Ask the provider to explain the additional decision value of each proposed procedure.

Make the next review feasible for the person who will do it

A proportionate audit should leave a small team with a maintainable evidence habit. For the catalogue workflow, that could mean retaining the approved source reference, final text and reviewer decision for the relevant publication events. For the meeting workflow, it could mean preserving the agreed task confirmation and correction status. The exact record depends on the claim; the business does not need to archive every intermediate model response by default.

Walk through the proposed habit with the employee who will maintain it. Can they identify the correct source version? Can they record a correction without overwriting the original evidence? Can a substitute perform the task when the usual owner is absent? If the answer is no, the control may be too complicated or poorly integrated into the actual work. Simplify the process while preserving the evidence needed for the decision.

Small teams also need a clear exception route. A staff member who cannot verify a product statement should know whether to hold it, remove the uncertain claim or ask the authorised owner. A meeting summary with disputed attribution should not be circulated as an agreed task list merely because the normal reviewer is busy. The audit should examine whether the proposed route is understandable and practically available.

Use restrictions as real operating choices

When evidence is incomplete, a business can sometimes reduce consequence by narrowing what the tool may receive or do. For example, a catalogue assistant may remain limited to draft text from approved records, with publishing performed through a separate reviewed process. That restriction needs implementation and communication. A written rule that staff routinely bypass is not evidence that the risk has been contained.

Document what would justify removing the restriction. The owner may require a corrected source process, suitable tests and evidence that review occurs before publication. This turns a temporary limit into a decision with a path forward. Without a closure condition, temporary restrictions can either become permanent by neglect or disappear informally when staff find them inconvenient.

The same reasoning applies when the business stops using a tool. Identify what data and outputs remain, who needs access to records and what supplier settings or requests are relevant to the agreed exit. Do not assume that cancelling a subscription resolves every evidence or data-handling question. The appropriate steps depend on the service and the organisation's actual responsibilities.

A useful final readout should let the owner explain each material decision in ordinary language: this use can continue under these conditions; this use is restricted for this reason; this evidence is missing; this person owns the next step. If the owner cannot explain the outcome without the consultant present, improve the handover before considering the assessment complete. Clear ownership and evidence are more durable than a complicated document library that the business cannot use.

A small business should be able to operate the agreed follow-up process without depending permanently on the person who wrote the audit report.

After the report: maintain the decisions

Assign each action an owner and a closure condition. Retain the tested version and relevant evidence, and record changes that affect the conclusion. A new supplier feature, an automatic publishing setting or a different data source can turn a previously narrow use into a more consequential one. The owner should know when to seek another review rather than treating the report as permanent approval.

Use retesting to verify material corrections, and keep a simple record of unresolved risks and restrictions. Do not mark an issue closed merely because a policy was written or a vendor ticket was submitted. The evidence should show that the relevant condition changed and that the agreed control now works in the tested circumstances.

For an audit scoping discussion for a small business, bring a short inventory, two or three material concerns and a masked example of the workflow. Open for AI's AI auditing services can be scoped to the evidence and decisions that matter most. The result should fit the business's capacity while preserving the discipline needed for a credible conclusion.