AI can help a small business work through a crowded recruitment inbox, identify sales prospects worth calling first, spot a suspicious order or support a lending decision. Used well, it can save time and bring more consistency to routine work. Used carelessly, it can quietly deny someone a job, service, credit option or fair price with no real opportunity to understand or correct what happened.
That is the line UK small business owners need to recognise: AI can assist a person to decide, or it can make a significant automated decision about an individual. The distinction is not determined by what the software supplier calls its product, whether it uses a large language model, or whether a manager clicks “approve” at the end. It depends on the real process and the effect on the person.
Since 19 June 2026, all data-protection provisions in the Data (Use and Access) Act 2025 have been in force. The Act widened the circumstances in which organisations can make significant decisions solely through automated processing, but it retained important safeguards for people. The ICO’s DUAA overview is clear that innovation does not remove the need for appropriate protections.
For SMEs, the practical message is straightforward: do not wait for a complaint to discover that a “helpful score” was effectively an unchallengeable rejection. Build human judgement, clear explanations, correction routes and fairness testing into the process before the tool goes live.
AI assistance is not the same as an automated decision
AI assistance, sometimes called decision support, gives a human useful input. It might summarise CVs against an agreed job description, flag inconsistent information in a loan application, estimate a lead’s likelihood of converting, or highlight an unusual payment pattern. A properly equipped person then considers the recommendation, the underlying evidence and the individual circumstances before making the final call.
A decision becomes solely automated when there is no meaningful human involvement. Under the amended UK GDPR rules, a significant decision is one with a legal effect or a similarly significant effect on the individual. The ICO’s summary of the legal change confirms both points and explains that organisations must also consider the role of profiling when judging whether human involvement is meaningful.
“Meaningful” is the vital word. A manager who is expected to accept every recommendation, sees only a traffic-light rating, lacks access to the facts behind it or has no time or authority to overturn it is not providing a meaningful safeguard. A human click after an algorithm has already filtered out everyone below a threshold is likely to look like rubber-stamping rather than a genuinely human decision.
Conversely, a person does not have to ignore the AI. They can use it as one input, provided they are trained to question it, can access relevant information, have enough time to assess the case and can genuinely accept, reject or amend the recommendation. Design the workflow to prove that this happens in practice, not merely in a policy document.
Where small businesses can cross the line
Recruitment and candidate screening
Using AI to draft interview questions, remove duplicate applications or organise CVs is one thing. Automatically rejecting applicants because their CV lacks selected keywords, their video interview score falls below a cut-off, or their career history does not resemble past hires is much riskier. Losing the chance to compete for a job is capable of being a similarly significant effect.
A safer recruitment design is to let AI identify points a recruiter may wish to examine, while ensuring a trained hiring manager reviews the whole application against job-related, pre-set criteria. The manager should be able to see why the tool produced its result, consider reasonable explanations for gaps or non-standard experience, and record their own decision. Do not use protected characteristics, proxy signals or irrelevant personal information as a shortcut to perceived “fit”. Employment decisions also remain subject to equality law, regardless of whether a human or software made the error.
Credit, payment terms and affordability
A model that automatically declines credit, reduces a credit limit or offers markedly worse payment terms can have an obvious material impact. The same applies if an AI-generated risk label effectively prevents a customer from accessing a product. An SME offering trade credit, instalment payments or subscription services should treat this as more than a back-office efficiency project.
Use clear eligibility criteria, retain the inputs and outcome for each case, and give people a practical route to provide additional context. A recent missed payment, a thin credit file or an incorrect address match may not tell the whole story. A staff member reviewing a challenge needs authority to reach a different result, rather than simply repeating the model’s conclusion.
Pricing and insurance-style risk signals
Dynamic pricing can be legitimate, but individualised pricing based on profiles deserves careful scrutiny. If data about a person’s behaviour, location, device or inferred characteristics changes the price they pay or the product they can access, assess the impact rather than assuming it is ordinary marketing. A small difference on a low-value optional purchase may be unlike a substantial price increase or exclusion from an essential service.
Set boundaries before deployment. Specify which data may influence price, which data is prohibited, the maximum permitted adjustment, who can approve exceptions and when a person must review the outcome. Your privacy information should not hide the practice behind vague language such as “we use advanced technology to improve your experience”.
Fraud flags and account restrictions
AI can be valuable for identifying patterns that may suggest fraud. But a fraud score is an alert, not proof. Automatically cancelling an order, freezing an account, refusing a refund or blacklisting a customer on the basis of a weak or opaque signal can cause real harm. It can also create a frustrating loop where the customer cannot discover what went wrong or offer evidence that the transaction was genuine.
Match the response to the confidence and consequence of the signal. A low-confidence alert might trigger extra verification; a high-impact restriction should normally receive prompt, informed human review. Keep a process for correcting mistaken matches, stale information and identity confusion, particularly where third-party data is involved.
Lead scoring and sales prioritisation
Lead scoring is often lower risk, but not automatically risk-free. A model that helps a salesperson decide which business contact to call first may be a sensible internal prioritisation tool. It is less likely to be a significant decision if no one is denied a service, opportunity or meaningful choice as a result.
The analysis changes when a score automatically suppresses offers, withholds access to a scheme, excludes a sole trader from a business opportunity or causes an individual to receive materially different treatment. Business-contact details are still personal data when they identify a person. Even where the decision is not significant, you still need a lawful basis, fair and transparent processing, data minimisation, security and a way to respect applicable data-protection rights.
Put a meaningful human review into the workflow
A credible review is a designed control, not an emergency escalation email. Decide at the outset which outcomes require a human decision on every case and which alerts need review before a customer or candidate experiences a negative effect. Higher-impact decisions should have stronger controls.
- Give the reviewer the full picture: show the principal factors, source data, confidence or limitations, relevant documents and the individual’s explanation where available.
- Give them authority: the reviewer must be able to change the result, grant an exception or send the case for further investigation.
- Give them capability: train staff on the tool’s purpose, known weaknesses, inappropriate uses and signs that an output may be unreliable.
- Give them time: a reviewer asked to clear 300 cases in an hour is being pushed towards automatic approval.
- Keep an audit trail: record the AI recommendation, the human decision, the reasons for disagreement where relevant, and any evidence supplied by the person.
This also improves operations. Reviewing overrides and challenges can reveal that a scoring threshold is too harsh, a data source is out of date or a particular scenario is confusing the model. That feedback is much more useful than treating staff intervention as a failure of automation.
Be transparent before a person has to complain
Transparency means telling people in plain language when their personal data is used in significant automated decision-making, what that means for them and what they can do about it. Under the DUAA safeguards, a person must be given information about a significant decision and be able to make representations, obtain human intervention and contest it.
You do not need to publish source code or hand over trade secrets. But a useful notice should explain the purpose of the system, the categories of data and main factors used, whether profiling is involved, the possible outcomes, the role of human reviewers, and how to challenge a result. The explanation should help a real person decide whether information was wrong, incomplete or misunderstood.
For example, a candidate could be told that screening assesses evidence against essential job criteria, that an automated tool may organise applications for a hiring manager, and that they can ask for a review if they believe relevant experience was overlooked. A customer receiving a fraud-related cancellation could be told what action was taken, how to contact a trained reviewer, what information may help reassessment and the expected response time.
A route buried in generic customer support is not enough. Make the challenge channel easy to find, free to use and workable for people who need reasonable adjustments or cannot use a digital form. Train frontline colleagues to recognise a request for review, rectification or an explanation and to send it to someone who can act.
Start with accurate, relevant data—not a persuasive dashboard
AI can produce a convincing score from poor data. That is why data-protection accuracy and model accuracy need separate attention. Data-protection accuracy concerns whether personal data is correct and, where necessary, current. Statistical accuracy concerns how well a model’s predictions perform against appropriate test data. The ICO’s AI accuracy guidance stresses that inferences are often statistically informed guesses, not facts, and that their limitations must be accounted for when decisions affect people.
Before switching on a system, map every input: CRM records, CVs, credit information, web activity, transaction history, third-party enrichment and human notes. Ask who supplied it, when it was updated, whether it is relevant to the purpose and whether it is likely to contain errors or historical bias. Do not let a supplier’s claim that its data is “enriched” substitute for this work.
Label predictions as predictions. Do not write “high fraud risk”, “poor cultural fit” or “unlikely to pay” into a permanent customer or staff record as though it were established fact. Set retention periods, correction processes and controls over who can see sensitive assessments. If a person disputes an input or inference, investigate it rather than assuming the model must be right.
Test for unfair outcomes before and after launch
An overall performance score can conceal serious problems. A recruitment tool may appear accurate in aggregate but reject suitably qualified candidates from one group more often. A fraud model may generate disproportionate false positives for particular communities. A lead score may systematically deprioritise small firms in certain locations because it has learned patterns from historic sales decisions rather than genuine commercial value.
Test the actual use case before launch and regularly afterwards. Compare error rates, false positives, false negatives, selection rates and outcomes across relevant groups where it is lawful and proportionate to do so. Examine proxy variables too: postcode, school, gaps in employment, language patterns, device type and purchasing history can reproduce disadvantage even when protected characteristics are not explicit inputs.
Testing is not a one-off supplier demonstration. Re-test when you alter the model, data source, threshold, product, customer base or human workflow. Monitor override and appeal rates, then investigate clusters. If the system cannot be tested adequately, explained sufficiently or controlled safely for the proposed impact, narrow its role or do not use it for that decision.
Use a DPIA and interrogate your supplier
A Data Protection Impact Assessment, or DPIA, should shape the project before procurement or rollout. The ICO says a DPIA is required for processing likely to create high risk and specifically highlights systematic, extensive profiling or automated evaluation with significant effects. Its DPIA guidance and checklist also flags scoring, automated decisions, access to services and innovative technology as key risk indicators.
Your DPIA should document the purpose, lawful basis, data flows, likely effects, alternative lower-risk options, human-review design, transparency wording, challenge route, data retention, security, supplier responsibilities and fairness testing. Keep it live: revisit it when the system or its context changes.
Ask suppliers direct questions. What data is used to train and operate the tool? Can you see the factors behind an individual output? Can you export records for a review or complaint? What known limitations and bias tests exist? Will the vendor reuse your data? Where is it processed? Can thresholds be adjusted, and can you switch the tool off safely? A contract does not transfer your accountability for how your business uses personal data.
Make safeguards the business case
AI should help your team make better decisions, not make it harder for people to understand, correct or challenge decisions that affect them. Start by classifying each use case, identify whether the effect could be legal or similarly significant, and build safeguards proportionate to the risk. Where in doubt, seek specialist data-protection and employment or consumer-law advice before deploying a high-impact system.
The strongest SME approach is not “human in the loop” as a slogan. It is a real person with information, training, time and authority; clear notice to the people affected; a prompt route to challenge; accurate data; and ongoing evidence that the system is not producing unjustified unfair outcomes. Put those controls first and AI can support growth without sacrificing trust.
