Artificial intelligence is no longer developing as a single category of software. It is becoming a moving stack of capabilities: chatbots, copilots, document analysers, workflow automations, coding assistants, research tools and increasingly autonomous agentic systems that can take action across business processes.
For small and medium-sized businesses, this creates an uncomfortable problem. The tools are becoming more useful at exactly the same time that they are becoming harder to control.
A few months ago, the main AI governance question was relatively simple: should staff be allowed to use tools such as ChatGPT, Claude or Microsoft Copilot, and what information should they avoid putting into them? That question still matters, but it is no longer enough.
Modern AI tools are not just drafting emails or summarising notes. They can read documents, interpret contracts, search databases, write code, trigger workflows, analyse customer data, generate reports, make recommendations, interact with other systems and, in some cases, operate with limited human intervention. The boundary between “AI assistance” and “automated business activity” is becoming much less clear.
That is where governance becomes urgent.
The real challenge for SMEs is not whether AI is useful. It clearly is. The challenge is whether business controls, risk assessments, vendor due diligence and data protection processes are keeping pace with what the tools can now do.
New AI features are changing the risk classification
The problem with AI adoption is not just speed. It is that new feature sets can change the nature of the activity being performed.
A tool that was originally approved for low-risk content drafting may later gain the ability to analyse uploaded files. That changes the data protection position. It may then gain memory, retrieval, workflow automation or integration with email, CRM or project management systems. That changes the operational and security position. It may then gain agentic functionality, allowing it to plan steps, call tools and complete tasks across systems. That changes the control environment completely.
The business may still think it has approved “an AI writing assistant”. In reality, it may now be operating a multi-purpose decision-support and automation layer inside the organisation.
That matters because governance does not attach only to the name of the product. It attaches to what the product does, what data it touches, what decisions it influences, what systems it connects to and what harm could result if it fails.
This is why feature-led AI adoption can be dangerous. A business may approve a supplier based on one use case, while the supplier’s roadmap rapidly expands into areas that were never considered during onboarding.
Agentic AI breaks traditional approval assumptions
Autonomous and semi-autonomous AI agents are a good example of this problem.
A normal chatbot responds to prompts. An agentic AI product may pursue an objective. It may break a task into steps, decide which tool to use, call an API, search a database, draft a response, classify information, update a record or recommend an action.
That is commercially attractive because it moves AI from “assistant” to “operator”. It can reduce manual effort, speed up processes and connect work across departments.
But it also breaks several traditional business assumptions.
First, it makes the user less directly responsible for each individual step. The human may approve the objective, but not every intermediate action the agent takes.
Second, it can blur the difference between recommendation and execution. A tool that merely suggests a customer response creates one category of risk. A tool that sends the response, updates the CRM and triggers a follow-up workflow creates another.
Third, it creates chain risk. If the AI relies on a weak data source, misinterprets a policy, calls the wrong system function or acts on stale information, the failure may propagate across the workflow before anyone notices.
Fourth, it creates accountability gaps. When something goes wrong, the business still needs to answer basic governance questions: who approved the use case, what data was processed, what decision was made, what controls applied, what human review occurred, and what evidence exists?
Those questions cannot be answered by enthusiasm for a new feature. They require operational governance.
Why existing business controls are often too slow
Most business control frameworks were not designed for AI feature velocity.
Supplier onboarding is usually periodic. DPIAs are often completed at the beginning of a project. Information security reviews may focus on hosting, access control and certifications. Procurement teams may rely on standard supplier questionnaires. Legal review may examine the contract once, before signature.
That model assumes the product being approved is relatively stable.
AI products are not stable in that way.
A supplier can release new functionality every few weeks. A feature that did not exist during procurement may become central to the product six months later. A model upgrade may change output quality, reasoning behaviour, data handling assumptions or automation capability. A new integration may bring the tool into contact with data that was never included in the original assessment.
This does not mean businesses should avoid AI. It means AI governance needs to become more dynamic.
A one-off approval is not enough when the product’s capability envelope keeps expanding.
DPIAs need to become living AI control documents
Data Protection Impact Assessments are particularly affected by this shift.
Under UK data protection guidance, organisations should consider DPIAs when AI processing involves personal data and may create high risks to people’s rights and freedoms. The ICO’s AI and data protection guidance specifically highlights the need to consider DPIAs for AI systems and the accountability implications of AI use.
For SMEs, the practical issue is that many AI uses start small. A team may begin by using AI to summarise internal notes. Later, the same tool may be used to analyse customer complaints, employee performance notes, call transcripts, supplier documents or financial records.
At that point, the risk profile has changed.
A DPIA should not be treated as a one-time compliance form completed when the tool is first purchased. For AI systems, it should become a living control document that tracks:
what the AI tool is used for,
what categories of personal data are processed,
whether special category data may be included,
whether outputs influence decisions about individuals,
whether automated decision-making or profiling risks exist,
what human review is required,
what supplier data retention and training terms apply,
what technical controls are in place,
what residual risks remain,
and what changes require reassessment.
The moment an AI product gains new capability, the DPIA may need to be revisited. The issue is not only whether the supplier has changed its terms. The issue is whether the business has changed how the tool is used.
Vendor due diligence must move beyond standard supplier checks
AI vendor due diligence also needs to evolve.
Traditional supplier due diligence often focuses on financial stability, information security certifications, data processing terms, cyber controls and contractual protections. Those remain important, but they are not enough for AI.
AI due diligence needs to ask capability-specific questions.
Does the supplier train models on customer data?
Are prompts and outputs retained?
Can customer data be reviewed by humans?
Where is data processed and stored?
What subprocessors are involved?
Can the system connect to internal applications?
Can it take actions automatically?
Does it maintain memory?
Can users upload files?
Can outputs be explained or audited?
How are model updates managed?
What logs are available?
Can risky features be disabled?
Can access be controlled by role?
What happens when the model is wrong?
Who is responsible if the AI output causes harm?
The OECD’s 2026 Due Diligence Guidance for Responsible AI frames AI due diligence as an ongoing process connected to risk identification, prevention, mitigation, tracking and communication, rather than as a one-off procurement activity.
That principle matters for SMEs. A supplier may look acceptable at onboarding, but the risk may increase later because the product changes, the internal use case expands, or staff begin using the tool in ways leadership did not anticipate.
Shadow AI increases when governance cannot meet demand
There is another commercial reality: if official processes are too slow, employees will route around them.
If staff see powerful AI tools that help them work faster, and the business has no clear approval route, shadow AI will increase. People will use personal accounts. They will paste data into unapproved tools. They will upload documents to free services. They will rely on AI outputs without review. They will experiment in ways that are invisible to leadership.
This is not usually malicious. It is often a sign of unmet demand.
The answer is not to block everything. That rarely works for long. The better answer is to create a governance route that is fast enough to be usable and strong enough to protect the business.
For SMEs, that means simple, practical controls:
approved AI tools list,
clear permitted and prohibited use cases,
data classification rules,
mandatory review for higher-risk outputs,
supplier risk checks,
DPIA triggers,
AI use register,
escalation path for new tools,
board-level reporting on AI adoption and risk.
Governance should not feel like a separate bureaucracy. It should be embedded into how the business chooses, approves, deploys and monitors AI.
Business controls must follow capability, not brand
One of the biggest mistakes SMEs can make is treating AI governance as a supplier-brand issue.
The question is not simply whether ChatGPT, Claude, Gemini, Copilot or another platform is “safe”. The better question is: safe for what use, with what data, under what configuration, with what controls, for which users, and with what review?
The same AI platform may be low risk for drafting a public blog outline and high risk for analysing employee disciplinary records. It may be acceptable for internal brainstorming but unsuitable for customer-facing automated advice. It may be appropriate for summarising anonymised documents but inappropriate for processing identifiable health, HR or financial data.
Business controls need to follow the actual use case.
That requires a more granular approach to AI onboarding. Instead of approving a tool once and letting usage expand informally, businesses should approve combinations of:
supplier,
product,
feature set,
use case,
data category,
user group,
integration,
decision impact,
human oversight requirement,
monitoring requirement.
That sounds more complex, but it is actually more practical. It lets a business say yes to low-risk AI use quickly while applying stronger controls to use cases that genuinely need them.
Board reporting is becoming essential
AI governance should not live only with IT, legal or compliance. It needs board-level visibility.
This does not mean every AI tool needs to be discussed in board meetings. It means senior leaders need a clear view of the organisation’s AI exposure.
Useful board reporting should answer:
which AI tools are approved,
which business areas are using them,
which high-risk use cases exist,
which suppliers process sensitive or personal data,
which DPIAs are complete or overdue,
which vendors have unresolved due diligence issues,
which AI incidents or near misses have occurred,
which new features require reassessment,
and where shadow AI risk is increasing.
For SMEs, this can be proportionate. A simple AI governance dashboard is better than a complex policy nobody reads.
The key is visibility. Leadership cannot manage AI risk if it does not know where AI is being used.
The new governance model: continuous, evidence-based and operational
The next phase of AI governance will not be about writing longer policies. It will be about operational control.
Businesses need a repeatable way to assess new AI tools, classify use cases, review data protection risk, challenge supplier claims, approve deployment, monitor changes and produce evidence when needed.
That is especially important as agentic AI becomes more common. Once AI tools begin taking actions across workflows, the business needs to know not only whether the supplier is reputable, but whether the workflow itself is controlled.
The practical governance model should include:
an AI use case intake process,
risk classification by capability and data type,
vendor due diligence specific to AI,
DPIA screening and completion where required,
approval gates before deployment,
technical guardrails where available,
user training linked to real use cases,
logging and evidence capture,
periodic reassessment,
and board-level reporting.
This is how SMEs can adopt AI confidently without pretending that every tool is either harmless or unacceptable.
How Strategic AI Guidance can help
Strategic AI Guidance builds practical tools and advisory services to help organisations adopt AI with stronger governance, not slower adoption.
Our AI Vendor Due Diligence Assistant helps businesses assess AI suppliers more consistently by structuring the due diligence process around the risks that matter for AI products, including data handling, security, model behaviour, supplier controls, contractual risk, auditability and operational dependency.
AI Vendor Due Diligence Assistant:
https://www.strategicaiguidance.com/vendor-dd/
Our AI DPIA Companion supports structured assessment of AI-related data protection risks, helping organisations identify when a DPIA is needed, capture the right information, assess risks and generate a more defensible review process.
AI DPIA Companion:
https://www.strategicaiguidance.com/dpia/
These tools are designed for the reality businesses now face: AI products are changing quickly, governance requirements are becoming more complex, and leadership teams need evidence-based control without slowing useful adoption to a standstill.
Conclusion
AI features are moving faster than business controls. That gap is now one of the main risks in AI adoption.
The next generation of AI products will not simply generate better text. They will analyse, decide, connect, automate and act. That creates significant commercial opportunity, but it also creates new categories of operational, regulatory, data protection, cyber and accountability risk.
For SMEs, the answer is not to avoid AI or chase every new feature. The answer is to build governance that can keep pace.
AI adoption should be practical, controlled and evidence-based. Businesses that get this right will be able to use powerful new AI capabilities with confidence. Businesses that ignore the control gap may find that their AI usage has scaled long before their governance has caught up.