AI Safety for SMEs: How to keep your business safe in the agentic era
If you’ve been keeping up with AI for any amount of time now, you’ll know that AI agents are no longer just chatbots. They can take actions, send emails, write and deploy code, access other tools and systems, all while making their own decisions.
While modern AI agents have been a massive boost for business productivity, something that isn’t talked about as much as it should be is AI agent safety and security. How do you keep your business secure when you have AI agents handling tasks for you end-to-end? Because the real risk to businesses doesn’t come from AI when it stops working, but it comes from when AI is misused or manipulated, which can lead to sensitive information being leaked, unauthorised access, and actions being taken in your business by AI that you would never take in a million years.
And the part that shocks me about AI security is that most SMEs aren’t even talking about it. From most of the conversations that I’ve had, most SMEs are focused solely on the capabilities of AI and disregard cost and safety. If you’re running an SME, this post is your wake-up call.
A brief overview of AI security issues in business
Prompt injection
Attackers can trick AI models into ignoring the original system prompt or guidelines, which allows them to hijack what the AI does and what information is exposed. They do this using ‘jailbreak’ prompts, which use several different methods to make the AI ignore its original prompt.
Excessive agency
If your AI connects to different systems in your business, it’s likely that it has more permissions in those systems and apps than it really needs. This means that even without an outside attack, the AI has the option to delete files, read private emails, or make changes to the business that it really shouldn’t be able to make. The worst part of excessive agency is that you won’t find out about this until it bites you and the AI has decided to delete all your customer records, just for example.
Sensitive information leaking
AI agents often have a memory, and because of this, they can inadvertently leak personally identifiable information, which is a huge GDPR no-no, or proprietary business information. These things can enter the AI’s memory subtly just by employees using the system, or the AI having access to the company’s online workspaces like Microsoft Teams or Google Drive.
Employee undisclosed AI utilisation
Employees using public AI tools can unintentionally paste internal business information into the tools, which can then be used to retrain these AI models, which could leak this sensitive information. This is why it’s important to know which tools your employees are using and what those tools’ policies on input data are.
A mental model for secure AI in your business
When planning your AI implementation in your business, there is a mental model that it’s important to have in mind. That’s the principle of least privilege.
To put it simply, that means not giving the AI any more permission or access than it actually needs to perform its task. You can do this by checking what apps and systems the AI is connected to. Normally, it will use an API key or MCP server, which will give you fine-grained control over what that API key can be used to do, so, for example, you could set it to only read documents and not be able to delete them.
You have to think about it like this: if the AI turned against me today, what could it do? If the answer is “a lot of damage” then your permissions are too broad. You should apply this thinking every time you set up a new AI tool or integrate AI with your existing platforms.
Building a sensible AI safety stack
Human in the loop
Whenever you have AI making decisions in your business, especially if they’re big decisions, you should have a human approve them. This can be done via “human in the loop”. But here’s the risk: people are lazy, and what the AI says looks right, so the employee just approves it without really looking at it. It’s very important with humans in the loop that you don’t just rubber-stamp the AI’s decision, but actually take the time to understand the implications.
Human-AI SOPs
And that’s where human-AI SOPs come in. If you can train your employees not to just rubber-stamp decisions, feed the AI sensitive data, or abuse the use case of the AI, then you can create a safe and secure AI environment, which will ensure you don’t run into any issues.
Audit logs
Keeping internal AI audit logs, as in what the AI is thinking internally, is super important because if something does go wrong, you can pinpoint the exact time, date, and reason that it happened. Unfortunately in my experience, though I have never seen an SME do this before hiring me.
AI monitors
If you want to get really automated, you can use AI-AI monitors. These are other AI models that will monitor your model’s decision-making, perform sentiment analysis on outgoing emails, and escalate anything it thinks is an edge case to humans. You can think of these as your AI line managers.
Don’t build it yourself
It’s very easy to make an AI security mistake when you are connecting systems or building AI agents yourself. I would highly recomend either using a professional app end-to-end or hiring an AI expert to implement the AI into your business and connect the systems for you. With how quickly AI can make a business-changing decision these days, it’s not something to take lightly.
What kind of data should your AI touch?
If you’re not working with customer, financial, or proprietary business info, then you don’t really need to worry about how you use AI in your business, but if you do, then it’s important to think about how AI is using your data, and even more importantly, where your data is going.
GDPR compliance can become very complicated when you start sending sensitive data to cloud-based AI models for processing, and most AI platforms these days retain the right to train their models on your input data. This is a huge GDPR no-no. It’s just worth thinking about where your data is going, and how sensitive it is.
For sensitive data, I’d recomend running an AI model locally, which can be done on a relatively inexpensive Mac mini or AI unit. This way, all your sensitive data stays on-premises, which can reduce the compliance complexity significantly.
For the most sensitive data, for example, health records, consider whether AI should even touch this at all. We as humans are still learning the capabilities of AI and discovering new features every day.
One thing you can do today
If you’re already using AI in your business and you’re not sure if it’s secure, bringing in an AI consultant for a day can be a genuinely underrated move. One day of going over your AI infrastructure could save you weeks, if not months, of headache down the line. If, for example, you had an AI data leak.
I would recommend you ask them to audit the tools you’re using, the connections between them, including the privileges granted to the AI, and where the data you are using in AI is actually going. Hiring an AI consultant is inexpensive relative to the possible security, compliance, and brand reputation risks you’re carrying if something goes wrong.
The agentic era for SMEs is here, whether your business is ready or not, and AI safety isn’t about avoiding AI but using it with the same diligence you’d apply to any business system, tool or employee. The businesses that get this right now will have a real advantage over those that leave AI security up to chance.
Need an AI consultant? Hire me.
