Right now, someone in your office is probably using an AI tool you didn't know was being used. AI is being adopted into corporate spaces more every day, but every unapproved and unvetted AI tool is an open door for mistakes, mishandled data, and malicious behavior. Shadow AI is something to pay attention to and fix even if your office has not adopted AI and doesn't plan to. Your company may depend on it.
If your company has started to use AI tools, make an official AI policy as soon as you can and review any tools before you start using them.
Here's how to vet an AI tool before it becomes part of your workflow, with security and data privacy prioritized.
Is Shadow AI Already in Your Office?
Shadow AI is the use of AI tools at work without your IT team's knowledge or approval. This may not seem like a big deal, but it means that any company data put into that tool is at risk. It's similar to someone emailing client files to their personal email. If you don't know where the company data is going or how it's being used, that's a problem.
The use of shadow AI is already widespread. In 2024, Microsoft and LinkedIn found that 78% of AI users were bringing their own AI tools to work. Beyond security concerns, shadow AI use also puts your money and reputation at risk. In IBM's 2025 study of 600 organizations that experienced a breach, one in five reported a breach involving shadow AI. Those with high levels of shadow AI saw an average of $670,000 more in breach costs. Customer personally identifiable information (PII) was exposed in 65% of breaches involving shadow AI, compared to 53% of breaches overall.
The creep of unknown AI tools happens quietly and comes in many forms. One person might use a free chatbot to rewrite an email. Someone else might decide to use a browser extension to summarize PDFs. A coworker might use an AI note-taker during meetings and calls without disclosing its usage. Your company may not choose to adopt AI, but shadow AI means it's being used whether you know it or not.
If you're unsure about what AI is being used in your organization, SpireTech can help. We run AI audits as part of our AI Services. Book an IT consultation with us to see if we'd be a good fit to work together.
How to Vet AI Tools While Prioritizing Security
How does an organization adopt AI tools safely? Before anyone on your team uses a new AI tool with work data, it needs to pass these questions. If there isn't a clear answer, err on the side of safety and find another solution.
Data privacy and security
What happens to the data put into the AI system?
This is the most important question on the list. Find out whether your inputs are used to train or improve the vendor's models, and whether that's the default or a setting someone needs to turn off. Find out whether your data is shared with third parties, including the company behind the underlying AI model, how long it's retained, and whether it's actually deleted when you ask or cancel.
Model training and data sharing are why we so strongly recommend paid AI tools for business use. Business plans generally don't use your data to train models by default, and they give admins more control over access and data retention. If you already use Microsoft 365, Copilot has the added benefit of working within your existing Microsoft 365 security and compliance settings. Even with paid tools, always check the privacy settings and terms rather than assuming your data is protected.
Is the data encrypted?
Data should be encrypted both in transit (while it's being sent) and at rest (while it's stored). This is a baseline expectation, and any reputable vendor should be able to confirm it quickly.
Where is the data stored, and who else handles it?
Ask which countries your data is stored in and which third parties (subprocessors) handle it. Many AI tools are built on another company's model, so your data may pass through more than one vendor. Reputable vendors publish a subprocessor list.
Can the vendor's employees see your data?
Some vendors allow staff to review conversations for quality or safety. Find out whether that happens, under what circumstances, and whether you can opt out.
Is there documentation for the security controls used?
Security controls are the specific safeguards a vendor uses to protect data, such as multifactor authentication, encryption, access reviews, logging, and backups. Frameworks like the current CIS Controls lay out which safeguards an organization should have. SpireTech's cybersecurity is designed to follow CIS IG1.
When vetting a vendor, look for proof that an independent auditor has checked their controls. The two you'll see most often are SOC 2 Type II, an audit report showing the vendor's controls worked over a period of time (usually six to 12 months), and ISO 27001, a certification for how the company manages information security.
You don't need to understand the technical details behind these, though it certainly helps. What matters is that an outside expert has reviewed how the vendor protects your data and confirmed those safeguards actually work. Ask for the report or certificate, and check that it covers the AI product and plan you're using, not just another part of the company.
Can we control who has access?
Look for single sign-on (SSO) with your existing accounts, multifactor authentication, and admin controls. IT should be able to approve or block the tool, manage who has access, and remove access automatically when someone leaves, rather than relying on each employee to manage their own account.
What permissions does the tool ask for?
Don't assume that a tool asks for the minimum amount of permissions it needs to function. As always, follow the principle of least privilege: give the tool only the access it needs to do its job. (Apply the same thinking to the permissions your employees get, as well.) For example, an extension that can "read and change all your data on all websites" has the potential to see your banking portal, management system, and everything else. Ask yourself whether the access requested matches what the tool actually does. Browser extensions, meeting bots, and email plug-ins are the sneakiest offenders.
What can the tool access and do on your behalf?
Many AI tools now connect to email, calendars, SharePoint, and other systems, and some can take actions like sending messages or editing files. The more a tool can reach, the more damage a mistake can do. There's also a newer risk called prompt injection: an attacker hides instructions in an email, document, or web page, and the AI follows them when it reads that content. Ask three things: what can it read, what can it do, and what happens if an attacker controls something it reads? Only connect what the tool actually needs, and require a person to approve actions like sending, sharing, or deleting.
Model behavior and reliability
How was the model trained, and on what data? What is the underlying AI model?
Training data shapes what an AI model can and can't do, and it also carries legal risk. A model trained on copyrighted content, biased datasets, or unverified sources can create problems for you, not just the vendor.
If the tool is built on someone else's model, ask which one. Whether a tool uses Claude, GPT, or another model may be a deciding factor, since your data may also pass through that model provider and be covered by their terms.
Does the tool provide audit logs?
Audit logs show what the AI did, when, who asked it, and what data it touched. They matter for compliance, for incident response, and for figuring out what went wrong after the fact. You may not ever need them, but it's far better to have them and not need them than need them and not have them.
Compliance and legal exposure
Does the AI tool follow the regulations to which you need to comply?
Depending on your industry and clients, your organization may be bound by HIPAA, GDPR, CCPA, state privacy laws, or professional confidentiality rules. Your vendor needs to meet the same standards. Ask for documentation. For anything touching health information, ask whether they'll sign a business associate agreement (BAA).
What is your liability if something goes wrong?
Many AI vendor contracts include broad indemnification clauses that push liability back onto the customer. Read the fine print. Ask what protections exist if their tool violates privacy laws, produces discriminatory output, or causes an operational failure.
Vendor stability and support
What's the incident response plan?
Breaches happen. What matters is how fast the vendor detects one, how they respond, and how quickly they inform you. Ask for a written incident response plan and specific notification timelines, plus what assistance they'll provide if your data is involved.
What happens to your data if you leave?
Don't assume that your data isn't valuable to the vendor. Are you able to export your data in a usable format? How long does the vendor keep it after you cancel? Is deletion verified and documented?
How does the tool handle updates and model changes?
AI models change, and sometimes in ways that break your workflows or quietly change the tool's behavior. Ask how often the model is updated or retrained, what data is used, and whether you'll be notified before major changes roll out.
Build an Approval Process (and Make It Easy to Follow)
Deciding not to use AI may work for your team. It also has the potential for employees to just be quieter about their AI use. Either way, give your team a clear framework, even if it's just an AI-use agreement that spells out what is and isn't allowed.
An approval process for AI tools might look like:
- Publish an approved tools list. Make it easy to find and keep it current. Our recommended AI tools gives you three good options to consider.
- Create a request process. A short form where staff name the tool and what they want to use it for.
- Define what data is absolutely off-limits. Client PII, financial records, privileged communications, and unreleased designs should never go into an unapproved tool. Data privacy needs to remain a major consideration.
- Favor tools inside your existing security boundaries. If you're already dialed into Microsoft 365, Copilot with a business license works within your tenant's permissions and compliance settings. It makes it easier to adopt, change, and oversee. Keep in mind that Copilot can see anything the user already has access to, so clean up overshared files and SharePoint and OneDrive permissions before rolling it out. (SpireTech can help with that cleanup.)
- Audit regularly. Even if your organization has AI governance policies, you need to perform regular audits for unsanctioned AI. Sign-in logs, browser extension inventories, and network monitoring can show you what's actually in use.
How to Get Started
If you'd like help, SpireTech provides Portland businesses with AI services and cybersecurity. We'll begin with an AI audit, inventory the AI tools already in your environment, review vendor terms with you, and set up controls so your team gets the productivity boost without the data exposure.
Schedule a meeting to review your AI tools and build an approval process that works for your business.
Frequently Asked Questions
Are free AI tools okay to use for work?
It depends on the task. Low-risk work, like brainstorming a headline or rewording a generic email, is generally fine. Anything involving client data is not and should absolutely be avoided.
Free tools often use your inputs to train their model. If you do approve a free tool, set a clear rule about what kind of data can and cannot go into it.
What should I do if someone already put client data into an unapproved AI tool?
First, find out what data was shared, which tool it went into, and when. Review the tool's settings and terms to see whether the data can be deleted or excluded from training, and request deletion if possible. Involve whoever manages your compliance obligations because some client contracts and regulations require you to report data exposure. If you're not sure where to start, contact your IT provider.