Scarlet Risk logoSCARLET RISK

← Blog

Five AI risks growing companies overlook

Scarlet Governance · Nov 10, 2026

Share

Growing companies adopt AI tools faster than they notice the new risks those tools create. The five that get overlooked most often are shadow AI (tools nobody approved), customer data leaving through a chat box, AI features that turn on inside software you already pay for, AI tools with standing access to email and files, and not being able to answer honestly when a customer or insurer asks what AI you use.

That is not a technology problem first. It is an ownership problem. Sales started using a summarizer last month. Operations turned on meeting recaps because the button appeared. Nobody updated the vendor list. Then a security questionnaire asks which AI tools touch customer data, and the room goes quiet.

We think about this a lot while building Scarlet, which ships AI features of its own. The tools are useful. The gap is that most growing companies treat AI like a personal productivity trick instead of a new place data can go.

1. Shadow AI: the tools nobody put on a list

Find out which AI tools are creating risk in your business. Get your free Risk Snapshot.

Get your free Risk Snapshot

The UK National Cyber Security Centre describes shadow AI as the use of AI that is not captured in an organization's approved systems and processes, a form of shadow IT. In a September 7, 2026 paper, NCSC notes that workplace policies have not always kept pace with how quickly people start using these tools, and that when official guidance lags behind, staff can end up turning to unapproved tools.

You cannot manage what you cannot name. A personal chatbot account, a browser extension that "writes emails for you," or a contractor's favorite summarizer can all touch customer information without ever appearing on an expense report.

If you do not already have a simple inventory, start with our guide to a shadow AI inventory for teams without a compliance hire.

2. Data leaving through the chat box

This is the risk people feel in their stomach after they hit send.

An illustrative example: a salesperson pastes a customer contract into a free chatbot and asks it to summarize the indemnity clause. A hiring manager pastes interview notes. An engineer pastes a config file that still has a password in it.

NCSC's shadow AI guidance is direct on this point. Employees who transfer sensitive or proprietary information to consumer AI services will likely reduce the organization's visibility and control over that information, because that information may be stored, retained, or used to improve the service unless specific privacy controls are in place.

In January 2024, CISA joined international partners, including the Australian Cyber Security Centre, on user-facing guidance called Engaging with Artificial Intelligence. That guidance lists privacy and intellectual property threats among the AI-related risks people should manage when they use these systems.

The practical rule is simple enough to write on one page: what data never goes into a chat box (customer files, employee records, credentials, unpublished financials, anything you would not put in an email to a stranger), and which approved tools, if any, are allowed to see the rest.

3. Vendor features that quietly turn on

A lot of AI risk does not look like a new purchase. It looks like a toggle.

The CRM you already pay for adds an assistant that can read your pipeline. The video meeting tool offers recaps, which usually means recording the call. The help desk begins drafting replies from past tickets. Nobody filed a vendor request, because the vendor was already on the list. The product just changed what it can see and what it can send out.

Treat a newly enabled AI feature like a change to an existing supplier, not like a harmless software update. Ask four questions before anyone leaves it on:

  1. What data can it read?
  2. Is that data used to train a model, and can you turn that off?
  3. Where is it stored, and who on the vendor's side can see it?
  4. Who in your company approved it, and when will you look at it again?

Our one-page vendor risk questionnaire is a lightweight way to capture those answers for digital suppliers, including the AI features hiding inside tools you already use.

4. AI tools with standing access to email, files, and calendars

Some AI tools do not wait for someone to paste text. They connect to your calendar, your inbox, or your shared drive and keep that access.

A note-taker joined to every meeting can see customer calls, pricing discussions, and personnel issues. A browser extension can read the page in front of you. An app that was approved once, so it could "help with email," may still have permission to read every message in the tenant.

This is the same class of problem as any vendor with standing access: the risk lives in the permission, not in the marketing page. NIST's AI Risk Management Framework, and its July 26, 2024 Generative AI Profile (NIST AI 600-1), ask organizations to identify the generative AI systems they use and manage the information security, data privacy, and intellectual property risks those systems create. You cannot do that if you do not know which tools already have keys to the building.

If you run Microsoft 365, the enterprise applications list in Entra is a practical place to start. Flag anything with mail or file access that nobody can name an owner for.

5. You cannot answer the AI questions on the next questionnaire

Security questionnaires and insurance applications are not only about firewalls and backups anymore. Some now ask which AI tools touch customer data, whether that data is used for training, and whether anyone has written down the rules.

If the honest answer is "we are not sure," the form stalls. MSPs answering these forms on behalf of clients run into the same gap. We wrote about how to handle client AI tools on security questionnaires, and about what to do when a customer sends you a security questionnaire.

Answer the same way you should answer the rest of the form: Yes (and you can show evidence), Partial (with what you do instead), or Not yet (with a date). A truthful "not yet" beats a confident "yes" you cannot explain on the follow-up call. Buyers who review these for a living can tell the difference, which is the gap we covered in what buyers ask versus what compliance dashboards show. MSPs answering these across many clients can reuse the same library approach in our MSP security questionnaire playbook.

Got a security questionnaire, or an AI question you cannot answer yet? See your gaps first. Get your free Risk Snapshot.

What to do this week

You do not need a new department. You need a list, a one-page rule, and an owner.

  1. Make a short inventory. Approved tools, personal accounts people admit to using, browser extensions, and AI features inside software you already pay for. For each one, note what data it can touch and who owns it.
  2. Write the chat-box rule. What never gets pasted. Which tools are approved. What happens when someone wants a new one.
  3. Review standing access. Note-takers, inbox add-ins, and anything with file permission. Turn off what nobody can explain.
  4. Draft your questionnaire answer before the next form arrives. If you use an MSP, ask them to help you map tools they manage versus tools the business adopted on its own.

CISA's user guidance also flags generative AI hallucinations: these systems can produce confident answers that are wrong. Do not let a chatbot be the only reviewer of a contract, a security control, or a questionnaire answer. A person who knows how the business actually works still has to sign off.

Where Scarlet fits

We built Scarlet Intelligence for growing companies that adopted the tools first and are only now being asked to account for them. Scarlet Governance keeps your vendor list and risk register in one place, and Scarlet Comply helps you keep policies and evidence close to what is actually true, so the next questionnaire does not start from a blank page.

Start here: Get your free Risk Snapshot. It stays free, takes about three minutes, and shows your biggest cyber, vendor, AI, and compliance gaps and what to fix first. If you would rather talk it through, grab 30 minutes with me.

Pete Macias, Founder, Scarlet Risk

Sources

Find out which AI tools are creating risk in your business. Get your free Risk Snapshot.

Get your free Risk Snapshot