The Agentic Post
Breaking
Perplexity vs ChatGPT Search vs Google AI Mode: Which AI Search Is Best in 2026?  Â·  What Are Vision-Language-Action (VLA) Models? How Robots Learn to Follow Instructions  Â·  How to Translate Documents With AI Accurately (and When Not To)  Â·  MCP vs A2A: The Two AI Agent Protocols Explained in Plain English  Â·  Is Grok Accurate for News? What Studies Show About AI Search Citations  Â·  Is AI a Bubble? 7 Warning Signs Investors Are Watching in 2026  ·  
Home/AI Safety
What Is Shadow AI? How to Manage Employees Using Unapproved AI Tools

What Is Shadow AI? How to Manage Employees Using Unapproved AI Tools

AI Safety

Shadow AI is staff using AI tools the company has not approved. The risks, why banning it fails, and a practical 6-step plan to bring it into the open safely.

Shadow AI is when employees use AI tools their company has not approved, like pasting customer data into a personal chatbot account or installing an AI browser extension. It happens in almost every organisation because the tools are useful and easy to access. Banning them rarely works; the better approach is to offer approved tools, set clear rules and make it easy for staff to ask.

Why is shadow AI a risk?

  • Data leaks: confidential or personal data may be stored or used for training by a provider you have no contract with
  • Compliance: privacy laws and client contracts may forbid sharing certain data with third parties
  • Security: unvetted extensions, plugins and agents can be malicious, as in the fake Claude Code extension case
  • Wrong answers: unchecked AI output can end up in reports, code or client emails
  • No audit trail: if something goes wrong, nobody knows what was shared

Why does banning AI tools not work?

Because people use them to get work done faster, and they can use personal phones and accounts. A ban pushes use further out of sight. It also signals that the company is not helping staff learn a skill the job market increasingly expects, as covered in AI literacy as a job skill.

How do I manage shadow AI in 6 steps?

  • 1. Find out what people use. Run an anonymous survey and check network and expense data, without punishing honesty
  • 2. Offer an approved option. A business AI plan with data protections removes the main reason for shadow use
  • 3. Write a short acceptable use policy saying which data can and cannot go into AI tools; see our acceptable use policy guide
  • 4. Create a fast approval path for new tools, measured in days, not months
  • 5. Train people on safe prompts, privacy and checking output
  • 6. Review regularly, since tools and risks change every few months

What data should never go into an unapproved AI tool?

Customer and employee personal data, passwords and API keys, unreleased financial results, source code under client contracts, health or legal records, and anything marked confidential. When in doubt, ask or remove identifying details first.

What about AI agents and browser extensions?

These are a bigger risk than chatbots because they can act, not just answer. An agent with access to email or files can be tricked by hidden instructions on a web page. Require approval for any tool that connects to company systems, and see our agent autonomy risk guide.

Who should own the AI policy?

Usually a small group from IT or security, legal and a business team, so rules are safe but practical. For the wider rollout, read AI governance frameworks and common enterprise AI rollout mistakes.

Up Next
What Is Quantization in AI? How Models Shrink to Run on Your Laptop

What Is Quantization in AI? How Models Shrink to Run on Your Laptop

AI Glossary

Quantization explained simply: how AI models store numbers with less precision to cut memory and cost, what 8-bit and 4-bit mean, GGUF files, and the quality trade-off.

Quantization is a way to shrink an AI model by storing its numbers with less precision, for example using 4 bits per number instead of 16. A 4-bit model takes roughly a quarter of the memory, runs faster and costs less, usually with only a small drop in quality. It is the main reason you can run capable open models on an ordinary laptop or phone.

How does quantization work, in simple terms?

An AI model is billions of numbers called parameters. Normally each is stored very precisely, like writing a price as 3.14159265 dollars. Quantization rounds it to something like 3.14. You lose a tiny bit of detail on each number, but across billions of them the model behaves almost the same while taking far less space.

What do 16-bit, 8-bit and 4-bit mean?

  • 16-bit (FP16 or BF16): the usual full-quality format; about 2 bytes per parameter
  • 8-bit (INT8 or FP8): about half the memory, with almost no quality loss for most tasks
  • 4-bit (INT4): about a quarter of the memory; the popular sweet spot for running models at home
  • Below 4-bit: very small files, but quality drops noticeably

Rough rule: a 7-billion-parameter model needs about 14 GB at 16-bit, 7 GB at 8-bit and around 4 GB at 4-bit, plus extra memory while it runs.

What is a GGUF file?

GGUF is a popular file format for quantized models, used by tools like llama.cpp, Ollama and LM Studio. When you download a model you will often see names like Q4_K_M or Q8_0; the number is roughly the bits per parameter. Higher is better quality and bigger. Our guide to running open-source models locally shows how to pick one.

Does quantization make AI models worse?

A little, and it depends on the task. Casual chat and summaries rarely suffer at 8-bit or good 4-bit. Precise tasks like maths, coding and long reasoning are more sensitive, so they benefit from higher precision. Bigger models usually handle quantization better than small ones, which is why a 4-bit large model often beats a full-precision small one.

Why do AI companies use quantization too?

To cut inference costs. Serving models at lower precision means more users per chip. New chips include fast support for 8-bit and even 4-bit maths for this reason. Combined with designs like mixture of experts, it is how giant models get affordable, as in our colibri guide where trillion-parameter models run in 4-bit.

Which quantization should I choose?

  • 8 GB of RAM: a small model (3 to 8 billion parameters) at 4-bit
  • 16 GB: up to roughly 14 billion parameters at 4-bit, or small models at 8-bit
  • 32 GB or more: mid-size models at 4-bit or 8-bit

Start with a Q4_K_M or similar, and step up if answers feel sloppy. Find more terms in our AI glossary.