The basics

What is AI and how does it work?

The simple version

AI is a non-deterministic (randomised) predictive engine - think autocorrect, but trained on almost everything ever written. It predicts the most helpful next word, sentence, or idea based on your input and its training data. It doesn't "think" or "understand" - it pattern-matches at enormous scale.

Why it can get things wrong

AI is trained to be helpful and to always produce an answer. When it doesn't know something for certain, it may confidently make something up. This is called hallucination - it's not lying, it's over-eager. Always verify important facts from primary sources.

Data frozen in time

AI models have a training cutoff date - they don't automatically know what happened recently. Many can use tools like web search or database queries to fetch up-to-date information when needed, but check whether your tool supports this.

Key terms to know

GPT Generative Pre-trained Transformer - the architecture behind most large language models. Not a brand; a type of AI technology.

Token How AI measures and charges for text - roughly three-quarters of a word. Millions of tokens are available for a few pounds per month. A paid subscription beats pay-as-you-go for predictable costs and access to the best models.

Model The specific trained AI system you're talking to (e.g. GPT-4o, Claude Sonnet, Gemini Pro). Different models have different strengths, costs, and data cutoffs.

Prompt Your input - the question or instruction you give the AI. Better prompts produce better results. Be specific, give context, and describe what a good answer looks like.

"Treat AI like a brilliant but reckless intern - genuinely capable, enthusiastic, and fast. But always needs a human to review the output before it goes anywhere."

Who's who

The key players and their tools

Anthropic Claude (Haiku, Sonnet, Opus), Claude Code - known for safety and nuanced reasoning. Our top recommendation for most business tasks and writing.

OpenAI ChatGPT, Sora (video), Codex (code) - the original household name. GPT-4o is fast and capable with a wide ecosystem of integrations.

Google Gemini (1.5, 2.0), NotebookLM - excellent for long documents and research. Deep integration with Google Workspace and Gmail.

Microsoft Copilot (powered by OpenAI) - built into Microsoft 365. Best starting point if your team already uses Teams, Word, Excel, or Outlook.

Meta Llama 3 - open-source models you can download and run yourself. No subscription required, but technical setup needed. Can be trained on your own data.

DeepSeek (China) DeepSeek R1, V3 - powerful and open-source, strong for technical tasks. Data privacy considerations apply - review before using with business data.

Beyond the big names

There are thousands of AI tools worldwide - image generators (Midjourney, DALL-E, Adobe Firefly), voice tools, coding assistants, and specialist business tools. Many are built on top of the models above. Open-source models can be trained on your own data and run entirely within your own infrastructure.

Quick task-to-tool reference

Making a website or appBolt.new or Lovable

Excel or Google SheetsCopilot (MS) or Gemini

Summarising or researchingChatGPT (quick) or NotebookLM (deep)

Writing docs or emailsChatGPT or Claude

Marketing visuals & videoMidjourney, Kling, or Veo

Coding or technical buildingClaude

Getting started

What we recommend - and building safely

1

Experiment - especially leaders. AI has to be seen as OK to try, fail, and learn. If leadership isn't experimenting, the rest of the team won't either.

2

Give everyone a paid licence to ChatGPT or Claude. Better models, reliable data access, fixed monthly cost. Avoid pay-as-you-go surprises and unlock the best possible tools.

3

Find AI ambassadors - willing and able learners who bring ideas back to the wider team. Show and tells, shared prompt libraries, and regular demos all help build momentum.

Generally safe to build

Landing pages with no data captured or stored
Internal read-only tools (reports, project organiser)
Using AI to generate content, images, emails
Summarising public documents or meeting notes

Proceed with care

Sending data via AI - is only the right data shared?
Tools that create or update records - can you trust it?
API keys - treat like passwords, secure properly
Recruitment or assessment tools (EU AI Act - Aug 2026)

Always remember

Don't assume AI is correct - ask what it's doing AND what it's not doing
Always keep a human in the loop for anything that matters
Run security scans over anything you build or commission
Watch for AI slop - is this really good enough to show customers?
You build it, you own it. Delegating the task doesn't delegate the responsibility.

Print it. Share it. Use it.

Your AI policy template

Establish what is permitted

Be explicit about which tools are approved and in which contexts. Approved consumer AI tools for individual productivity; stricter controls for anything touching client data. Let AI ambassadors explore more freely - they are your learning edge.

Use with care

AI tools are not private. Never share passwords, API keys, private client data, or proprietary code. All AI output must be critically reviewed for accuracy, context, bias, and appropriateness before use or publication.

Confidentiality first

If unsure whether something is safe to share, don't share it. Treat all external AI tools as public spaces - even paid or "private" accounts. No external system is as secure as simply not sharing sensitive information.

Beware of integrations

Connecting apps together can expose data in unexpected ways. APIs may share more than expected. Always review what data is being shared when setting up integrations, and seek approval if client or sensitive data is involved.

Do

Use dummy data when testing AI tools
Discuss ideas, not confidential specifics
Use secure tools for passwords and API keys
Ask your team leader before using AI with client data
Keep a human review step before publishing

Don't

Paste API keys, passwords, or private data into AI tools
Upload client materials without explicit permission
Assume a paid account guarantees full privacy
Enable integrations without understanding their data reach
Treat AI output as fact without checking the source

If sensitive data may have been exposed, report it immediately to your team leader - acting fast limits damage.
Full guide: glo-ai-guide.netlify.app/index-v3.html