Learn
Help me understand.
Plain explanations of how AI actually enters an organisation — agents, MCP, permissions, state and governance.
AI Agents
Managing multiple AI agents
Once an organisation uses several AI tools and agents, it needs one inventory of what is connected, consistent rules for what each may do, shared context so work is not duplicated, and a single record of what they did. Managing each tool separately does not scale.
What is an AI agent?
An AI agent is an AI system that can take steps towards a goal, choosing and using tools such as search, email or a database, rather than only replying with text. Because agents act, businesses need to decide what each one may see, do and do on whose behalf.
AI Fundamentals
AI Governance
How do I audit AI actions?
Record every request an AI makes to your systems at the point where it reaches them: who it acted for, which agent, what resource, what action, which rule decided it and what happened. Chat histories alone are not an audit trail.
What is AI observability?
AI observability is the ability to see how AI is actually being used across an organisation: which agents are connected, which systems they reach, what they did, what was approved or refused, and where AI adds value or risk. It is about organisational insight, not surveillance.
What is an AI control plane?
An AI control plane is the layer where an organisation decides and enforces how AI may use its systems: which agents are connected, what each may see and do, which actions need approval, and what gets recorded. It separates those decisions from the AI tools themselves.
What is human-in-the-loop approval?
Human-in-the-loop approval means an AI can prepare an action, such as an email, payment or publication, but a named person must approve it before it happens. It keeps accountability with people while letting AI do the preparation.
AI Memory
Can one AI share context with another?
Not natively. Each AI tool keeps its own conversations and memory. To share context between Claude, ChatGPT, Copilot and others, the context has to live somewhere independent that every authorised AI can read from and write to.
What is organisational AI memory?
Organisational AI memory is shared, governed knowledge of an organisation’s current work (decisions, status, facts and artifacts) that any authorised AI can use. Unlike a personal chat memory, it belongs to the organisation, carries provenance and respects permissions.
AI Security
AI access to confidential data
AI can safely work with confidential data only if access is enforced outside the AI: classify the data, limit each agent to what the task and person need, keep restricted categories out by default, check provider terms, and record every access.
How do I revoke AI access?
List every place an AI holds access, then revoke at the source: remove app consents and tokens, rotate any shared keys, disable the agent’s identity and check the audit trail. Revocation stops future access; it cannot recall what an AI has already retrieved.
Staff sharing client data with AI
When staff paste client information into AI tools, the firm can lose control of confidential data. The fix is a clear rule, an approved AI route that respects client boundaries, and a way to give AI the client context it needs without copying files into chats.
Business AI
How do I give ChatGPT access to company data?
Decide which data ChatGPT needs, connect it through a connector rather than uploads, scope access to each person’s permissions, keep sensitive systems out at first, and make sure every retrieval is recorded. Start with one use case, not the whole business.
Employees using ChatGPT at work
Employees are likely already using ChatGPT and other AI tools, often with personal accounts. Banning it rarely works. A better response is a clear policy, an approved account, rules for what data may be used, and a governed way to connect AI to company information.
Delegation
External AI Collaboration
B2B AI agent collaboration
B2B AI agent collaboration is one organisation’s AI working with another organisation’s information or agents, for example a client’s AI requesting documents from its accountant. It works when each side exposes only defined information and capabilities through a governed relationship.
Can clients use their own AI with your data?
They can, safely, if you share a defined portion of the information through an access layer that decides what their AI may retrieve. Sending files or opening your workspace to a client’s AI gives up that control; sharing a governed projection keeps it.
Identity
How do I use Microsoft Entra ID with AI agents?
Use Entra ID to sign people in to AI tools and gateways with single sign-on, control which apps may be consented to, manage who has access through groups, and revoke centrally. Entra establishes identity; you still need rules for what each agent may do on a person’s behalf.
What is AI agent identity?
AI agent identity means an agent is recognised as a distinct actor, separate from the person it works for, so systems can tell which agent made a request, apply rules to that agent specifically, record its actions and switch it off without affecting anyone else.
Industry Guides
Integrations
Can AI access a SQL database?
Yes. AI tools can query SQL databases through a connector or MCP server. The safe pattern is read-only access through a restricted account, limited to specific tables or views, with queries logged, rather than handing an AI a full database login.
Can ChatGPT access SharePoint?
Yes, ChatGPT can work with SharePoint content through a connector that authenticates to Microsoft 365 on someone’s behalf. The questions that matter are whose permissions it uses, which sites and libraries are in scope, and whether anyone can see what it read.
Can Claude access Xero?
Yes — Claude can work with Xero data through a connector, typically using the Model Context Protocol. The important questions are who holds the Xero credentials, what Claude may read or change, and whether a person approves financial actions.
Can Claude send Outlook emails?
Technically, yes. With a connector to Microsoft 365, Claude can draft and even send Outlook email. Most organisations should let AI create drafts freely and require a person to approve anything that actually leaves the building.
MCP
How do I secure an MCP server?
Authenticate every client, keep upstream credentials on the server, expose narrow tools rather than raw access, validate inputs, treat retrieved content as untrusted, require approval for consequential actions, and log every call. MCP defines the connection, not the security policy.
What is MCP?
MCP, the Model Context Protocol, is an open standard that lets AI applications such as Claude connect to external tools and data through "MCP servers". It standardises how an AI discovers and calls capabilities; it does not by itself decide what the AI should be allowed to do.
Permissions
Publishing
Can AI publish to your website?
Yes. AI can draft content, update pages and trigger deployments if it is connected to your content system and hosting. The safe pattern separates drafting from publishing, and requires a person to approve before anything goes live.
How do I publish company information for AI?
Keep your company facts in one structured, approved source; publish them as clear, server-rendered pages with consistent terminology and structured data; add machine-readable feeds or an API where useful; and keep everything current and dated.
Structured company data for AI
Structured company data holds facts about your organisation (clients, services, projects, decisions, locations) as defined records rather than prose in documents. AI can retrieve, filter and act on records reliably, and access can be controlled field by field.
What is an AI-readable website?
An AI-readable website presents its information so that AI systems can find it, understand it and describe it accurately: server-rendered pages, clear definitions, consistent terms, semantic structure, structured data, stable URLs and current, dated content.