Glossary

Software that uses an AI model to carry out a task in steps: it reads, decides what to do next, and uses your tools to do it.

An agent can take on work a person now does by hand, such as sorting requests or updating records. Its permissions, set in advance, decide what it may do alone and what waits for a person.

A company designed around AI agents: they do a large part of its work, and its people direct them and answer for the results.

An application programming interface: the way software talks to a system, to ask it for data or tell it to do something.

When a tool has an API, other software can work with it without anyone copying data by hand. It is the first thing to check before connecting a tool to another, or to an AI agent.

A record of who did what in a system, and when.

It answers “who changed this?” after the fact, for people and AI agents alike. Customers and auditors can ask to see one.

Computers and storage rented from a provider such as Amazon, Google, Microsoft or Cloudflare, and used over the internet.

You pay for what you use, and the provider keeps the machines running. The major providers let you choose the region where your data is stored, and some clients ask about it.

Customer relationship management: the tool that holds your customers, prospects and every exchange with them, such as HubSpot or Salesforce.

When the team keeps it up to date, anyone can see where a deal stands without asking. It is also where an AI agent finds a customer’s history before it writes to them.

A screen that shows a business’s key figures at a glance, updated from its data.

A good dashboard replaces the weekly report someone builds by hand in a spreadsheet. It is worth knowing where its figures come from before trusting them.

Keeping the same information identical across two or more tools automatically, so it is typed once.

Without it, the same order or customer is entered twice and the copies drift apart. A sync also decides which tool is the reference when two copies differ.

One place that gathers a company’s data from all its tools for analysis and reporting, such as BigQuery or Snowflake.

It answers questions that cross tools, such as sales against stock against margin, that no single tool can answer alone.

Where software stores its information in an organised way, so it can be found, sorted and updated quickly.

A spreadsheet used as a database starts to fail when several people edit it at once and one wrong paste can overwrite everyone’s work.

Electronic data interchange: a standard way for companies to send each other orders, invoices and delivery notes directly between their systems.

Some large retailers and distributors require it from their suppliers. It replaces the order that arrives by email and is typed in again on the other side.

A way for software to compare texts by meaning, so it finds the right passage even when the wording differs.

They let an AI agent search your documents by meaning rather than exact keywords. RAG usually relies on them.

Enterprise resource planning: the central software that runs orders, stock, purchasing and accounting in one place, such as SAP, Odoo or Microsoft Dynamics.

In a firm that has one, it holds the figures everyone else relies on. Connecting other tools and AI agents to it ends typing the same data twice.

The General Data Protection Regulation: the European law on how organisations collect, store and use personal data.

It applies to firms that handle personal data about people in the EU, customers and employees included. Before an AI tool reads personal data, you need to know where that data goes and who can see it.

When an AI model states something false with confidence, because it writes the most likely-sounding answer rather than a checked fact.

It is why an agent should answer from your own documents and cite them, and why an important output is checked before it goes out.

Everything around an AI agent that sets how it works: what it may do alone, what waits for a person’s yes, what it knows, how its work is checked and what is logged.

The same AI model can work in any company. The harness is what makes it work the way yours does: the rules you would give a new hire, written down so they are followed every time.

A connection that lets two tools exchange data or trigger actions in each other automatically.

When a firm runs on several tools, people copy data from one to the next. Each integration removes one of those steps.

Large language model: the kind of AI model behind ChatGPT or Claude, trained on large amounts of text to read and write.

On its own, it knows only what it learned in training. Connected to your tools and documents, it can answer and act on your company’s own information.

Building software with visual tools and ready-made blocks, and writing code where the blocks stop, on platforms such as Retool.

It suits the internal tools a team needs quickly, such as an approval screen or a stock view.

The Model Context Protocol: the open standard that lets any agent, Claude, ChatGPT or one you built, use a company’s tools.

With an MCP server, a tool works with whichever assistant your team prefers, without a separate connection for each.

Building software by assembling visual blocks, on platforms such as Airtable or Zapier.

A team can set up simple tools and automations itself. It reaches its limits as the logic, the volume or the security needs grow.

The standard behind “Sign in with Google” buttons: it lets one app use your account in another without ever seeing your password.

When an AI tool asks to connect to your email or files, this is usually how: you approve what it may access, and you can withdraw that access at any time.

Optical character recognition: software that reads the text in a scan or a photo, such as an invoice or a delivery note, and turns it into data.

It ends retyping paper documents. Paired with an AI model, it can also tell which figure is the total and which is the VAT.

The rules that decide who, or which software, can see, change or delete what in a system.

For an AI agent, a simple set is: what it may read, what it may edit, and what needs a person’s yes first.

The instruction or question you write to an AI model.

Context, an example and the format you expect get better answers than a one-line request. A prompt a team reuses often can become a skill.

Retrieval-augmented generation: the agent first searches your files and records, then answers from what it found, citing it.

It keeps an assistant’s answers on your own procedures and product sheets rather than general knowledge, which cuts down hallucinations, and every answer can be checked against its source.

Software as a service: software you rent by subscription and use in a browser, such as Slack, HubSpot or Notion.

There is nothing to install and updates arrive on their own. Your data lives with the vendor, so how to get it out, by export or API, is worth checking before you sign.

A login a tool gives to other software rather than to a person. Whoever uses it sees everything it sees.

AI agents and integrations usually run on one. It should get only the access its job needs, a named owner, and the same review as any other account.

A packaged set of instructions, sometimes with files, that teaches an AI assistant to do one task the way your company does it.

It turns know-how that lives in one person’s head into steps every colleague’s assistant follows the same way.

Single sign-on: one company login that opens all your work tools.

When someone leaves, turning off one account closes their access everywhere. Some tools offer it only on their higher plans, so it is worth checking before you buy.