What is an AI assistant?

An AI assistant is software that uses a large language model to do work on your behalf: it understands requests in ordinary language, holds context about you or your business, and produces useful output — answers, drafts, summaries, schedules, research — without being programmed for each specific task.

The defining feature is delegation. A search engine gives you materials to work with; an AI assistant does a portion of the work itself. The practical question when evaluating one is not 'how smart is it?' but 'what can I hand it, and how much checking does its output need?'

The market uses these words loosely, which suits vendors and confuses buyers. Working definitions that hold up:

• Chatbot — answers questions in a conversation window, usually from a fixed script or knowledge base. Reactive; no memory of you; no tasks.
• AI assistant — converses and does work: remembers context, drafts, summarises, researches, reminds. Works with you, at your instruction.
• AI agent — an assistant given a goal and the ability to take multi-step actions with less supervision: it plans, executes, checks, retries.
• Copilot — an assistant embedded in a specific tool (code editor, spreadsheet, CRM), scoped to that tool's work.
• AI employee — vendor language for an agent packaged around a role ('AI sales rep'). Sometimes real, sometimes a chatbot in a suit; evaluate the autonomy and permissions, not the job title.

Reliably, today: answer questions from documents they've been given (with sources, if well-built); draft and revise emails and documents; summarise long material; do structured web research; extract data from messy input; remind and follow up; triage and route incoming messages; schedule against calendars.

Unreliably, or only with careful engineering: multi-step autonomous actions in business systems; anything requiring precise arithmetic at scale; judgment calls with legal or safety consequences; work where a confident-sounding wrong answer is costly and hard to catch.

The honest general rule: assistants are strongest where output is checkable and the cost of a miss is low — and they get dangerous exactly where output looks authoritative but isn't checked.

Four things, none of them the model: memory (does it actually know your context, durably?), grounding (does it answer from verified facts and cite them, or improvise?), permissions (is what it may do explicit and auditable?), and escalation (does it know what it doesn't know, and hand off gracefully?).

When you compare assistants — ours or anyone's — those four questions expose the difference between a tool and a demo faster than any feature list.