AI assistant: what it is, which kind you need, and what it costs

Virtual Minds
A business owner at a busy desk looks uncertain while exploring an AI assistant on his laptop.

An AI assistant is software that takes an instruction in ordinary language and does something useful with it — drafts the email, summarises the sixty-page contract, pulls the numbers out of the spreadsheet, books the slot. Underneath it is a large language model wrapped in three other things: a memory of your context, a set of tools it is allowed to call, and rules about what it may do without asking you first. The model is the engine. The wrapper is what decides whether the assistant is a toy or part of your working day.

The short answers to the questions most people arrive with: yes, there are genuinely free personal AI assistants (ChatGPT, Claude, Gemini, Copilot and Perplexity all run usable free tiers, with caps on volume and on the best models); a personal paid plan sits in the low tens of dollars per user per month across the major vendors; and a business assistant that touches your CRM, inbox and documents costs considerably more, because the cost is in the integration and the governance, not the chat window. Below, the four types, how to pick, and how to tell a real one from a demo.

What AI assistants actually are

Strip the branding and every assistant is the same four parts.

  • A model. The reasoning engine. Most assistants run on a frontier model from OpenAI, Anthropic, Google or Meta — often several, routed by task.
  • Context. What it knows about you: the current thread, uploaded files, your calendar, your product catalogue, prior conversations. Retrieval decides what gets pulled in.
  • Tools. The actions it can take. Search the web, run code, query a database, send a message, create a ticket. An assistant with no tools can only talk.
  • Guardrails. What it may do unsupervised, what needs a human approval gate, what it must never touch. This is the part buyers skip and then regret.

The difference between a chatbot and an assistant is tools plus guardrails. A chatbot answers. An assistant acts, and something sensible happens when it is wrong.

The four kinds of AI assistant

1. General-purpose chat assistants

ChatGPT, Claude, Gemini, Copilot, Perplexity. Broad, strong at writing, reasoning, summarising and code. They know nothing about your business until you tell them, and they forget unless you turn on memory or connect your files. Best first move for almost everyone: pick one, use it daily for a month, and only look further when you hit a wall you can name.

2. Device and home assistants

Siri, Alexa, Google Assistant and the newer screen-and-speaker devices. Optimised for hands-free, short commands: timers, lights, music, quick facts. They are improving as LLMs replace the old intent-matching, but they remain weaker at long reasoning and anything multi-step. Judge them on latency and reliability, not cleverness.

3. Embedded assistants inside software you already pay for

The assistant inside your spreadsheet, design tool, IDE or helpdesk. These win on context — they can already see the document — and they lose on breadth. A design assistant that understands your room dimensions will beat a general chat model at that one task, which is the logic behind narrow products like AI design assistants.

4. Operational assistants and agents

These run work rather than answering about it: qualify an inbound lead, chase a document, reconcile a list, draft and route an approval. They need access to your systems, a record of every action, and a human gate on anything irreversible. This is where most business value sits and where most deployments fail — usually because nobody defined what the assistant does when it is unsure. The architecture questions matter more than the model choice.

Free, paid, or built: what you get at each level

TierWhat you getWhere it stops
Free consumerDaily chat, file uploads, web search, voice, limited use of the strongest modelsMessage caps, no business data controls, limited or no integrations
Paid personalHigher limits, best models, memory, connectors to drive and mail, image and code toolsStill per-person. Nothing shared, nothing audited, no team workflow
Team / business seatsAdmin controls, data excluded from training, shared assistants, SSOGeneric. It knows your documents, not your process
Custom buildAssistants wired into your CRM, billing and inbox, with approval gates and loggingNeeds a real integration budget and an owner inside the business

On price: personal plans from the major vendors sit in the low tens of dollars per user per month, and business seats are typically a step above that. Custom work is quoted on scope — number of integrations, volume, and how much human review you want in the loop. If you are sizing a build, our breakdown of what AI lead generation costs shows how those line items actually assemble.

A small-business employee types a plain-language prompt into an AI assistant to pull data from a spreadsheet.
A simple typed instruction is how most AI assistants receive and act on your work.

How to choose an AI assistant without testing twelve of them

  1. Write down three real tasks. Not "help me be productive." "Turn Monday's call notes into a follow-up email." "Find the renewal clause in this contract." "Draft a reply to this complaint in our tone."
  2. Run the same three tasks through two assistants. Same input, same day. Differences show up in ten minutes that no review article will tell you.
  3. Check where your data goes. Consumer free tiers are the wrong place for client files. Business tiers normally exclude your data from training — confirm it in writing before you upload anything sensitive.
  4. Test it on something it should refuse. Ask for a figure it cannot know. A good assistant says it does not have the number. A bad one invents one, and now you have a liability.
  5. Decide what it may do alone. Drafting, yes. Sending to a client, probably not yet. Put the gate in before the first mistake, not after.

For the deeper evaluation criteria — context windows, tool use, memory, evaluation harnesses — see our guide to what makes an assistant advanced.

Where AI assistants still fail

Three failure modes recur across every deployment we have shipped. First, confident wrong answers on facts the model cannot verify — mitigated by retrieval against your own documents and by citing sources, never by asking the model to try harder. Second, silent drift: the assistant works for a quarter, the underlying model updates, and the output changes shape with no alert. You need evaluations running on a schedule, not a one-off accuracy check at launch. Third, adoption. An assistant that lives in a separate tab gets used twice. One that answers inside the inbox or the CRM where work already happens gets used daily.

There is also a quieter cost: review time. If every output needs a careful human read, you have moved the work rather than removed it. The assistants that earn their seat are the ones where you stop checking after two weeks on a specific, bounded task — and keep checking on the rest.

Business assistants: what we build and what we have learned

Virtual Minds has run consumer AI apps past 500,000 users — Room AI, Reshot AI and others — all on one internal platform, Cortex, which we describe as the AI operating system our products run on. On the business side, LeadsMind is an operational assistant rather than a chat window: it researches accounts, drafts the outreach, and routes replies to a human before anything commits. We run our own outbound on it, which is the only endorsement worth much.

The pattern that holds across both: the model is rarely the constraint. The constraint is clean context, a short list of tools, and a clear rule about when a person steps in. Teams that get those three right do well with almost any frontier model. Teams that get them wrong buy a better model and stay stuck.

If you are weighing whether to assemble this internally or have it run for you, we wrote up the trade-off in in-house AI team vs operated AI system. If you just want a good personal assistant and nothing more, start free, give it a month of real work, and upgrade only when the limits annoy you.

  • ai-assistant
  • ai-agents
  • productivity
  • buying-guide
  • cortex