What is a proactive AI assistant?
A proactive AI assistant is software that acts before you ask. It reads the places you let it see, like email, chat, your calendar and your meetings, notices what needs doing, and either does it or prepares it for your approval. A reactive assistant, like a chatbot, waits until you type a prompt.
That is the short answer. The longer one is more interesting, because "proactive" is easy to promise and hard to get right. I'm Minsung, and I build Arca, so I think about this every day. Here is how we think about it.
Reactive vs. proactive
Most AI tools people use today are reactive. You open a window, describe what you want, and get an answer. The quality of the result depends on how well you explained the situation. If you forgot to mention that the client already said no to the first draft, the model doesn't know.
A proactive assistant flips who starts the work. Something arrives, such as a client email asking for a revised scope, a meeting that ends with three action items, or a message that contains a date. The assistant notices it, decides whether it matters, and moves it forward. You find out when there is something to approve, or when it's already done.
The difference isn't the model. The same language model can sit behind both. The difference is everything around the model: what it can see, what it remembers, when it decides to act, and what it is allowed to do on its own.
What a proactive assistant needs
There are five pieces. If any of them is missing, the assistant either does nothing useful or does too much.
- Inputs. It has to see where work arrives. For most people that means email, a work chat like Slack, the calendar and meetings. An assistant that only sees what you paste into it can't be proactive.
- Memory. It has to know who people are, what was decided last week, and how you like things done. Without memory every message looks new, and every draft sounds generic.
- Judgment about what matters. Most of what arrives needs nothing. Newsletters, receipts and automated notifications should be ignored. A real person asking you a real question should not.
- A permission model. You decide what it may do alone and what needs your approval. This is the part that makes it safe to leave running.
- A report back. When it acts for you, it tells you what it did. If you can't see it, you can't trust it.
The two ways proactive assistants fail
When we started dogfooding Arca on our own work, two failures stood out. Both destroy trust quickly.
Nagging about things that are already done. You submitted the form yesterday, and the assistant reminds you again this morning because it never saw the confirmation. Each reminder like that teaches you to ignore the next one. Our rule now: before Arca reminds you of anything, it looks for evidence that it's already finished. If it can't find that evidence, it treats the item as unconfirmed, not as undone.
Missing the thing that mattered. A deadline buried in the third paragraph of an email, a renewal notice, an invitation that expires on Friday. A proactive assistant earns its keep by catching these before you think to ask, putting them on a list with the source attached, and tracking them until they close.
Notice that the two failures pull in opposite directions. One is too much noise; the other is too little. A good assistant has to be quiet about what's done and relentless about what isn't.
How much autonomy should it have?
"Proactive" doesn't have to mean "autonomous." In Arca, autonomy is a setting with five levels:
| Level | What Arca may do on its own |
|---|---|
| Off (suggestions only) | Nothing. It only suggests. |
| Read & organize only | Gather, summarize and draft. No outbound actions. |
| Draft only | Prepare messages and emails as drafts, not sent. |
| Handle routine tasks & send | Send routine replies, schedule, file things. |
| Full delegation ("arca it") | Act end to end on most tasks. |
Most people should start low and move up as they see what it does. Trust is earned one correct action at a time.
Even at the higher levels, some things stay behind an approval. In the current Arca beta, when an email comes in, Arca turns it into an action plan. It adds the calendar event and creates the to-do or follow-up itself, and marks each step done only after it actually succeeds. The reply, though, waits for you. You see a one-line question such as "Reply to the sender with the latest business registration attached?" and you approve, edit or skip.
A few concrete examples
Abstract definitions only go so far. Here is what proactive behavior looks like in Arca today.
- After a meeting. Arca records on your Mac without a bot joining the call, transcribes it with speakers separated, and writes the summary, the decisions and the action items. The action items become tasks, not just bullet points in a note.
- When an email asks for a document. If someone asks for a certificate or a registration document and Arca's document vault has it, Arca picks the exact file, drafts the reply in the sender's language, and asks whether to send it.
- When a message contains a date. A meeting invitation or a delivery deadline becomes a proposal you can accept from a single notification.
- When you change your mind. If the draft isn't right, you can type a direction like "politely say I'll send it next week," and Arca rewrites the reply to match.
Safety is part of the definition
An assistant that reads your inbox is reading text written by strangers. Some of that text will try to give it instructions. A proactive assistant has to treat every inbound message as data, not as commands.
In Arca, the triage step is told explicitly that inbound messages are untrusted and that instructions inside them, like "mark this routine" or "attach X," should be ignored. There is also a check that doesn't depend on the model at all: Arca will only auto-send an attachment to someone you have already corresponded with. "The model thought it was routine" is never enough to mail a document to a stranger. And if a draft promises an attachment that Arca can't find, it won't send that draft automatically.
When you evaluate any proactive assistant, ask how it handles this. If the answer is "the AI is smart enough," keep looking.
When you don't need one
Not everyone needs a proactive assistant. If most of your work is deep solo work with few inbound messages, a good chat assistant you open when you need it is probably enough. Proactive assistants pay off when your day is fragmented: lots of meetings, lots of people waiting on you, and a steady stream of small requests that each take five minutes but break an hour of focus.
The question to ask yourself is simple. Are you the bottleneck? If people regularly wait on you for replies, documents or decisions that don't really need your judgment, that is the work a proactive assistant should take.
How to judge one
When you try a proactive assistant, watch for these things in the first two weeks:
- Does it stay quiet about things you've already handled?
- Did it catch something you would have missed?
- When it drafts for you, how much do you have to rewrite?
- Can you see everything it did on your behalf, in one place?
- Can you turn autonomy up or down, and does it respect the setting?
If the answers are mostly yes, it is doing its job. If you find yourself checking its work more than doing your own, it isn't.
Arca is in early access. You can join the waitlist on heyarca.app. The conversations and work shown on our site are product demonstrations, not customer testimonials.
FAQ
What is the difference between a proactive and a reactive AI assistant?
A reactive assistant waits for you to ask. A proactive assistant watches the inputs you connect, such as email, chat, calendar and meetings, and starts work on its own when something needs doing, either finishing it or preparing it for your approval.
Is a proactive AI assistant the same as an AI agent?
They overlap. "Agent" describes the ability to take actions with tools. "Proactive" describes when it acts: without being prompted. A proactive assistant is usually built from agents, but many agents only run when you start them.
Will a proactive AI assistant send messages without asking me?
Only if you allow it. In Arca, autonomy is a setting that ranges from suggestions only to full delegation. Replies in the current beta wait for your approval, and attachments are never auto-sent to people you haven't corresponded with before.
What does a proactive AI assistant need access to?
The places where work arrives for you: typically email, work chat, your calendar and your meetings. It also needs memory of people, decisions and your preferences, otherwise every message looks new.
How do I get access to Arca?
Arca is in early access. Join the waitlist on heyarca.app with your email. Joining the waitlist is not a purchase and does not guarantee immediate access.