A canned reply is a saved answer that a person chooses to send, a typing shortcut with a human still doing the reading. A chatbot is a decision tree built in advance, a menu wearing a conversation, and the person who did the deciding is whoever built the tree, months before your customer ever turned up. An AI agent reads the actual message in front of it and composes an answer in the moment, so the deciding happens at the time, by the machine. Three tools, and the difference was never really the year each one was invented. It comes down to who does the deciding, and at what point they do it.
That distinction sounds like philosophy and is actually the most practical thing on this page, on account of it predicting exactly how each of the three goes wrong, and the going-wrong is what you are really shopping on. Anybody can show you the demo where the thing works nicely. What you actually needed was a look at the case it handles badly. No vendor runs that one for you, ourselves included in that.
We sell AI agents, to be plain about it early, and no chatbot at all, so you would expect this page to crown the newest of the three. It will not quite do that. Two of the three earn their keep on most desks, and one of them costs nothing and gets no marketing from anybody, funnily enough, which ought to make you curious about it.
The upgrade story the market tells, and the bit it skips
The way the three get sold is as generations. The canned reply is the old thing from the flip-phone years, the chatbot is the middle thing, the AI agent is the new thing, and the natural reading of any generation story is that each one replaces the last, the way nobody keeps a fax machine beside the scanner.
The story reads well on a slide and it is wrong in one load-bearing place. The three were never three attempts at the one job. They are three different jobs, and every desk we have ever looked inside still has all three kinds of work arriving in it: the question with one right answer that never changes, the bounded task a customer could do their own self if a form held their hand, and the message that has to be read by something, machine or person, before anybody can even say what it is about. Retiring the canned reply because the AI arrived is like retiring the kettle because the oven is new. Different work, and the whole of it still turns up every week.
So the honest question is never which generation to buy. It is which of your mail is which, and nobody selling any of the three has seen your mail. You will do that sorting yourself at the end of this page, in about twenty minutes, and it settles most of the argument before a demo gets booked.
The canned reply, where a person still does the reading
Strip it back and a canned reply is a human decision with the typing taken out. Somebody reads the mail, recognises it, and reaches for the saved answer instead of writing the same four sentences a ninth time. The judgement is exactly as good as the person holding it, which is to say as good as your best hire on her good days, and the cost of the whole arrangement is nothing beyond the writing of the texts.
It fails in small, human ways. The saved answer goes stale when the policy changes and nobody went and updated it, so a customer in March gets told the rules from October, politely and with confidence. Or the wrong one gets pasted at speed on a busy Friday, greeting a man named for somebody else entirely, and he screenshots it, and fair is fair, you earned that one. Staleness and haste, the pair of them, and both have the same fix, which is treating the saved answers as things with owners rather than things that got written once and left alone a good long while.
What the canned reply cannot do is scale past the people. Every send still costs a reading, so the queue moves at the speed of the humans, and that is the entire case for the other two tools, plainly said. It is also why this one gets no marketing: there is nothing to meter, nothing to subscribe to, and no vendor makes a shilling off you writing better saved answers. Which is exactly why it should be the first thing you fix, not the last.
The chatbot, where the deciding happened months ago
A chatbot is a menu that learned manners. Underneath the chat window sits a tree somebody went and built: if they say order, ask for the number, look it up, read the status back. Track a parcel, book a slot, reset a password, cancel a thing. For bounded tasks of that shape the tree is genuinely the right tool, quicker than mail, quicker than a person, no reading required because the builder did all the reading in advance, once, for everybody.
The characteristic failure is the case that is not in the tree. You have stood in it your own self, some wet Tuesday night, typing a perfectly clear sentence at a widget that answers with the very same four buttons it opened with, none of the four being your problem, and round you go again. The bot is not being stupid. It is being exactly what it is, a menu, and your case was not on the menu, and the conversation shape of the thing is what makes the moment so maddening, since a menu that looks like a menu never promised to understand you in the first place.
So the buying question for a chatbot is not how clever it sounds in the demo. It is how big the tree is against the true spread of your questions, and what the exit looks like from inside the loop. A bot that hands a stuck customer to a person, carrying the transcript along so nobody has to repeat themselves, is doing a decent job of the doorman work. A bot with no exit at all simply holds people until they give up on you. That damage lands in your customer satisfaction metrics long before the bot’s own numbers admit to a thing.
The AI agent, where the machine reads in the moment
The new one, and the one we sell, so the salt is on the table where you can see it.
An AI agent does the thing neither of the others attempts: it reads the message that actually arrived, the whole rambling three-paragraph version with the order number in the wrong place, works out what is being asked, and writes an answer to the very same, in the moment, with nobody’s tree and nobody’s paste involved. Ask it something no builder anticipated and it does not loop, on account of there being no tree to loop inside of. That is the genuine step forward and this page will not pretend otherwise.
The new failure mode is the price of it. A canned reply is wrong the way a sign is wrong, visibly, checkably. A chatbot is wrong by absence, your case missing from its menu. An AI agent can be wrong the way a confident colleague is wrong, fluently, plausibly, in complete sentences, about your actual situation, and a customer has no way of telling that answer from a right one until the parcel does not come. Nothing about the fluency signals the error, and that is a different animal to anything support has dealt with before, no two ways about it.
Which means the entire trustworthiness of the arrangement lives in one place: what the agent does when it is not sure. Not the average case, the boundary. Does it know it is unsure. Does it send anyway. Does it hold the draft for a person, does it route the hard thing to a human with its homework attached, does anybody ever see the ones it almost sent. A vendor who cannot show you that boundary, working, on an awkward example, is selling you the confident colleague and hoping you never meet him on a bad day.
One desk, three piles, all three tools
Whichever of the three you point at the queue, the queue its self does not change, and the honest split of most support mail has all three kinds sitting in it at once.
There is the pile with one right answer that never changes, where is my refund policy, do you ship to such a place, and that pile belongs to saved answers and to a decent customer self service page, on account of a question with one answer wanting writing once, properly, somewhere findable. There is the bounded-task pile, track it, book it, reset it, which is chatbot country in a widget world, and in a mail world it is mostly a lookup your helpdesk automation can route or a link your reply can carry. And there is the pile that needs reading, the ramblers, the edge cases, the angry, the ambiguous, and that pile splits again between the machine-readable middle of it and the genuinely human end of it, the escalation material, the judgement calls.
The proportions are yours and nobody else’s, mind you. A desk that is eighty per cent tracking questions has a different shopping list to a desk that is eighty per cent disputes, and the two desks reading the very same vendor page will come away needing opposite things, which is why the sorting exercise at the bottom of this page is worth more than the rest of the reading put together.
What our own version does, said in one paragraph
Maxdesk is a help desk that reads one shared address, the ticketing system puts states and owners on the mail, and the Elite plan at $99 a month on the workspace adds our AI customer support agent, which is two agents rather than the one. The triage one reads what arrives and sets the priority, files the category, hands the ticket to whoever is carrying least. The resolver one drafts the reply, and here is the boundary made visible: where its confidence is high it sends, and everywhere else the draft sits still. Somebody on your side reads it, approves it, or bins it. The held drafts are the part we would point at first, truth be told, on account of them being the evidence that the thing knows the edge of its own competence. The sending is the half that demos well.
The three honest caveats on our side of it
The AI comes only at $99, bundled with the rest of Elite, and there is no buying the resolver on its own, so if the agent is the whole of what you want you are paying for the bundle to get at it, which is the exact tiering habit we grumble about in other people’s pricing. Said here so you do not go and find it out at the checkout.
There is an allowance on it as well. Elite includes five thousand AI responses a month and paid packs past that, meaning the agent is metered where the rest of our bill is flat, and a metered line deserves to be read as a cost line by whoever signs off on it.
Then the channel, which is the awkwardest of the lot on a page like this one. Mail is the only door we watch. No widget, no phone, no portal. The chatbot section above concerns a world we do not sell into at all, and we wrote it anyway because the comparison is the search you typed, and an honest three-way beats a two-way with the awkward leg missing. If your questions arrive at half nine at night on a pricing page, that is widget work, and you should go and buy a widget from somebody who makes a good one.
Twenty minutes with last month’s mail, before anybody’s demo
Open last month in whatever holds your record, your email management software or the desk its self, and go and deal the most recent hundred conversations into the three piles by hand. Same answer every time. Bounded task. Needed reading. Rough is fine, and the whole of it goes quicker than it sounds, a few seconds each once your eye is in.
Then look at the sizes. A fat first pile means your cheapest win is saved answers with named owners and a knowledge base page per repeat offender, and that project costs a weekend and no invoice. A fat second pile in a mail-first shop mostly wants better links and lookups before it wants anybody’s bot. And a fat third pile is the one worth pricing an AI agent against, ours or anybody’s, armed now with the only vendor question that separates the field: show me, on a hard example, exactly what it does when it is not sure. Watch the screen at that moment rather than listening to the pitch over it. Whoever you buy from, that boundary behaviour is the thing you live with every day after.