What an AI Copilot Is, and What It Is Not

An AI copilot is a customer support tool that reads the ticket, writes a draft reply, and then stops and waits for a person to check it, fix it and send it. It sends nothing on its own. That last sentence carries the whole of the difference between a copilot and an AI agent, and it goes missing from most of what gets written about either of them.

The blurring is nobody’s doing on purpose. Vendors say assists, and they say autonomous, and both of those words survive a demo comfortably, so a buyer comes away with an impression of degree, a sense that the one is a smaller version of the other. What actually sits between them is a procedural line about who signs the answer.

Worth saying before any of the rest of it. We sell a help desk and there is a copilot inside the dearest of our three plans, so we have a side in this and you should read the page accordingly. What we have tried to write is the version we would want handed to us, which means the part where it does not pay for its self is in here as well.

What an AI copilot does on a ticket, from landing to sent

Take one ticket arriving in the mailbox at ten past nine, Gmail or Outlook sitting behind the desk, makes no odds which of them. The copilot reads it, and it reads the thread underneath it, and generally whatever history that customer has with you that the desk can see from where it sits. Then it goes and writes a reply into the box, the whole of it, greeting and answer and sign-off, sitting there waiting on somebody.

Nothing has gone anywhere at that point. That is the part worth holding on to. A person reads what came out. Either it is right and they send it. Or it is close, and they fix the two things wrong with it and send that, and on most desks running one of these that covers the bulk of the traffic. Now and then it is nonsense and they bin the thing and write their own, and that happens less often than people walking into it expect.

Around the drafting sit a handful of smaller features doing the very same job, which is putting a first version of something in front of a person. A summary at the top of a thread that has run to 30 messages, so whoever picks it up on Thursday need not read the lot from the start. A translation, both directions. A rewrite of something written angry at half four. A suggested macro, where the desk already holds the answer and the copilot has gone and found it rather than making somebody remember it was in there.

What it cannot do is the finding out that happens away from the desk. Where the answer sits in Stripe, or on an order in Shopify, or on a record in your CRM, or with the engineer who knows why that integration broke on Tuesday and has the whole of it written up in Jira, the copilot writes you a well-mannered paragraph asking the customer to bear with you while somebody goes and checks. Useful, mind you. Not the thing anybody thinks they are buying. It cannot go and look.

What the difference is between an AI copilot and an AI agent

An AI agent does the very same reading and the very same drafting and then sends the thing by its self, without asking anybody first. That is the entire distinction. Not the model underneath it, not the cleverness, not the price on the plan, it is the sending.

Read the sending row and the failure row together, seeing as the two of them are one fact said twice. Everything else on that table falls out of where the send button sits, and a vendor page that spends its length on the model and skips over the row has answered a different question than the one you came in with.

Most products shipping both run them as two modes of the one thing rather than two purchases, ours included, and the seam between the modes is a confidence setting. Above the bar it sends. Below the bar the draft waits for a person, which is copilot behaviour arriving by another road. Where that bar belongs and who ought to own it is a question with a page of its own coming, and our own arrangement of it sits over on the AI support agent page.

Why the signature is the whole of the difference

Fair is fair, the capability underneath the two is frequently the very same model reading the very same ticket off the very same articles, and a vendor telling you otherwise is selling you a word. What changes between them is who owns the answer once it has gone out.

A copilot leaves that with the person who pressed send. They read it, they decided it was right, their name is on the reply, and where it turns out wrong there is somebody who can say what they were thinking at the time. An agent moves it somewhere else entirely, up to whoever switched the thing on and set the bar, and that decision got taken once, in a settings screen, in March, and then applied a few thousand times over without anybody going back to look at it.

The ITSM crowd went and had the very same argument about change approvals a good long while back, and it settled about where this one settles, which is the machine preparing the thing and a person putting their name under it.

For a good deal of what lands on a support desk that trade is worth making and we would say so plainly. Where the tracking number has got to, whether the invoice was issued, has the plan renewed. Nobody needs a human signature on any of that, and the customer would far rather have the answer at twenty past eleven at night than a person’s initials on it in the morning.

Then there is the other pile. A refund, an account closure, a discount somebody has asked for twice, a deletion request under GDPR, anything at all where the reply admits something or promises something. What the person adds there is not the writing, on account of the copilot having already done that. It is a decision about whether this is a thing your company wants to have said, which is a judgement about your own business rather than about the English, and no confidence score anywhere holds an opinion on that.

Where a copilot saves you time, and where it quietly does not

An ordinary ticket runs 11 minutes end to end, near enough, and the 11 wants breaking up honestly before anybody counts a saving off it. Call it 2 minutes reading the thread and working out what is being asked, which is regularly not the thing that got typed. Then 6 minutes finding out, meaning a search, a look at the account, a message to somebody in Slack or Teams who knows. The writing itself is 3 minutes.

A drafting copilot goes straight at the 3. It does nothing whatever to the 6, seeing as the answer is not anywhere it can read, and it takes a little off the 2 where it summarises. Now hand back 40 seconds for reading the draft properly against the account, and the honest saving comes out a bit under 2 minutes on an 11 minute ticket, call it 18%. Which is real and worth having. It is nothing like the half everybody has in their head walking into a demo, mind you, and the gap between those two figures is what the disappointment comes out of.

The shape changes entirely once the writing is the expensive part rather than the finding. A long explanation of a policy, given for the fifth time this week. A patient answer to somebody upset who deserves better than four lines. Anything that has to go out in a language nobody on the desk speaks. There the copilot is doing the actual job and the saving is large, no two ways about it, and the same goes for the summary on a thread that ran to 30 messages between the three of you and a customer who forwards things.

So the question to put to your own queue is a narrow one, and you can answer it in an afternoon with last month’s mail. Of the minutes your desk spends, how many go on composing and how many go on finding out. Where composing is a quarter of it, a drafting copilot buys you a slice of a quarter. Where composing is the most of it, go and buy the copilot. That is the whole of the test, funnily enough.

What to read on a draft before you send it

The review is the whole of the value and it is also the first thing to decay. Six things, and they take about 40 seconds between them once they have become habit.

1. Every number in the draft, checked against the account rather than against the draft. Dates, amounts, windows, plan names.
2. Any sentence stating policy. Ask where it came from and whether that is still policy this month, since the copilot is quoting something it read somewhere.
3. Anything written in the past tense about an action. “We have refunded you”, “your account has been upgraded”. The copilot cannot do the doing and will cheerfully say it is done.
4. The customer’s name and account details, on account of a draft built off a similar ticket carrying the other customer’s particulars more often than anybody expects.
5. The tone against the last message in the thread. A cheerful draft sitting under a furious message reads worse than no reply at all.
6. The question that was actually asked, read once more, against the answer that came back. This one catches the most and gets skipped the most.

None of that wants a process document or a training day. What it wants is somebody senior enough to say out loud, in about week four, that the reviewing has gone soft and everyone has started pressing send.

Why a bad copilot draft does not look like a bad answer

A bad canned reply announces its self. Wrong name at the top, a paragraph about a product the customer has never had, formatting out of another decade, and whoever is about to send it catches the thing in a second and a half, on account of it looking wrong before it has even been read.

A bad copilot draft looks correct. It is in your house voice, having read a thousand of your replies to learn how you talk. It runs the right length, it opens the way your desk opens, it apologises in the right places and the paragraph breaks fall where a person would put them. Underneath all of that it says your refund window is 30 days when you moved it to 14 back in the spring, and the only way anybody catches that is by knowing the answer already and reading closely enough to notice the draft disagrees with it.

That is the risk in the whole of this and it is not the one the market talks about. Nobody sends a draft that reads badly. People send drafts that read beautifully and are quietly wrong about a fact, and the better the writing gets the less likely a tired reviewer at four on a Friday is to stop over the number in the third paragraph.

The decay is the ordinary human sort and there is no shame in it anywhere. Week one, every draft gets read like a contract. Week three, the first eleven were perfect so the twelfth gets a glance. By month two the glance has quietly become the process, nobody wrote it down and nobody decided it, and the desk is running an autonomous agent while believing it has a copilot, truth be told.

Why the knowledge base underneath decides the ceiling

A copilot can only write out of what it can read. Your published help articles, your past replies, whatever notes sit on the ticket, and nothing else besides, so the quality of the drafts is a fact about your writing rather than a fact about the model.

Which lands two ways round. Where a desk has a real knowledge base with current answers in it, the drafts arrive close enough that the review is a genuine 40 seconds, and the thing pays for its self inside a month. Where the answers live in the heads of two long-serving people, or scattered through Google Workspace documents that nobody has linked to since, the drafts come out plausible and general and hollow, and every one of them wants rewriting, at which point the copilot has gone and added a step to the job.

Nobody selling one of these will tell you that the first month’s work is a writing job. It is, though. Take the questions your desk answers most, write the answers down properly, put them where the desk can read them, and the same effort feeds customer self-service at the same time, so the hours are not spent twice.

There is a stale-article trap sitting in the middle of it as well. A wrong article gets read by a person maybe once a quarter and read by the copilot on every single ticket it touches, day in, day out, which turns one bad paragraph from a slow leak into something that arrives in customer inboxes at scale.

Nine replies in a fortnight, and the sentence behind all of them

Say the refund window moved from 30 days to 14 in the spring. Somebody updated the terms page and the macro, both of them properly, and an old help article about returns kept the 30 in the third paragraph, on account of nobody having opened that article in a year and a half.

The copilot reads it every time a refund question lands. It drafts the 30 into the reply, in your voice, in the right place, and the reviewer sends it, seeing as it reads exactly like the answer they would have written their own selves. Over a fortnight that goes out 9 times. Nobody notices until a customer comes back on day 22 quoting the reply, and then it is a credit somebody has to approve and an apology somebody has to write.

Here is the check, and it costs about 10 minutes. Take a distinctive phrase out of the article you suspect, six or seven words of it that nobody would type twice by accident, and search your sent mail for it across the last month. Then do the same for the two or three articles that carry the numbers customers argue with, the windows and the fees and the eligibility. What comes back is a count of how far one stale paragraph travelled while everybody was doing their job properly.

The fix is not a better model and it is not a stricter review. Go and fix the article, on the grounds that it was the source, and then put the articles carrying numbers on a list somebody reads through twice a year. 10 minutes now and then, against the copilot repeating the very same wrong sentence with great confidence until somebody complains loudly enough. It will not tire of saying it.

Which numbers move when a copilot goes on, and which one to watch

Two things move quickly and neither of them settles the question. Composing minutes fall, which pulls average handle time down, the AHT on your dashboard, and that page sets out at length why the number falling is not on its own good news. The SLA clock is untouched by any of it, mind you, seeing as a draft waiting on a reviewer is a ticket still waiting. Replies also get longer and better mannered, which nobody expected and which reads well in a sample.

The number to watch beside it is the reopen rate, and beside that the count of second contacts arriving within 48 hours, seeing as a reply to a closed thread often lands as a fresh ticket and stays invisible to the first measure. Where handle time comes down and both of those hold steady, something genuinely improved. Where handle time comes down and the second contacts climb, the drafts are answering the question asked instead of the question meant, and CSAT usually says so a fortnight later.

What our own copilot does, and what it costs

Plainly said, up here rather than buried at the foot. Maxdesk puts the copilot on Elite, which is $99 per workspace per month, and there is no AI whatever on the other two plans. Free at $0 and Pro at $20 have the desk, the automations, the knowledge base and the reports, and not one AI feature between the pair of them, and it seems fairer to put that up here than to leave you meeting it at the checkout.

What Elite ships is 2 AI agents and a copilot sitting beside them. The triage agent sets a priority on arrival, files the ticket by issue, then assigns it by looking at who is holding how much open work at that minute. The resolver drafts replies and sends on its own only where its confidence is high, holding the draft for a person underneath the bar, which is the seam described further up this page. The copilot and response assistant is the drafting and rewriting on your side of it, along with ticket summaries, live translations, suggested macros, and sentiment and urgency detection. The allowance runs to 5,000 AI responses in a month across the whole of it, there is a usage dashboard so nobody finds out at the end of the month, and extra packs exist where a busy month runs over. The numbers behind the lot of it come out over the API or in a CSV export, the same as the rest of your ticket data, where you would rather count it your own self.

Two things about the money, since a page about AI features written by a vendor ought to say them. The bill goes by the workspace and never by the agent, so putting another two people on the reviewing changes nothing about what you pay, and we do not charge per resolution either, which is an argument we have made at length on our page about per-resolution pricing. The other one is that AI sits on our dearest plan, so every paragraph up above arguing that a copilot is worth having is also, plainly, a sales argument, and you should weigh it knowing that.

For the ordinary sorting and routing, no AI is wanted at all. Automation rules tag, route and acknowledge on every plan including the free one, and where a rule has never once got it wrong, there is not much of an argument for handing that particular job over to something which is right most of the time. Where the two of them differ and which to reach for first is the whole of our page on AI agent vs chatbot.

The parts of this we would argue with ourselves

The review cost is real and we have not measured it for you. Call it 40 seconds a ticket up above and it will be 40 seconds on some desks and 2 minutes on others, and where your answers want checking against a system the copilot cannot see, it may be that a copilot buys you very little at all. That figure is the one to pull out of your own first month rather than out of anybody’s page.

The ceiling belongs to your writing and not to us. Where the answers are not written down, the drafts will be general, and the honest recommendation is to spend the month writing articles before spending anything on a plan, seeing as that work pays whether or not you ever switch an AI feature on.

Mail is the only channel we ship, whether that mail comes out of Google Workspace or Microsoft 365 or anywhere else at all, so a copilot here drafts written replies and nothing besides. Somebody who rings you, or who walks in and asks, gets no help whatever from this.

Under about four people we would leave the whole of it alone. A desk of three already knows its answers and can write them faster than they can review a draft of them, and the $99 is better spent elsewhere for another year.

Then the data question, which every vendor should be asked and which we will answer for ourselves. An AI feature anywhere means the text of your tickets gets read by a model, ours included, so ask us and ask everybody else what is kept, for how long, and whether it trains anything. Where you work under HIPAA, or somebody is going to put a SOC 2 question to you about it in the spring, get the answers in writing rather than off a page like this one. Our retention runs 3 months on Free, 12 on Pro and 24 on Elite, and the free workspace also carries ads, our name on outbound mail, and only so many automations switched on at the one time. Those plan figures were right on 4 September 2026 and there is a re-check in the diary for February 2027.

Where to start, if you are weighing one up

Do not start with a demo, on account of every one of them being run on a queue that suits the tool. Start with 20 of your own tickets from last month, picked across the spread of what you actually get, and for each one write down where the minutes went, roughly, between reading and finding out and composing. That is an hour and it settles the question better than any trial does, since it tells you which of the three piles a drafting copilot would even touch.

Then take the three questions you answer most and write the answers down properly, in articles, with the numbers correct as of this week. Most desks turn something up in that hour worth fixing, and it tends not to be whatever they sat down expecting to find. If the composing pile turns out to be small and the articles turn out to be thin, you have saved yourself $99 a month and learned where the real work was hiding.

And if the composing pile is large, try the drafts against your own last month rather than reading another page about them. Ours is on Elite, the desk underneath it is free forever with unlimited agents, and the honest test of any of this is whether the drafts arrive close enough that the 40 seconds is genuinely all it takes.

Common questions about AI copilots

What is an AI copilot in customer service?
An AI copilot reads the ticket and writes a draft reply, then waits for a person to check it, correct it and send it. It never sends on its own. Alongside drafting, most copilots summarise long threads, translate, rewrite tone and suggest an existing macro where the desk already holds the answer.

What is the difference between an AI copilot and an AI agent?
The sending. A copilot drafts and a person sends, so the answer carries that person’s name. An AI agent drafts and sends by itself once it is confident enough, so accountability moves to whoever configured it. The model underneath is often identical; the difference is procedural rather than technical.

Does an AI copilot actually save time?
Less than most people expect, and it depends where your minutes go. On an 11 minute ticket split into 2 minutes reading, 6 finding out and 3 writing, a drafting copilot only touches the 3, and a real review hands about 40 seconds back. The saving is large where composing is the bulk of the work, such as long explanations or replies in another language.

Can an AI copilot send replies on its own?
No. A tool that sends on its own is an AI agent, whatever the marketing calls it. Many products ship both as two modes of one feature, with a confidence setting deciding which behaviour applies to a given ticket, so the question to ask a vendor is where the sending threshold sits and who owns it.

What does an AI copilot need to work well?
Current written answers it can read, and reviewers who keep reading. Draft quality is a fact about your knowledge base rather than about the model, so a stale article gets repeated on every ticket it touches. Check any number, policy claim or past-tense promise in a draft against the account before sending.

Does Maxdesk have an AI copilot?
Yes, on the Elite plan at $99 per workspace per month, which also carries 2 AI agents and up to 5,000 AI responses a month. The Free plan at $0 and Pro at $20 have no AI features at all. Email is the only channel Maxdesk supports, so the copilot drafts written replies only.