An AI email client lands at the layer where your working day happens. So the first thing most firms do after buying one is go looking for what it replaces. That instinct is a fair one, because a new per-seat line needs justifying. Start somewhere else. Most of a small firm’s stack does work an assistance layer leaves alone, and a small, nameable part of it overlaps. The quiet pieces holding the firm up are the ones most at risk in the enthusiasm of week two, because a line nobody defends in a meeting is an easy line to cut.
This guide takes the keep side of that decision. What survives the addition, what honestly stops earning its line, and the test that sorts the rest before anything gets canceled.
- An assistance layer keeps no record of its own. That’s what makes it cheap to add and cheap to leave. It’s also why the products that do hold a record stay.
- Grade every line by what it would take with it if it stopped. That sort takes about twenty minutes, and it lands somewhere different from a sort by price.
- The subscriptions that genuinely stop earning are few and specific. In most firms it’s one, occasionally two. If the purchase rests on that saving alone, the arithmetic is thin, and it’s worth saying so out loud.
- The product nobody opens duplicates nothing and breaks nothing. It needs a different question. Ask what it was bought for and where that job sits now.
- The things worth keeping hardest never had a subscription. Your retention decision. Your arrangement for covering each other. And if the Safeguards Rule reaches you, a security program that belongs to the firm rather than to any vendor.
What a layer on the mailbox cannot hold
Start here, because half the keep decisions fall out of it.
An AI email client signs in to the mailbox your firm already licenses, and works on the mail that’s already there. An assistance layer isn’t a mail provider, so everything underneath it stays put. Where email fits in a small firm’s stack sets out the four separate things a firm buys under the single word “email.” An assistance layer is the last of those four. The domain stays where it was bought, and so do the seats, the retention tier and the archive.
One thing follows from that, and it’s worth being plain about. A product that leaves your mail with your provider leaves everything else there too. Your client record lives where it lives. So does your document store, and so does the matter file. What this layer does is read correspondence and act on it. That’s a real job and a narrow one. A product at this layer offering to be your filing system as well is describing a second copy of your record, and a second copy is a liability rather than a saving.
If your firm handles clients’ financial information, a rule sits behind this, and it’s why a tool change stays a tool change. 16 CFR 314.3(a) requires you to “develop, implement, and maintain a comprehensive information security program that is written in one or more readily accessible parts and contains administrative, technical, and physical safeguards that are appropriate to your size and complexity, the nature and scope of your activities, and the sensitivity of any customer information at issue.” The program belongs to your firm. Software is a component you describe inside it. Adding a tool changes the description, and the obligation stays with you. That holds for every purchase at this layer. You also take on a separate duty for each vendor, under a different provision, and the stack overview works through 16 CFR 314.4(f) against the rule’s own words.
So the honest starting position is a short one. Adding an email client cancels nothing. Things get canceled because the exercise made you look at the list.
Grade every line by what it would take with it
Go down the subscription list and ask one question of each product. If we stopped paying at the end of the month, what leaves with it? That grades a stack faster than a feature comparison does. It also grades it in the order that matters, because each grade carries a different cost when you get it wrong.
It holds the record, and nothing else does. The client file. The matter history. The books. The signed documents. The place your firm goes to answer what was agreed in March. These stay. Price becomes a fair question once you have a specific plan for where the record goes instead. Until then, a cheaper product in this grade is a migration rather than a saving. Migrations get charged once, at somebody’s real hourly rate, against a saving that recurs.
It holds a copy. A tool that logs, mirrors or syncs something that also lives somewhere else. Reporting layers, note-takers, anything that receives a feed. These are genuinely negotiable. The test is whether the copy does work the original can’t. Sometimes it does, and then you keep it and you can say what the copy is for. Often it’s a habit from a period when the original was harder to search than it is now.
It holds nothing you couldn’t rebuild. Its state is a preference, a layout, a set of shortcuts. Cutting it costs an afternoon of annoyance. This grade is where most of the money that can safely move actually sits.
An assistance layer sits in the third grade, and it’s worth knowing why, because the same reasoning works on every product you’ll ever grade this way. Ask whether a product writes into a store you already own, or into one the vendor owns. Gmail labels and Outlook categories are stores you own. Google’s developer guide describes a label plainly. “You can use labels to tag, organize, and categorize messages and threads in Gmail. A label has a many-to-many relationship with messages and threads” (Gmail API labels, checked September 7, 2026). Microsoft’s Graph reference draws the Outlook equivalent the same way. “The user defines categories in a master list, and can apply one or more of these user-defined categories to an item,” and the name of one is “a unique name that identifies a category in the user’s mailbox” (outlookCategory, checked September 7, 2026). A product that puts its judgment there leaves its work inside a subscription you already pay for.
That cuts both ways, and the second way is the one to weigh before you sign. The product keeps some things on its own side. The ranking it learned, the summaries, the task list, the corrections you fed it. That’s the vendor’s store, and it leaves when you do. The personal side of that departure is worked through in how to switch to an AI email client. The stack-level version is a question you put to the vendor before you connect, rather than at the point of canceling. Which of this sits in my mailbox, and which of it sits in your database? A straight answer is a good sign about the rest of the relationship. The wider list of questions worth putting to anybody holding client correspondence is in what to ask before you connect a tool to client data.
One warning rides along with the same test. Some products pull the correspondence into their own system, which moves them from the third grade to the first quietly, before anybody has decided to move them. Which tools in your stack actually reach the mailbox sets out the five shapes of that. Read it when you’re unsure which of yours are which.
The lines that genuinely stop earning
Being specific here helps more than being generous, so here’s the short and honest list.
Scheduling is the real overlap, and it’s the one with a number attached. If your firm pays per seat for a booking product, that line and an assistance layer running the back-and-forth itself cover overlapping ground. Calendly’s pricing page lists a Free plan holding “one event type” and one calendar connection, Standard at “$10/seat/mo” and Teams at “$16/seat/mo” (Calendly pricing, checked September 7, 2026). Six people on the Teams plan is $96 a month, or $1,152 a year. That’s a real line to hold up against a new one, rather than a feature to enjoy on top of it. Compare honestly in both directions. A dedicated booking product does things a general email layer leaves alone, and what a booking link should do marks out where the separate link is still the better instrument. The move to avoid is paying for both because nobody ever asked.
The follow-up scaffolding usually empties out on its own. The send-later add-on. The reminder plug-in. The tracker that tells you nobody replied. The personal task app that exists mostly for things that arrived by email. These are small amounts of money, and several are free, which is exactly why they persist. Watch them rather than canceling them on day one. The tell is a month going by with nobody opening one. Email task management without writing it down and where later actually goes are the two guides that describe what replaces them.
Some of what you stop using was free all along, and it’s still worth stopping. Two of these already run inside your platform. The first is whatever split your provider does on its own. Microsoft’s Focused Inbox splits the inbox in two, so that “your most important email messages are on the Focused tab” and the rest sit on an Other tab, using “emails and contacts with whom you interact” and filtering “noisy sources like automatically generated or bulk email” (Focused Inbox for Outlook, checked September 7, 2026). Its coverage has always had limits, and Microsoft says so plainly. “If you’re using a shared mailbox, Focused Inbox will not work. This is by design.” So when two things sort one inbox, pick the one your people read by. Then the answer to “did you see it” is the same for everybody.
The second is the wall of rules somebody built over four years. It has a ceiling most firms never discover. Microsoft allows a mailbox 256 KB for its enabled rules in total, within a valid range of 32 KB to 256 KB. Microsoft is blunt about the direction that number moves. You “can’t increase the amount of space that’s allocated to store Inbox rules in Exchange Online, but you can decrease it” (Inbox rules quota, checked September 7, 2026). Spend that budget on the rules that route a message to a place a person chose. The rules that were somebody’s attempt to guess importance are the ones doing a job twice. Which of yours is which is a question of evidence rather than tidiness, and how to switch to an AI email client has the mechanics of pruning them.
And here’s what keeps earning, whatever the sales page implies. The books. The client record. The document store. Practice management. Payroll. Signing. The password manager. The shared address that three people watch. And if your inbound genuinely behaves like a queue with assignment and response times attached, the help desk built for that. Each of those is a job that happens outside the mailbox, and the nine jobs a small firm covers in software is the frame that keeps the distinction straight.
Count the result before you build the business case. For most firms of five to ten people, the honest total of what leaves is one subscription and a handful of free add-ons. That’s a fine outcome. The reason to buy is the hour, which never appeared on an invoice in the first place, and what an AI email client costs is where that arithmetic is done properly.
The product nobody opens
Somewhere on your list is a subscription that stands on its own. It duplicates nothing, it connects to nothing, and everything keeps working if it stops tomorrow. Nobody has opened it in months. Cutting overlapping subscriptions works through the four shapes of genuine overlap, and the checks that stop a cancellation taking something with it. This one falls outside all of them. It passes every safety test by being inert. That’s exactly why it survives audit after audit. It’s too quiet to force a decision.
So ask a different question. Ask this one: the problem this was bought for, is it still real, and where is it being solved now? There are four answers, and each one tells you what to do.
The problem went away. The client it was for left, the service line closed, the reporting requirement changed. Cancel it. That’s the cleanest cut available in any stack, and it takes a minute.
Somebody solved it another way and kept quiet. A spreadsheet, a different product, a habit. Find out which. You may be paying for the abandoned answer and the working one at the same time, and the working one may itself be a subscription you’ve yet to look at.
It’s being solved by hand, in the inbox. This is the interesting answer. It’s also why the question lands better the week you add an assistance layer than it does in January. Think of the product bought to chase documents, track who owes what, or remind somebody about a deadline. Using it took more discipline than the problem was worth, so it sat there. The job went into somebody’s head, carried between messages. That’s the job the layer you just added is pointed at. Cancel the shelf product, then watch whether the work gets picked up. That last part is the one to verify.
The problem was never real. Somebody bought it in a week when it seemed prudent. Cancel it. Note who bought it and why, because the pattern is worth interrupting more than the charge is.
One more thing turns up in this pass, and it deserves naming. A product nobody opens is sometimes a product exactly one person opens, quietly, for something they never mentioned. That’s a reason to ask. The time to ask is before the renewal date rather than after the cancellation.
What you keep that never had a subscription
The most valuable things in a small firm’s operating stack sit off the card statement. They’re also the ones an enthusiastic adoption is most likely to disturb.
The retention decision. How long the firm keeps client correspondence, decided deliberately and set on the platform. An assistance layer leaves it exactly where it is, so keep it in plain sight. If you were on the wrong tier for holds last month you’re on the wrong tier this month, and where email fits in a small firm’s stack is where that tier boundary is worked out on both platforms.
The arrangement for covering each other. Whatever your firm does so a client waiting on an answer gets one while the person who owns the relationship is out for a week. That’s a practice rather than a product, and it slips easily once the software appears to be watching the inbox. Software watching an inbox is one thing. A colleague picking something up is another. Covering for each other is the version of this written for the smallest firms.
The offboarding list. The document naming every product somebody has a seat on, produced from the inventory rather than from memory. Add a product, add a line to the list the same afternoon. A stack grows one forgotten seat at a time, and the mechanics of keeping that list current are in cutting overlapping subscriptions.
The convention about where things get written down. If your firm’s rule is that anything agreed with a client ends up in the matter file, that rule survives the arrival of summaries and extracted tasks. Those are working notes, and the matter file is still the record. Say so once, out loud, at the point of adoption. The failure here is a quiet one. A summary gets treated as the file for six months, and the firm finds out when somebody asks for the file.
The person who answers. An assistance layer changes who does the sorting. Accountability for the reply stays where it was, and every judgment that was a person’s judgment last month still is. Say that plainly to a team, because the other belief spreads on its own.
Keep paying until a busy week has been through it
Cancel nothing in the first month. That’s the whole of the timing rule, and it’s cheap. One extra month of one duplicate is the least expensive insurance in this exercise.
The evidence you need arrives later than the enthusiasm does. A quiet week tells you about a quiet week, and in a quiet week everything looks optional. Run the overlap through one full cycle of whatever your firm’s busy actually is. A month-end, a filing deadline, a closing, the week the returns go out. Then look at two things. Whether anybody went back to the old product during it, and whether the job it covered got done. Those two answers settle it together, and both of them arrive well after day three.
Two mechanical notes make the waiting cheaper. On the new purchase, take a monthly term for the first stretch if the vendor offers one. Then an annual discount, taken before you knew anything, stays out of the overlap decision. On the old one, the cancel decision has a date rather than a moment, and on an annual term that date is the renewal. The order that keeps what you already paid for is set out in cutting overlapping subscriptions, including the difference between canceling and letting a term lapse. Read it before you touch anything.
Then put the review in the calendar for ninety days out, with the shortlist of candidates on it. The saving that goes missing is rarely the one somebody argued against. It’s the one everybody agreed to and nobody scheduled.
Running the test on Point
Point is an AI email client. An article about keeping should run Point through the same three grades as everything else, so here it is.
Point sits in the third grade, and it sits there on purpose. You connect the Gmail or Microsoft 365 mailbox your firm already licenses. The domain, the seats, the tier, the retention rules and the archive stay as they are, and your provider is still the place the mail lives. What Point decides gets written back into the mailbox itself. Connect Gmail and that judgment arrives as real Gmail labels. Connect Outlook and the same thing arrives as categories. Put a thread away and it’s archived in the mailbox, the same as any archive you do yourself. That’s what makes Point cheap to leave. The labels, the categories, the archived threads, the drafts you sent and the calendar entries Point created are ordinary items in an account you own. The layer is what leaves with Point. The ranking, the summaries, the task list and the corrections Point learned from you. That’s the honest cost of stopping, and it’s worth pricing before you start. It’s the same question this guide says to put to any vendor.
What Point plausibly takes off your bill is short. Scheduling is the one with a number. Point runs the back-and-forth itself, finds a time several people can make, and publishes availability clients can book against, reconciled with the calendar you actually keep. Hold that against the per-seat booking line above. After that it’s the small free scaffolding. An ask made three paragraphs into a message is lifted out as a dated task, with the thread still attached. A watch you set in plain words tells you when the signed document lands, so you can stop going to look. Your other subscriptions keep writing to you, because every product at this layer leaves that alone. What changes is the order you meet them in, relative to the client who’s waiting. The weighing happens before you open the list, and how that ranking works is a subject of its own.
Weigh the controls in the same column as your other vendor answers. How much Point does on its own is set per kind of action, rather than by one switch. Each kind has three positions. Suggest only, prepare and wait, or handle it. Every action type starts in the middle out of the box. At the highest setting Point proceeds without checking with you, which is why the dial is the page to read before you raise any of them. Actions land in a log with times against them, and most can be reversed from it. One limit is shared by every product in this category. A message sitting on the recipient’s server is beyond recall. Home and work mail run in separate lanes. A second business you operate is walled off from the first. A seat is added or removed as somebody joins or leaves. Pricing is private beta, per seat, so it belongs in the column that multiplies when you hire.
Say the boundaries once, because they settle several keep questions above. Your record lives in your CRM and your document store, and Point leaves both of those exactly where they are. A shared or co-managed inbox with queues, assignment and response time reporting is a different product, so a practice whose inbound really is a support line should keep the thing built for that. Compliance stays a platform matter, so retention, holds and discovery sit there. And Point sits outside the nine jobs a firm has to cover. Everything Point does is the full inventory to check against your own week, and the version written for your trade opens closer to it.
Common questions
Does an AI email client replace anything in my software stack?
Usually one thing, sometimes two, and never the ones people expect. The realistic candidate is a per-seat booking product, because a client that runs the scheduling back-and-forth itself covers overlapping ground. After that it’s free scaffolding. A send-later add-on, a follow-up tracker, a personal task app kept for things that arrived by email. Your books, client record, document store, practice management, payroll and signing product all stay, because those jobs happen outside the mailbox.
Can I cancel Microsoft 365 or Google Workspace if I buy an AI email client?
No. The client signs in to the mailbox that subscription provides, so canceling the platform ends your email rather than replacing it. Your domain, addresses, storage, retention tier, archive and holds are all bought from the provider, and they stay there. An AI email client is an additional per-seat line on top of the platform.
What happens to everything the tool did if we stop paying for it?
Split it in two. Whatever the tool wrote into your mailbox stays, because Gmail labels and Outlook categories are properties of your account rather than the vendor’s. A message it archived is archived. A draft you sent is sent. A calendar entry it made is in your calendar. Whatever the product kept on its own side goes with it, which typically means the ranking, the summaries, the task list and whatever it learned from your corrections. Ask a vendor which of its work falls on which side before you connect, rather than at the point of canceling.
We pay for a tool nobody uses. Do we cut it?
Almost certainly. Ask one question first, because the answer is occasionally worth more than the saving. The problem it was bought for: is it still real, and where is it being solved now? If the answer is “by hand, in somebody’s inbox,” you’ve found a job that an assistance layer is pointed at, and the cancellation and the adoption are the same decision. If the problem went away or was never real, cut it today.
How long should we run the old product and the new one together?
Through one full cycle of whatever your firm’s busy period is, rather than one calendar month. A quiet week leaves the question open, because in a quiet week everything looks optional. Two signals are worth watching. Whether anybody went back to the old product during the busy stretch, and whether the job it covered actually got done. One extra month of one duplicate is the cheapest insurance in the whole exercise.
Does adding an AI email client change our compliance position?
It changes what you have to describe, and what you owe stays the same. For a firm the Safeguards Rule covers, 16 CFR 314.3(a) puts a written, comprehensive information security program on the firm, appropriate to its size and complexity and to the sensitivity of the customer information involved. A tool is a component described inside that program. A new vendor is a service provider, governed by a separate provision of the same rule, and the stack overview works that one through. Retention and legal holds stay the platform’s job in every case.
What should we keep that is not software at all?
Four things, in roughly this order. The retention decision, and the platform tier that makes it possible. The arrangement for covering somebody’s mail while they’re out. The offboarding list naming every product a person has a seat on. And the firm’s rule about what gets written into the client file. Summaries and extracted tasks are working notes, and the client file is still the record. Say that once at adoption, rather than discovering it six months later.
The short version
- An assistance layer sits on the mailbox you already license. The domain, the seats, the retention tier and the archive stay with the platform, whatever you buy at this layer.
- Grade each subscription by what would leave with it. Things holding the only copy of a record stay. Things holding a copy are negotiable. Things whose state you could rebuild in an afternoon are where the movable money is.
- The test for any product is where it writes. Gmail labels and Outlook categories are yours. A vendor’s ranking, summaries and task list belong to the vendor, and they leave when you do.
- What genuinely stops earning is short. A per-seat booking product, and a handful of free follow-up add-ons. Calendly’s Teams plan at $16 a seat a month is $1,152 a year for six people, and that’s a line worth holding up against the new one.
- Two free things also stop earning and are worth resolving anyway. Whatever split your provider already runs, since two sorters produce an inbox neither of them describes. And the triage rules sitting inside a rules quota Microsoft caps at 256 KB and will not raise.
- The product nobody opens needs a different question from the duplicates. Ask whether the problem it was bought for is still real, and where it’s being solved now. If the answer is “by hand, in the inbox,” that’s the job you just bought a layer for.
- Keep what never had a subscription. The retention decision, the coverage arrangement, the offboarding list, the rule about what goes in the client file, and the person accountable for the answer.
- Cancel nothing in the first month. Run the overlap through a genuinely busy week, then decide. Put the ninety-day review in the calendar so the saving actually happens.
The rest of this collection carries the parts around it. Cutting overlapping subscriptions is how the list you graded here gets built in the first place. The layer underneath the one you’re adding is where email fits. Then which tools actually reach the mailbox separates the products that are connected from the ones that are merely noisy, and the nine software jobs is the frame all four sort into.