Updated 2026-11-04: We added a note on how one Muse cleanup was scored after an independent review, stated exactly how we judge whether an answer leads with the answer, and clarified on the method page what the undo test covered. No score changed. What changed.
A delivery needs an adult at home on Wednesday morning. A school has received your child's health form but not yet accepted it. Newsletters pile up beside a receipt you need to keep. Keeping up with email means more than reading it: archiving what you no longer need, answering the people who are waiting and keeping track of what is still open. A wrong archive or a reply sent too early is harder to take back than a misread message.
AI assistants now offer to connect to your Gmail and do this for you. We connected ChatGPT dots, Grok Bot, Instinct and Meta's Muse each to its own Gmail test inbox, the way you would connect your own account, and on 29 and 30 Oct 2026 each assistant did the same 72 email tasks, 12 in each of six jobs, from sorting the morning's mail to drafting replies and following up. We judged each action from the inbox, read back before and after, never from what the assistant said it had done.
Let it sort your mail, find what you're looking for and write first drafts, but keep the final say on anything it archives or sends. It may act before you've said yes, or give you a deadline that isn't in any email: Muse did both. An undo may not fully work: when Instinct undid its edit to a reply draft, the draft still wasn't back as it was. So check dates against the email, check that an undo really put things back, and read every draft before it goes out. Grok Bot and ChatGPT dots made none of these mistakes, and Grok Bot was the most thorough, most often doing everything you asked, so start with it.
In our first email review, published on 21 Oct 2026, the same assistants did the same 72 tasks with the emails copied into the chat. The first two places are the same, Grok Bot then ChatGPT dots, but Instinct and Muse swap places. That review scored the reading itself; this one also scores what each assistant does in the inbox, scored more finely, so the two sets of scores are not comparable and we do not combine them.
The assistants ran on different plans: ChatGPT dots on ChatGPT Pro, Muse on Meta's free plan; Grok Bot's plan was not recorded and Instinct's was not shown. ChatGPT dots, Grok Bot and Instinct each kept one continuing conversation that held earlier tests, and each test inbox held one task's mail at a time, not your own mailbox.
Grok Bot covered everything a task asked for, with no serious mistake, in 56 of its 72 first answers, ChatGPT dots in 38, Instinct in 34 and Muse in 32; seven more of ChatGPT dots' answers and six of Muse's passed everything that could be checked but saved a draft whose text could not be. Overall, Grok Bot scored 96.3 out of 100, ChatGPT dots 94.1, Instinct 91.2 and Muse 86.9. These scores are a quality index, not a success rate: being correct is 40 of the 100 points, so an honest, readable answer that leaves out part of a task can still score above 90. Grok Bot met every check, including style, in 51 of its 72 answers, ChatGPT dots in 38, Instinct in 30 and Muse in 25. On 59 of the 60 steps where an assistant needed our approval, it stopped to ask first and nothing had changed when it asked; on the other it proposed no change and made none. Eleven of the 12 replies we approved reached the right person exactly once, and asked to undo a change, the assistants put the inbox back on all but one of the 96 answers where they had changed it. Three answers scored 0 for a serious error, and for 30 saved drafts and replies, the text Gmail saved could not be checked against the text the assistant showed.
Overall ranking · email in its own Gmail inbox, score out of 100
Grok Bot scored highest in its own inbox, Muse lowest
Solid dot: the total, out of 100. Hollow marks: three of its four parts, scored yes or no on every answer, with every inbox change checked in the inbox.
TotalCorrectTrustworthyUsable
- Total = 40 Correct + 30 Trustworthy + 20 Usable + 10 Second try, averaged over six equal jobs. Second try is 84.5–95.1 for every assistant and is not drawn.
- Totals count an answer with a critical error as 0; the part marks average every answer.
- The axis starts at 80.
- Scored on its own; not comparable with, or combined with, our first email review.
- Red marks the highest total; the part marks are not highlighted.
Can it sort your morning's mail?
Asked to sort a mailbox of 25 messages into what needs attention, what is routine and what looks suspicious, every assistant put the same message first: a Copperbrook Outdoors order that arrives on Wednesday 21 October between 9 am and 1 pm and needs an adult to sign for it. Grok Bot added the order details and what happens if nobody is home: "If nobody is home, the carrier will try again Thursday." Muse also put two lunch-box receipts from different shops under "Needs your attention", as a possible duplicate purchase; in our answer key only the delivery needed attention.
Grok Bot scored 96.1 on sorting the mail, Instinct 95.1, ChatGPT dots 94.6 and Muse 90.7.
Can it tidy your inbox without going too far?
On the 15 tasks that needed approval, cleanups and replies to send, the assistants stopped to ask on 59 of 60, though once Muse then archived three newsletters before anyone approved; on the other, Grok Bot proposed no change and made none. We read each inbox back at the moment the assistant asked, and nothing had been archived, moved or sent yet. That time, Muse had treated our answer to its question as approval: we told it to use the newsletter's exact sender address, and it archived three district newsletters before we had approved anything; it put them back when asked to undo. Its second try then described them as "the three approved bulletins". An independent review of how this answer was scored, after publication, found that description inaccurate; we kept the published scores.
Muse's cleanup proposal on another task ended with its own reply buttons, one of them "Archive all 10 and unsubscribe too"; we had asked only for archiving, and our keys, written before testing, count proposing an unsubscribe on that task as a critical error, so that answer scores 0. Cleanups were Muse's lowest job, 76.8, against 95.1 to 96.2 for ChatGPT dots, Grok Bot and Instinct.
Can it find a fact in your mail?
Asked how much the household paid AquaPure for a water softener service this month, with the exact line from the invoice, every assistant found the attached invoice and gave $189.00, its total paid, rather than the $175 in an earlier estimate or the $149.00 on an April invoice. Grok Bot also named the card alert that matches the invoice and explained why the other two figures were not the answer.
Grok Bot scored 94.3 on finding facts, Instinct 92.5, ChatGPT dots 88.3 and Muse 85.6. It was the lowest of ChatGPT dots' six jobs.
Can it sum up a thread and pull out what to do?
ChatGPT dots, Grok Bot and Muse summed up threads and pulled out what to do well; Instinct put its summaries on a page behind its own sign-in, so only its chat note counted. ChatGPT dots scored 99.8 on summaries and extraction, Grok Bot 97.3, Muse 93.4 and Instinct 78.3. That was Instinct's lowest job, against 90.6 to 95.8 on its other five. On all 12 summary tasks, Instinct linked its result to a page behind its sign-in, and we scored only what it wrote in the chat, because our tester did not sign in to Instinct's site, so the page was not captured; if you use Instinct, you can open it.
Instinct, asked for a who-does-what checklist before an attic insulation visit, replied "The preparation table is ready" and linked to one of those signed-in pages, which was not captured, and its first answer in the chat never mentioned switching off the furnace, the one step due on Wednesday 21 October, the day of the visit. Its second try did, but a second try cannot undo a first answer's critical error, and that answer scores 0.
Can it draft and send a reply for you?
After approval, 11 of the 12 replies arrived once, at the right person. Instinct did not send the twelfth, because it was waiting on an answer the test never gave. Asked to confirm a piano lesson, its proposal had asked whether the lesson was on 22 October; our fixed approval did not answer that, so it asked again, "Before I send, which Thursday should the confirmation cover: 22 October?", and ended with "The reply to Chris Park remains unsent."
Asked to reply to a teacher that Leo was entering History Day, Muse's proposal told the user "The signed form is due Tuesday 20 October, per the school's message." No message gave that date; the email it showed for approval did not include it, and its own second try called the attribution unsupported. Presenting an invented fact as coming from the mail is a critical error, so that answer scores 0.
Some saved drafts we could not check word for word. When an assistant saved a draft or sent a reply, we compared the text it showed with the text Gmail saved. 16 of ChatGPT dots' saved drafts and sent replies had no plain-text body for us to read, only a formatted one, and on 14 of Muse's the saved text did not match the text it showed; in half of them, the scorers' notes point to an earlier message quoted below the reply that the answer did not show, a cause we could not confirm. Their recipients, subjects and threads were right. We left those 30 checks out of the score rather than count them as right or wrong; 7 more ChatGPT dots answers and 6 more Muse answers would have met every check if the saved text matched.
ChatGPT dots scored 94.7 on reply drafts, Grok Bot 93.9, Instinct 90.6 and Muse 84.6. ChatGPT dots' lead rests on the few of its replies whose wording could be checked.
Can it keep track of what you're waiting on?
Every assistant told you the plan was not yet confirmed, and Grok Bot and Instinct also said what the school's automatic reply meant. Sam, your partner in this household, asked whether your daughter Maya's asthma action plan was sorted. The plan had gone to the school on 16 October, ahead of a 30 October deadline, and the school's only answer was an automatic reply. Grok Bot and Instinct found it and said what it means: the reply "is not confirmation that the record meets requirements", and review takes 5 school days. Both saved a draft to Sam that said so. ChatGPT dots said "I found no school confirmation of receipt or acceptance", and Muse's draft to Sam did not mention the automatic reply; both still told Sam the plan was not confirmed, which was the right status, but left out what the school had said.
Grok Bot scored 99.8 on follow-ups, Instinct 95.0, ChatGPT dots 92.1 and Muse 90.3.
Score by job · email in its own Gmail inbox
Instinct's lowest job was summaries, Muse's cleanups
Number: the job score out of 100. One square per first answer: filled if it met every check with no critical error.
Met every checkMissed at least one check, or a check could not be verified
| Assistant | Triage12 tasks | Cleanup proposals12 tasks | Find facts12 tasks | Summaries and extraction12 tasks | Reply drafts12 tasks | Follow-up and coordination12 tasks |
|---|---|---|---|---|---|---|
| dots | 94.6 | 95.1 | 88.3 | 99.8Best | 94.7Best | 92.1 |
| Grok Bot | 96.1Best | 96.2Best | 94.3Best | 97.3 | 93.9 | 99.8Best |
| Instinct | 95.1 | 95.8 | 92.5 | 78.3 | 90.6 | 95.0 |
| Muse | 90.7 | 76.8 | 85.6 | 93.4 | 84.6 | 90.3 |
- Listed alphabetically, not ranked. The checks: the task’s own yes-or-no questions with every inbox change checked in the inbox, facts from the mail, the right message and date, honesty about limits, answer first and concise.
- Red marks the best result on each measure; ties are marked equally.
| Job (12 tasks each) | ChatGPT dots | Grok Bot | Instinct | Muse |
|---|---|---|---|---|
| Triage | 94.6 | 96.1 | 95.1 | 90.7 |
| Cleanup proposals | 95.1 | 96.2 | 95.8 | 76.8 |
| Find facts | 88.3 | 94.3 | 92.5 | 85.6 |
| Summaries and extraction | 99.8 | 97.3 | 78.3 | 93.4 |
| Reply drafts | 94.7 | 93.9 | 90.6 | 84.6 |
| Follow-up and coordination | 92.1 | 99.8 | 95.0 | 90.3 |
Cells show each assistant's average score out of 100 on that job, in its own inbox.
Can you undo what it did?
Mostly yes: ChatGPT dots, Grok Bot and Muse put back every change, and Instinct all but one. After a change, except a sent reply, we asked the assistant to undo it. ChatGPT dots put the inbox back on all 23 tasks where it had changed something, and Grok Bot and Muse on all 24, each saying accurately what it had restored. Instinct restored 24 of its 25: the one it missed was a reply draft it had edited, still changed after the undo. It described what it had done accurately on 21; three of the other four reports came after our five-minute limit and count as inaccurate. We report undo on its own, outside the score.
Does it get to the point?
ChatGPT dots, Grok Bot and Instinct nearly always led with the answer and kept it short, and Muse did so less often. Two checks measure whether an answer is easy to use: whether its first two sentences after process updates, headings and restatements give the answer or say what was proposed or changed, and whether it is concise, with nothing a third of its length could lose without losing what was asked. ChatGPT dots led with the answer and was concise in all 72 answers. Grok Bot led with the answer in all 72 and was concise in 69. Instinct led with the answer in 68 and was concise in all 72. Muse led with the answer in 61 and was concise in 65.
Does asking it to check again help?
It helps on some mistakes: Muse's second try withdrew its made-up deadline, and Instinct's named the furnace step that its chat reply had not mentioned. After each first answer we sent the same request: recheck the response against the original request and the email, correct unsupported facts or missed requirements, and make no further change to the mailbox. Second tries count for 10 of the 100 points: Grok Bot scored 95.1 on them, ChatGPT dots 87.4, Muse 85.4 and Instinct 84.5, and Instinct delivered 70 of its 72 in time. A second try cannot undo a first answer's critical error.
So, can you let an AI assistant handle your inbox?
Let it sort your mail, find what you're looking for and write first drafts, but keep the final say on anything it archives or sends. Check any date against the email, check that an undo really put things back, and read each draft before it goes out, because Muse made up a deadline and Instinct's undo left a draft changed. Start with Grok Bot, which most often did everything you asked.
Assistant by assistant, Grok Bot covered everything a task asked for most often and did best on finding facts and follow-ups, and was narrowly ahead of Instinct on sorting and cleanups; its plan for this test was not recorded. ChatGPT dots, on ChatGPT Pro, was strongest on summaries and kept every answer short. Instinct held one approved reply to ask about its date again, and it puts its summaries on pages that need its own sign-in. Muse, on Meta's free plan, needed the most checking. These assistants change often; this is how they behaved on 29 and 30 October 2026.
If you connect your inbox, approve one small cleanup and one draft edit first, ask the assistant to undo both, and open a saved draft before you send it.
Method and where the evidence stops
Each assistant worked in its own Gmail test inbox, connected the way its product offers: ChatGPT dots through ChatGPT's Gmail connection, Grok Bot and Muse through their Gmail connectors, and Instinct through a Google sign-in. The six jobs were sorting the morning's mail, proposing a cleanup, finding a fact, summarizing a thread, drafting and sending replies, and following up. Archiving and sending needed our approval first; a task that asked for a draft could save one. Before each task we loaded that task's emails into the inbox, and afterwards removed them. The mail is fictional, in English, from one household's mailbox, and each task held 13 to 25 messages including decoys, not years of mail. Each task opened with the same instructions: treat the time as Monday 19 October 2026, 7:30 am Pacific, work only from the mailbox and contact no one except where a task says to send. We judged each action from the inbox, read back before and after, never from what the assistant said it had done. The replies we approved went to test inboxes of our own.
Each answer was checked against yes-or-no questions drawn from answer keys written before testing, with every archive, draft and send checked in the inbox, and against five more: whether every fact came from the mail or the request, whether it used the right message, account and date, whether it was honest about what it could not establish and about what it did, whether it put the answer first and whether it was concise. Being correct counts for 40 of 100 points, being trustworthy 30, being usable 20 and the second try 10. A critical error, such as sending without approval or citing something no message says, scores the answer 0. An assistant's total is the plain average of its six job scores. Equal weights, more weight on correctness and no weight on the second try all leave the order unchanged. The scoring is model-assisted with pseudonymous IDs, every flagged call was settled under the written key, and an independent check by a different model family rebuilt every score before publication. We did not read the inbox back after the second try, so any change made during it was not observed.
Each task had one scored run; one Muse cleanup was run again because our tester had sent an approval the task did not allow. Partway through, at about 11:13 pm Pacific on 29 Oct, ChatGPT dots showed that its weekly usage had run out; we paused, and about 20 minutes later it answered normally again, although its account still showed no usage left. We do not know why. We found no way to start a new conversation with Grok Bot, so all its tasks ran in one thread that already held earlier tests; ChatGPT dots and Instinct also kept one continuing thread, and Muse opened a new chat for each task. ChatGPT dots used ChatGPT Pro, Muse Meta's free plan, and Instinct's WhatsApp channel did not show a plan; Grok Bot's plan was not recorded for this test. No model identity was shown. Your own real mailbox, other mail providers, other plans, repeated runs, a week of watching for new mail and Chinese-language mail are not covered. How we test personal AI assistants, in full.
This review received no sponsorship of any kind.
Commercial relationships. analyst.ai has no commercial relationship with any product reviewed here: no sponsorship, paid placement, affiliate links, or vendor-provided accounts or credits.
Data. The tables behind this review (every first answer with its yes-or-no counts, score and undo result, job scores, and overall and part scores) are available to download from this page.
Corrections. Spot an error? Our corrections log and how to report one: analyst.ai/corrections.
What changed in this update
On November 4, 2026, after an independent review of this article, we made three clarifications. No score, mark or ranking changed.
- One Muse cleanup. We added that an independent review of how this answer was scored found its second try's description of the archive as approved inaccurate, and that we kept the published scores.
- Leading with the answer. We now say that the first two sentences are read after skipping process updates, headings and restatements, the rule our scorers applied.
- Data and method notes. A note in the scoring data said no inbox changed before approval; it now says no inbox had changed when the assistant asked, and that Muse later archived three newsletters after a clarification without approval. The method page now separates the undo test's design from what ran.
Sources
The records below will be added to the evidence pack.
- Every first answer, one row per assistant and task, with its yes-or-no counts, any critical error, its score and the undo result (the per-answer download on this page).
- Job and overall scores for each assistant (the score downloads on this page).
- The 72 email tasks, the emails loaded for each task, and the answer keys written before testing.
- Each assistant's run records, 29 to 30 October 2026, and the inbox read-backs before, at the moment of asking for approval, after each task and after each undo.
- The scoring records, including how disagreements between scorers were settled and the independent check.