A school email moves a consent deadline, and an older message still shows the old date. A delivery problem needs a reply within five days, somewhere among a dozen promotions. A colleague's weekly update needs a friendly correction to one number. Keeping up with email means finding what changed, what is due and what you owe people, and one missed correction can mean a missed deadline or a reply you have to take back.
AI assistants offer to do this reading for you. We gave ChatGPT dots, Grok Bot, Instinct and Muse the same 72 email tasks each on 20 Oct 2026: sorting a morning's mail, finding a fact, summarizing a thread, pulling out actions, drafting replies and tracking follow-ups. We put the emails straight into the chat, the way you might copy a thread in, so that we measured the reading itself rather than whether an assistant could connect to your inbox. Nothing was sent or archived. Assistants working inside a real connected mailbox, including undoing a change and a week of watching for new mail on their own schedule, are the subject of our upcoming email review (how we are testing it).
Yes. Every assistant found what mattered on 69 to 72 of 72 email tasks. In practice, every one of them noticed that an urgent 10 a.m. request had been withdrawn, told you a cancellation fee wasn't in the mail instead of inventing one, and refused an email asking it to forward a child's medical records; none claimed to have sent anything. They differed in how much of the answer you have to read: Grok Bot most often led with the answer and suggested the next step, ChatGPT dots most often kept it short, and Instinct often buried a correct answer in long text. On an overall score that counts those habits alongside the basics and a recheck, Grok Bot scored 96.0 out of 100, ChatGPT dots 93.6, Muse 86.0 and Instinct 83.6.
These results describe one run per task in the setups we tested. Plans differed, model versions were undisclosed, and every account or conversation kept earlier history, so they do not show how an assistant does on a fresh account. Grok Bot also needed incomplete input resent in 3 of 72 tasks; its scored answers came after complete delivery.
Personal email · overall ranking
Grok Bot leads the email scorecard
Overall ranking, out of 100. One dot per assistant; the score includes the recheck.
- Mail was supplied in the prompt; connected inboxes were not tested.
- Plans and retained conversation histories differed; these are results for the tested setups.
- Red marks the best result on each measure; ties are marked equally.
Can it find the facts without making them up?
The tested assistants usually found the required information, including when the right answer was that information was missing. A fictional broadband provider mentioned an early termination fee but supplied neither the fee amount nor readable contract terms. ChatGPT dots opened with: "The cancellation cost cannot be determined from this packet." Instinct also said the fee amount was unavailable.
If you are deciding whether to cancel a service, an invented fee is the error that costs you. Across the 72 tasks, Grok Bot met every required check on 72 answers, ChatGPT dots and Instinct on 71 and Muse on 69. Meeting the checks still left room for a weak sentence: a reply draft could get every fact right and add a promise nobody asked for.
None of the assistants touched a message it was told to leave alone or claimed to have sent anything, on any of its 72 answers. Because the emails were in the chat, that covers what they wrote and proposed.
Does the answer tell you what matters first?
Grok Bot most often put the answer up front. In the corrected school-consent task, it opened with the form to return and explained that the deadline matched the mailbox's reference date, so it was due right away. It then distinguished the corrected form from the old one and flagged that the return method was missing.
Grok Bot led with the answer on 70 of 72 answers, ChatGPT dots on 63, Instinct on 40 and Muse on 37. Where a task called for a useful next step, such as asking for a missing form to be resent, Grok Bot suggested it on 19 of 23 answers, Instinct on 17, ChatGPT dots on 11 and Muse on 7.
Instinct was best at telling you whether a date means today, tomorrow or next week: it did so on 14 of 17 answers where it mattered, against Grok Bot's 12, ChatGPT dots' 9 and Muse's 7. Its school-consent answer said the corrected deadline fell on the day of the mail, though it also brought in its own later clock date.
Personal email · usability checks
Grok Bot leads on direct answers; dots on brevity
One square per applicable answer. Columns have different denominators.
PassedDid not pass
| Assistant | Answer first72 answers | Time in context17 answers | Useful extra23 answers | Concise72 answers |
|---|---|---|---|---|
| dots | 63/72 | 9/17 | 11/23 | 63/72Best |
| Grok Bot | 70/72Best | 12/17 | 19/23Best | 61/72 |
| Instinct | 40/72 | 14/17Best | 17/23 | 11/72 |
| Muse | 37/72 | 7/17 | 7/23 | 45/72 |
- Listed alphabetically, not ranked. Time and useful-extra checks apply only where the task calls for them.
- Red marks the best result on each measure; ties are marked equally.
How much of the answer will you have to read?
ChatGPT dots most often avoided padding. In the broadband-cancellation task, it stated the missing fee immediately, then listed what was known and what still had to be checked. Instinct reached the same basic conclusion but repeated source material, listed unrelated messages, and added separate references and action notes. Instinct's answer ran to 487 words, against 257 for ChatGPT dots.
ChatGPT dots kept its answer free of padding on 63 of 72 answers, Grok Bot on 61, Muse on 45 and Instinct on 11. Padding here means text you could cut without losing anything the task needed; a long answer the task required, such as a full table or a finished draft, did not count against it.
Some answers turned background mail into a to-do. Muse avoided that on 69 of 72 answers, ChatGPT dots on 67, Grok Bot on 66 and Instinct on 65. When an assistant proposes a cleanup, check which messages it picked as well as how much it wrote.
Personal email · pass rate and scorecard
Basic passes cluster; usability scores spread out
Filled dot: scorecard out of 100. Hollow dot: first-answer pass percentage. Each assistant answered 72 tasks.
- Listed alphabetically, not ranked. The two measures answer different questions; their gap is not an error rate.
- Red marks the best result on each measure; ties are marked equally.
Will asking it to check again fix the answer?
Rarely. Asking it to check again sometimes improved a draft. Muse's first draft explaining a delayed forecast included: "I will let you know if that timing changes." On recheck, it identified that sentence as a promise the request had not authorized and removed it. The revised email draft kept the explanation of the delay and continued to omit confidential staffing details.
Across all 72 tasks, though, the recheck changed little and sometimes made an answer worse. Grok Bot's answers met every required check on 72 tasks after the recheck and Instinct's on 71, the same as their first answers; ChatGPT dots' fell from 71 to 70 and Muse's from 69 to 67.
Before you send a draft, check recipients, commitments and confidential details against the original mail yourself; don't rely on the recheck to catch them.
So, can it read your email so you don't have to?
Yes. Each assistant we tested found the facts, the deadlines and the messages to leave alone almost every time: the withdrawn 10 a.m. request, the cancellation fee that wasn't in the mail, the request for medical records it should refuse. The handful of misses were a deadline not flagged as today, a wrong message number and a draft that repeated a claim the mail contradicted, so read dates and drafts before you act on them.
If you want the answer up front with the next step already suggested, Grok Bot did that most often; keep its input-delivery interruptions in mind if you copy in long threads. If long explanations are your frustration, ChatGPT dots stayed concise most often. Instinct put dates in context best but added the most text, and Muse usually met the required checks but less often led with the answer or suggested a next step. We have not checked that anyone can reach these exact assistants today.
For your own trial, use a thread with a corrected deadline, a request whose answer is missing and a draft with confidential background, and check the answer against the mail before acting.
Method: how we tested and where the evidence stops
Each assistant answered 72 tasks, with 12 in each of 6 job types, followed by a fixed recheck. The test was in English, with fictional, self-contained mail. ChatGPT dots used ChatGPT Pro; Grok Bot used X Premium; Muse's recorded plan was free and was not re-inspected for this run; Instinct's WhatsApp channel did not show a plan. ChatGPT dots, Grok Bot, and Instinct reused persistent conversations. Muse opened a new chat per case but retained account history. No model identity was shown.
The article reports the resulting scores and examples. The full method and the personal-assistant testing method document the checks, how answers were scored and the alternative weighting. The scored outputs were collected on 20 Oct 2026. The supplied mail used its own reference date, so deadlines were assessed against that date. Each task ran once after complete input delivery. Time saved, repeated daily use, other plans, fresh conversation histories, Chinese-language mail and the chance of success on an unseen task are not covered in this review.
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 scored first answer, overall scores and pass counts, each check with its denominator, and scores by type of task) are available to download from this page.
Corrections. Spot an error? Our corrections log and how to report one: analyst.ai/corrections.
Sources
The records below will be added to the evidence pack.
- The 72 email tasks, the mail given with each task, and the answer keys.
- Each assistant's first answers and recheck answers, and a record of how each assistant was set up.
- The scoring of every first answer and recheck, and how disagreements between scorers were settled.
- The overall-score judgments for every answer, the scoring guide, and the decisions on disputed scores.