AI email tools priced against a real inbox: 144 threads, 1 human

Every AI email tool sells the same promise: it reads your inbox, tells you what matters, and drafts the replies. Before paying for that, it’s worth knowing what a real inbox actually holds. So we counted one. Over eight days in August 2026 we took a full census of a live personal Gmail account, every thread, every sender, using the Gmail API rather than a screenshot or a memory.

The answer, up front: 144 threads arrived, and exactly one of them came from a named individual. One message went out of the account in the same eight days. That’s the workload these tools are being asked to price, and it explains why the cheapest option on test beat the ones charging $30 to $120 a seat.

What follows is the census, then the list prices taken from each vendor’s own pricing page in August 2026, then the strongest argument for paying anyway. That last part matters, because a sales inbox and a personal inbox are different animals and the pricing is aimed at the first one.

What eight days in a real inbox actually contained

The window ran from 13 to 20 August 2026 inclusive. We pulled sender, date and label for every thread the inbox received, and counted messages rather than estimating them. Gmail’s result estimator returned the same placeholder figure for every query that still had pages left, so every count below comes from enumerating those pages rather than from asking Gmail for a total.

Measure	Count	Per day
Threads received	144	18.0
Messages received	164	20.5
Messages sent from the account	1	0.1
Distinct sending addresses	73
Distinct sending domains	65
Threads from a “no-reply” style address	45 (31%)
Threads Gmail marked Important	49 (34%)
Threads from a named individual	1

Take the no-reply figure literally, because that’s how we counted it. Forty-five of the 144 threads came from an address whose local part spells out no-reply, noreply or donotreply. Those senders have told you in the address itself that a reply goes nowhere. A drafting assistant has no job on any of them.

The rest are barely different. Ten addresses accounted for 57 threads, so 40% of the inbox came from eight organisations: a supermarket, a bank, a gym chain, a hotel booking site, a health food retailer, a cashback service, an academic submission system and an SEO newsletter. Two of them send from two addresses each.

Every one is a bulk sender. That leaves the human count, which is where the test really turns.

One thread arrived from an address in firstname.lastname form at a company domain, with a colleague copied in. Five more came from role addresses at a local council and two customer service desks, where a person plausibly typed the message. Even counting all six as human correspondence, that’s six threads in 144, or 4%. The other 138 were generated by software.

What the AI email tools charge, and what they charge for

We took every price below from the vendor’s own pricing page in August 2026. They’re quoted in the currency each vendor lists, because converting them would mean picking an exchange rate we can’t source. Google and Microsoft publish UK pricing in pounds; Shortwave, Fyxer and Missive publish in dollars.

Tool	Cheapest AI-bearing plan	Top plan	Headline AI feature
Gmail, free account	£0	£0	AI Overview thread summaries
Google Workspace	Starter, £5.90 user/month	Plus, £18.40 user/month	Gemini assistant in Gmail
Microsoft 365 Copilot	£13.80 user/month, paid yearly	£19.32 user/month, monthly commitment	Copilot across Outlook and Office
Missive	Starter, $14 user/month billed yearly	Business, $36 user/month billed yearly	AI assistant and automations
Fyxer	Starter, $22.50 user/month billed annually	Professional, $50 user/month monthly	Drafts replies in your voice
Shortwave	Business, $30 seat/month	Max, $120 seat/month	AI search, filters and instant replies

The interesting thing isn’t the spread, it’s what the money buys. Fyxer’s Starter plan leads on “Drafts replies in your voice”, and Shortwave’s tiers are sorted by how much AI you’re allowed to use: its Business plan is documented as roughly 150 to 300 standard requests a day, with Premier at twice that and Max at six times. Shortwave also names the models behind each tier, from Claude Haiku 4.5 up to Claude Opus 4.6 with adaptive thinking.

Read that against the census and the mismatch is obvious. A daily allowance of 150 to 300 AI requests is being sold to an inbox that received 20.5 messages a day and produced one outgoing message in eight days. The quota isn’t the binding constraint. The supply of things worth writing is.

A reply-drafting engine priced at 300 requests a day, pointed at an inbox that produced one outgoing message in eight days, is not underpowered. It’s aimed at somebody else.

Rundowns AI

Missive is the honest outlier here, because it doesn’t really claim to be a personal assistant. It’s a shared team inbox, priced from $14 a user a month on yearly billing, with the AI assistant sitting alongside rules and analytics. Its CEO says as much on the pricing page itself.

As a small team ourselves, we believe in accessible pricing. Your local car repair shop probably uses Missive. So does a fancy tech company like Maple. We want everyone to be able to enjoy their inbox.

Philippe Lehoux, CEO at Missive, on the Missive pricing page

That framing is the useful one. These are team tools with AI attached, and the per-seat meter is the giveaway. We’ve written before about how per-seat billing prices a workforce rather than a workload, and email is the clearest case of it we’ve measured.

The summary feature Google already gives away

Thread summarisation is the feature every one of these products leads with, so it’s worth checking what it costs at Google. According to Gmail’s own help page for AI Overview conversation summaries, Gemini in Gmail can generate an overview at the top of a thread that synthesises the key points and replies, and suggest reminders for tasks with a due date. The requirement is not a subscription.

A work, school, or personal Google Account: This feature is available to all Gmail users globally.

Gmail Help, “Catch up on email threads with AI Overview conversation summaries”

You have to turn on Smart features in Gmail settings, and Google warns on the same page that Gemini features may suggest inaccurate or inappropriate information. But the summarising itself is free, on any account, worldwide. That reframes what the paid tiers are selling, because it isn’t summaries.

The paid line sits further along. Google’s help page for Gemini in Gmail lists suggesting responses, drafting mail, pulling information out of previous emails, reading Drive files and creating calendar events, and states that the feature requires an eligible Google Workspace or Google AI plan. Google’s consumer AI plans page is narrower still: it offers Gemini “to help you proofread in Gmail”, and notes that asking your Gmail inbox for an AI Overview is US only.

So a UK reader on a paid Google AI plan gets proofreading, not an inbox agent. That geographic split is the sort of detail vendor comparisons skip, and it changes what the money buys. It’s the same pattern we found looking at what free AI tiers actually cost you: the free version does the popular thing, and the paid version does the thing that’s harder to demo.

Gmail’s free classifier flagged 49 threads, and 48 were machines

Triage is the other half of the pitch, so we measured the free baseline for that too. Gmail’s importance classifier, which costs nothing and has shipped for years, marked 49 of the 144 threads as Important. That’s 34% of the inbox, which sounds like useful filtering until you look at who sent them.

Forty-eight of the 49 came from role or automated addresses: two banks’ notification systems, a payment processor’s receipts address, a data science platform’s no-reply, an academic submission system, a ticket marketplace’s orders address and a local council. The single human thread in the whole census was in there too, correctly. So the classifier is not broken. It’s answering a different question, namely which mail has consequences, not which mail needs you to write something.

That distinction is where an AI email tool could genuinely earn its money, and none of the pricing pages we read commits to it. They sell drafting, search and summaries. On this inbox the binding problem was that 138 of 144 threads needed no response at all, and the useful action was archiving, unsubscribing or ignoring. That’s closer to the kind of automation that survives contact with real work than to a writing assistant.

The case for paying, which rests on a different inbox

Here’s the strongest version of the argument against everything above, and it’s a good one: this is a consumer inbox, and none of these products is priced for it. A work inbox looks nothing like it.

Microsoft has the numbers, drawn from its own telemetry rather than a survey. In a Work Trend Index special report published on 17 June 2025, based on aggregated and anonymised Microsoft 365 productivity signals, Microsoft reported that the average worker receives 117 emails daily, most of them skimmed in under 60 seconds. It also found that 40% of people online at 6am are reviewing email for the day’s priorities.

That’s about 5.7 times the daily message volume we measured, on our arithmetic, and it arrives with a job attached. At 117 messages a day, a £13.80 monthly Copilot licence costs a fraction of a penny per message received, and the reply-drafting quota that looked absurd against our census starts to look tight. The counter-case is real and it holds.

Even so, the same Microsoft report points the other way on one measure. Mass emails with 20 or more recipients rose 7% over the previous year, while one-on-one threads fell 5%. The composition of a work inbox is drifting towards broadcast too, which is the same direction our census sits at an extreme of. Tools that price on drafting are betting against that trend.

What we could not test, and what would change the answer

Three limits, stated plainly. First, this is one inbox over eight days, which is a census rather than a sample of anything. A single account can’t tell you what typical looks like, and we haven’t claimed it does.

Second, we did not run a paid trial of every product and score its output. Nothing here measures whether Fyxer’s drafts sound like you or whether Shortwave’s search beats Gmail’s. Those are real questions and this piece doesn’t answer them. What it measures is the input each tool would have had on a real inbox, and the list price of processing it.

Third, one price we wanted is missing. Superhuman’s pricing page renders only through JavaScript and returned no plan prices to us, so it’s absent from the table rather than estimated from secondary write-ups. Google’s consumer AI plans page also served its feature list without the pound figures attached.

The loser on this test is the top of the Shortwave range. Max at $120 a seat a month buys six times the AI usage and 50 AI filters, against an inbox with 73 distinct senders and one human correspondent. Nothing in the census would consume that. Fyxer at $22.50 and Missive at $14 are cheaper misses rather than better fits, because they’re still metering a workload that barely exists here.

What would change the conclusion is a shift in the ratio, not in the software. If the human-written share of an inbox climbs from 4% towards the numbers Microsoft reports for knowledge work, the per-seat maths inverts and the drafting tools earn their meter. The measurement to repeat is the one we ran: count the threads, count how many came from a person, and divide the monthly price by the second number.