The human becomes the scarce reader
For thirty years, every email was written for one reader. It now has two.
The argument in one line. Email assumed that a person reads it. Soon every customer relationship will have two readers: the person and their AI agent. Utility will increasingly be delegated to the agent; experience will not. Brands will have to earn the person through choice, ownership, progress and belonging. Convince the agent. Earn the human.
As machines become better at consuming messages, the objects left for humans may become more valuable. The previous essay, Issue a Collectible, introduced those objects. This one explains why they will matter more.
1
The second reader arrives
The person was always the endpoint. That assumption is ending.
Everything about email design assumes a human reader. The subject line is written to catch an eye in a list. The preview text is the second glance. The hero image, the layout, the colour of the button and the urgency of the call to action all exist because a person, in a second or two, decides whether a message deserves more of their attention.
That person now has help. In January 2026 Google brought Gmail into what it called the Gemini era, with AI Overviews that summarise long threads and answer questions about the inbox. Its new AI Inbox view reorganises email around summaries, topics and to-dos rather than individual messages, and in May it showed a conversational version that lets people ask their inbox questions aloud. Microsoft’s Copilot in Outlook assigns incoming email a priority with a reason, and can pin, flag, archive or delete messages on command.
None of this is yet the autonomous customer agent of every AI forecast. Today’s inbox assistants mostly summarise, sort, search and draft, and a person still approves the consequential actions. The next stage is the one that matters for marketing: software that compares offers, files receipts, tracks deliveries, schedules renewals and prepares a dispute, within permissions the person has set. Many people will switch these features off. The direction is still clear.
The consequence for brands is simple to state. Every message now has two possible readers with different needs, and most brands still write for one of them.
Figure 1 — Two readers. The same email now meets a person, who wants meaning and things to keep, and an agent, which wants facts and actions.
2
The agent’s lane
Sell and Notify increasingly become structured, machine-readable and machine-actionable.
An agent wants different things from a person. It wants facts, prices, dates and terms, stated plainly enough to compare, and a way to act on the person’s behalf. It is indifferent to design, and less susceptible to manufactured urgency. A real expiry date matters; “Hurry!” does not.
Of the four jobs of brand email, Sell, Notify, Deepen and Reactivate, the first two move into the agent’s lane first.
Notify goes first. Confirmations, statements, alerts and renewal notices are already data dressed as prose. An agent reads them, files them and acts on them: the receipt becomes a record, the alert becomes a trigger, the renewal notice becomes a comparison. Most people will be glad never to open another shipping update.
Sell follows. An offer used to be a nudge aimed at a person. To an agent, an offer that was once persuasion becomes a set of comparable attributes: effective price, expiry, exclusions, eligibility, delivery, cancellation terms, and how well it fits what the person has said they want. A discount is one variable among these, not automatically the deciding one.
That creates a new kind of competitive pressure. Human-facing marketing often benefits from ambiguity: “up to”, “from”, urgency without context, terms behind a click. Comparison needs precision, so agents are likely to penalise exactly those habits. A sender that wants to be chosen by the customer’s software may find that clarity is not only a compliance virtue but a way to win.
My work on NeoMails points the same way from the brand’s side. The more an action can be completed inside the email, with its terms stated and its outcome confirmed, the less a journey depends on a person clicking through screens. That is not a claim that agents will operate interactive email the way people do. It is a claim about the direction of travel: from persuasion towards structured intent, meaning what is offered, on what terms, and how to accept.
For brands, the task in this lane fits in three words: convince the agent. That means clean data, honest terms, clear permissions and actions that can be carried out safely.
3
The contested lane
If information can be handed to an agent, it may no longer hold a human.
This is the turn in the argument, and it applies to my own work before anyone else’s.
An agent can summarise a newsletter as easily as a receipt. The Brand Digest was built to give customers a reason to open between purchases, and its Read block, three useful things from the customer’s world, is exactly the kind of content an agent condenses into a line of a morning briefing. If the Digest earned attention only by informing, the agent would take that attention and pass on a sentence of it.
This does not make usefulness worthless. Accuracy still builds trust, and an agent choosing what to surface may prefer sources that have proved reliable, which rewards editorial quality. A well-made publication may be the one the agent quotes. But being quoted is not being read. Usefulness can earn the brand a place in the agent’s briefing without earning the person’s own open. The editorial advantage moves away from summarising what happened, which machines will do well, towards taste, framing, original access and a distinctive voice.
Underneath this lies a larger shift. AI will separate utility from experience. Utility is increasingly delegable. Experience is not. A renewal notice is utility. A price comparison is utility. A receipt is utility. Completing something I chose to collect is experience. The question for anyone who wants human attention between transactions stops being “is this worth knowing?” and becomes “is this worth doing, owning or finishing myself?”
Figure 2 — Utility and experience. The test is not whether an agent can do something, but whether delegating it keeps the value of doing it.
4
The human lane
Choice. Ownership. Progress. Belonging.
The question is not whether an agent can do something. Agents will be able to choose, catalogue, track progress and arrange meetings. The question is whether delegating it preserves the value of doing it. An agent can complete my tax return for me, and I am delighted. An agent can complete my crossword for me, and it has destroyed the point.
Four things can be assisted by an agent but lose something when they are entirely delegated.
Choice. Not every choice counts. Window or aisle is a preference I want my agent to remember and apply; that is routine choice, and it will be delegated. Picking which of three photographs should lead next week’s edition, or which Card to give my sister, is different. Routine choice gets delegated. Expressive choice stays human. The Digest’s Do block works when it asks for the second kind.
Ownership. An agent can list what is in an album. It can remind me what I hold. The small pleasure of opening a Card and deciding to keep it belongs to the person. Ownership is felt, not reported.
Progress. State that changes because of what the person did. The agent can tell me I have 21 of 24 Cards. It cannot feel the satisfaction of completing the Set for me. The arithmetic can be delegated; the motivation cannot.
Belonging. Other people: a duplicate that is worth something to a friend, a trade between siblings, a Set completed alongside someone else. An agent may find the friend who has my missing Card and arrange the swap. It cannot be the friend.
This is where the Digest’s Do and Keep blocks, and Cards, find a deeper rationale. They were designed to create return between purchases. In a world of two readers they do something more: they are the part of the relationship that stays with the person when the utility is handled for them. They appear here as consequences of the argument, not its subject. Any brand, selling anything, faces the same choice between messages an agent can absorb and experiences it cannot.
Figure 3 — Three lanes. Sell and Notify move to the agent; the Read block is contested; choice, ownership, progress and belonging stay with the person.
The more utility gets delegated, the more valuable human experience becomes.
5
The two-part email
Future messages may carry two parts: one structured for the machine, one designed for the person.
Email already carries more than one version of a message: a plain-text part, an HTML part, and for some senders an interactive part. Mail providers have read structured markup for bookings and orders for years, to show flight times and delivery dates above the message. The idea of one message with several forms is not new. What is new is who the forms are for.
The machine part needs more than structured data. An agent acting on a person’s behalf has to know several things before it acts. Is this really the brand it claims to be? Is the offer authentic? What has the person authorised me to do: read the due date, compare the price, renew automatically, spend money, or only draft a reply? Must I ask before cancelling? What record will prove what I did? So the machine payload is data, provenance, permissions and an executable action, together. Reading and acting have to be kept apart: an agent may be allowed to extract a date but not to renew a policy.
The human part carries experience: a choice only the person should make, an object they can keep, progress they can see, a story worth spending attention on, another person who matters. It may become smaller as the machine handles more of the routine. What remains has to justify being experienced rather than summarised.
Each part is weaker without the other. A message optimised only for persuasion may be poorly represented by the agent that reads it. A message optimised only for machine processing may give the person no reason to experience it. A brand that wants both a transaction and a relationship needs both parts.
None of this should be one company’s format. What fields a renewal should expose, how a sender states an executable offer or an expiry, how provenance travels, what permissions a person can delegate, and how a message marks the part that deserves the person’s own attention: that is an agenda for an open standard, bigger than any one email provider or brand.
Figure 4 — The two-part email. Data, provenance, permission and action for the agent; an experience only the person can use.
6
What brands should do now
Get both lanes right before the agents force the question.
Make Sell and Notify legible. State prices, dates and terms in full and in plain language. Make actions explicit and permissions clear. The brands that are easiest for an agent to understand will be the easiest for an agent to choose.
Stop relying on information alone to hold attention between purchases. Useful content still matters, but assume that any paragraph can be summarised on the customer’s behalf, and ask what the customer would lose if it were. Add things worth doing, owning and finishing.
Measure the difference. A message processed by an agent is not a message read by a person, and opens were already unreliable. The question worth tracking is not “was the message processed?” but “did the person choose to engage, and did that change what they did next?”
Show restraint in the human lane. If software is handling the routine, the answer is not to make every human-facing email louder. Fewer moments that deserve the person’s attention will do better than more messages trying to win it back from the agent.
7
What would prove it wrong
The scoring date is 31 December 2028.
Agents do not start acting in the inbox. If, by the end of 2028, customer-side agents have not begun to materially mediate inbox triage, offer comparison or routine service actions within permissions people have set, the two-reader architecture was premature. Summaries alone do not settle it; a summary is a reader, not an actor.
What “materially” should mean. I will not invent a percentage. The evidence is observable, and it has three parts that should be scored separately: capability, meaning inbox providers offering real delegation beyond summaries and drafts; adoption, meaning people using agents to decide what deserves their attention, compare options or complete routine actions; and sender behaviour, meaning brands designing communications for their customers’ agents. The thesis strengthens only if all three move.
Agents stay assistants. Liability and permission may keep agents organising and recommending but rarely acting. If so, Sell and Notify stay human-readable for much longer, and machine legibility becomes an improvement rather than a new lane. That would also be worth knowing.
Summaries do not reduce engagement. If agents summarise newsletters and human engagement with them does not fall, then information alone still holds people, and the contested lane was too pessimistic. People may want the source, not just the answer.
People delegate the experience too. If people happily let agents collect, trade and complete on their behalf, then even objects can be absorbed, and the human lane is narrower than I think.
The argument does not depend on autonomous agents arriving everywhere. It depends on a simpler shift that is already visible: software is consuming more of the inbox on the person’s behalf, and each step in that direction changes what is scarce. Information becomes easier to process. Human attention becomes harder to earn.
Convince the agent. Earn the human.
