Your Best Customer Wont Be Human

Last week I wanted to get movie tickets for The Devil Wears Prada 2 with my sister, her husband, and my wife. I told an agent the date, the cinema, and the constraints: good seats together, not too close to the screen, evening showing if possible, and premium format only if the price was sensible. I went to the gym. By the time I got back, the tickets were booked, paid for with a tokenised credential, the QR codes were in my wallet, and the calendar invite was on everyone phone. No cinema tabs, no seat-map wrestling, no twelve-step checkout. No are you sure you dont want popcorn the size of a carry-on. Just the thing, finished.

Okay, this didnt really happen, but that would have been really cool, right? We did watch the movie, though, and it was awesome.

For roughly 50 years, software has been built with one user in mind: a human at a screen. Every interface, every checkout flow, every dashboard, every CAPTCHA, every are you sure you want to do this modal exists because there is a person on the other side who needs to be guided, persuaded, reassured, or stopped from clicking the wrong button. That assumption is about to break. Software is about to have a second user, and that second user will eventually be the bigger customer.

Chatbots Answer, Agents Act

So what are agents, really? Agents are AI systems that act, not just answer. They take a goal, plan a sequence of steps, call tools, hit APIs, verify results, and try again when something fails. They run on a server somewhere, often without a human watching, and they finish things.

The difference between a chatbot and an agent is the difference between a recipe and a chef. One tells you what to do, while the other does it.

Software has had one user for 50 years

Look at any modern web application. The product card with the lifestyle photo. The hover state. The add to cart animation. The four-step checkout that asks for your email, then your shipping address, then your payment, then makes you confirm again. The cookie banner. The newsletter popup. The discount modal that fires 30 seconds after you land. All of it is designed for a human being with a wandering attention span and a credit card.

Ironically, and from a technical perspective, this experience really sucks for autonomous agents, even for those that have the ability to use a browser and click-about. It doesnt work, and worse still, it actively gets in the way.

An agent does not need a hero image. It does not need to be persuaded by social proof. It does not respond to urgency tactics or scarcity messaging. It needs structured product data, a clear price, a deterministic API for placing an order, and a way to authenticate a payment. Everything else is unnecessary friction.

I fundamentally believe that software is going to have to change to cater for this new class of users. Companies that figure this out first will stop competing on UX polish and start competing on something closer to API quality. The kind of thing engineers have wanted for years and marketers have spent two decades burying under a layer of conversion-optimised theatre.

If your product can only be bought by a human clicking through your funnel, you are going to find that a growing share of demand simply routes around you.

E-commerce is the first domino

Purchasing is the highest-friction, highest-volume, most-instrumented human workflow on the internet, which makes it the cleanest place for agents to take over first. Three things change at once.

Discovery. Agents will not browse, they will query. They take a constraint set, I want X, under $Y, delivered by Z, with at least a four-star rating and a 30-day return policy, and they evaluate options against it. SEO as it has existed for two decades becomes mostly irrelevant, because agents do not click through to a page and read your meta description. What replaces it is whether your product data is structured well enough for an agent to surface your item as the best match against a given set of constraints. Call it Agent Discoverability if you want a label. It is the new shelf placement.

Decision. Agents are immune to most of what marketing has spent the last 40 years optimising. They do not respond to brand colour palettes, lifestyle photography, celebrity endorsements, or aspirational copy. They evaluate on price, specs, reviews, return policy, delivery time, and whatever weights the user has set. Brand still matters, but only as a quality signal that survives compression into a structured field. Trusted seller, 4.7 stars, 99% on-time delivery is a brand. Live your best life is not.

Payment. This is where it gets real, because nothing happens commercially until money moves. The card networks and the fintechs have figured this out and are moving fast.

Mastercard launched Agent Pay in the UAE in November 2025, in collaboration with Majid Al Futtaim. The first transaction outside the US happened in Dubai. Mastercard is not bolting an agent onto a payment flow, they are designing an entirely new payment primitive for a world where the buyer is not a person. Tokenised credentials, transparent consent, programmatic spend limits, fraud signals tuned for agent behaviour rather than human behaviour. This is closed-loop infrastructure built by the people who already run the rails, and that is exactly the right posture for the first wave of this technology. Trust at the network level, not bolted on at the merchant level.

Stripe is moving on the same problem from the developer side, building agent-toolkit primitives so that any application can authenticate an agent, scope its spending, and settle a transaction without dropping the agent into a checkout page designed for human clicks. Link is doing similar work on the consumer wallet side. The shapes are different but the conclusion is the same, payments need a new contract for a non-human buyer, and the question is no longer whether that contract gets written, but who writes it and how widely it gets adopted.

My read is that the closed-loop networks win the first wave. They have the trust, the regulators on speed dial, the merchant relationships, and the fraud infrastructure. The open protocol approach matters more in the long term, but in the short term, if you are a merchant and you want agents to be able to buy from you with confidence, you want to be on rails that already have a hundred-billion-dollar fraud apparatus behind them.

There is a another path worth taking seriously, and that is bypassing the card rails entirely. Stablecoin infrastructure was, almost by accident, designed for exactly this. A wallet is a private key. A transaction is a signed message. There is no checkout flow, no merchant account, no card-not-present fraud category, no chargeback window, no human in the middle being asked to confirm anything. An agent with a scoped key and a spending limit can settle a payment in seconds, on a public ledger, with deterministic finality and a receipt that is the transaction itself. USDC, USDT, and the rest of the regulated stablecoin stack already do tens of billions in monthly settlement, and the same primitives that make them work for cross-border B2B make them ideal for agent-to-agent and agent-to-merchant flows. The interesting question is not whether agents will use stablecoins. They already do, in narrow developer contexts. The interesting question is whether merchants integrate them as a first-class payment option or keep treating them as a crypto sidecar, because the agents will route to whichever rail has lower friction, and lower friction for an agent means fewer humans involved.

Marketing is the second domino

McKinsey published a piece in April 2026 estimating that agentic AI will eventually power around 60% of tasks across the marketing process, with the highest concentration in creation (70%) and execution (70%). They identify six agent archetypes: content generator, knowledge, localisation, analyser, planner, and operator. Their estimates for the impact: 10x to 15x acceleration on campaign creation cycles, 10% to 30% revenue lift from hyperpersonalisation.

Believe the directional point even if you want to argue with the specific numbers.

The honest take: most marketing teams are not even close to this yet. They are using AI as a faster typewriter. Generate a headline. Make a hero image. Draft a LinkedIn post. Call it transformation. McKinsey has a phrase for this that I think is exactly right, the gen AI paradox: the technology is everywhere except the bottom line.

The shift that matters is when you stop using AI to do tasks faster and start letting agents own entire workflows. Not AI helps a copywriter draft a brief. An agent generates the brief, generates the variants, runs them against a synthetic audience, picks the winners, builds the campaign, deploys it across channels, monitors performance in real time, kills the underperformers, reallocates budget, and reports back. The human in the loop is setting the brief and adjudicating taste, not pushing pixels.

Most CMOs say they are experimenting. Far fewer have rebuilt a single workflow end-to-end. The gap between those two groups is going to be the gap between the marketing organisations that survive the next three years and the ones that get quietly disassembled.

Agents do tasks. Humans set the brief.

The lazy version of this argument is agents are better at repetitive tasks, humans are better at creativity. That is not useful, and it is also not true. Some humans are bad at creativity. Some agents are surprisingly good at it. The real distinction is sharper.

Agents are better when the task has a clear success function. Find the cheapest direct flight to Cape Town leaving Friday after 6pm. Identify which inbound leads match our ICP and route them to the right rep. Detect anomalies in the server logs. Anything where you can write down what good looks like in a way another system can verify, an agent will eventually do better than you, faster than you, and at lower marginal cost.

Agents are better when the cost of trying is near zero. Running 50 ad variants instead of 5. Generating 100 metadata candidates and picking the best. Polling an API every 30 seconds to catch a state change. The economics of try everything only work when the cost per attempt is small enough to be ignored. Agents make the cost per attempt small enough to be ignored.

Agents are better when the bottleneck is attention, not judgment. Overnight monitoring. Multi-timezone coverage. Anything that benefits from being always-on. Humans get tired. Humans take weekends. Agents do neither, and that alone is enough to make them better at a surprisingly large number of jobs.

Humans are better when the task involves defining the success function itself. What good means. What the brand stands for. What to bet the company on. Whether to fire someone. Whether to launch in a new market. Whether the agents output is actually any good. These are not tasks. They are judgments, and judgment compounds with lived experience in a way that no model has yet replicated. The job of a senior person in an agentic organisation is not to do the work. It is to define what the work is for.

Get this distinction right and the org chart starts to look different. Get it wrong and you either over-automate the things that needed taste, or you under-automate the things that were burning attention you should have been spending elsewhere.

Bullshit Metrics That Sound Smart (BSMSS)

Every technology shift produces a fresh batch of metrics designed to make decks look intelligent. The agentic wave is going to be no different, so let me get ahead of it and propose a few of my own. Some of these are useful, some are at least more honest than what we have now. All of them are easier to put on a slide than to defend in a board meeting, which is exactly what you want from a good BSMSS.

Return on Token Burn (RoTB). Output value divided by inference cost. I learned this one the hard way when I was burning $15 a day having an agent check my email, because I had pointed a frontier model at a task a 7B parameter open-weight model could have handled in its sleep. RoTB forces the question every team is going to have to answer: are you using the right model for the job, or are you using the most expensive one because it makes you feel better.

Agent Task Completion Rate (ATCR). Of the tasks an agent attempts, what percentage finish without human intervention. This is the honest version of automation rate. Most teams quoting an automation number are quietly doing a lot of human cleanup behind the scenes. ATCR makes that visible.

Time to First Useful Action (TTFUA). From input to the first action that has a consequence in the real world. Not first response. First action. The difference matters. A chatbot that replies in 200ms but takes six exchanges to do anything is slower than an agent that thinks for 30 seconds and then files the ticket, books the room, and pays the invoice.

Agent Discoverability Score. For brands and merchants, how well-structured is your data for agent consumption. The new SEO. If your product page is a beautifully designed React app with no structured data, no clean API, and no machine-readable inventory feed, your discoverability score is approximately zero, no matter how much your homepage cost.

Human-in-the-Loop Cost (HITL). Cost per human intervention. Tells you whether your agents are actually saving time or just shifting the work to a different person at a different desk. A workflow with a 99% ATCR and a $200 HITL is genuinely automated. A workflow with a 70% ATCR and a $5 HITL is a junior employee with extra steps.

Pick two or three. Put them on a dashboard. Argue about which numbers are moving. The point is not that these metrics are perfect. The point is that the existing ones, conversion rate, CPM, CPL, CAC, were built for a funnel where the buyer was a human. The funnel is changing. The metrics need to change with it.

Parting shot

Software has had one user for 50 years. It is about to have a second one, and within a decade, that second user will be doing more transactions than the first.

The companies treating agents as a niche use case are the companies who treated mobile as a niche use case in 2008. They will spend the next few years explaining why their roadmap does not need to change. They will lose, slowly, then all at once.

The agent that booked my tickets did not feel anything when it confirmed the seat. It does not care if the booking page has a parallax scroll. It cared about the price, the terms, and the API. That is the customer of the next decade, and you do not get to opt out.

Sannah Rajpurohit is a software engineer and management consultant helping banks do AI in East Arabia for Mastercard. Sid Wahi runs Thynker, an AI consultancy focused on several sectors, including media.