Your Next Customer Is a Machine
A working guide for marketing leaders whose buyers are starting to send software to shop for them.
You cannot flatter a machine. For thirty years, marketing has been the craft of winning a person's attention. A bright button. A clever headline. A deal that ends at midnight.
That still works on people. The trouble is that fewer of your buyers are showing up as people.
More often now, a buyer types one sentence into an AI assistant and lets it do the looking. It reads the web, weighs the options, and hands back a short list with a reason for each. The person picks one and moves on. They never see your homepage.
The thing that chose you was software. And software does not care how your page looks. It cares whether your facts are clear.
Quick definition: an agent is a program you trust to go do a job for you, like finding the right laptop or booking the right flight. When the agent becomes the shopper, the old rules bend. Here is what is changing, and what to do about it this quarter.
The click is dying, and the machine is taking its place
The web was built for human eyes. Its main reader is changing.
The old internet ran on attention. A platform showed you things, and your eyeballs were the prize. Agents do not have eyeballs, so the tricks built for humans slide right off them.
Here is what an agent ignores:
Urgency banners like "only two left"
Color-coded buttons and clever layouts
A beautiful page with thin facts underneath
What it reads instead is plain data: price, specs, availability, proof. This is not a far-off idea, either:
Gartner projects that by 2028, about 33 percent of business software will include agent technology, up from less than 1 percent in 2024.
Imperva found that automated software passed half of all internet traffic in 2025, the first year machines outnumbered humans online.
The contest used to be for the best-looking page. Now it is for the clearest, most trustworthy data.
Your customer is sending a proxy
There are two kinds of AI helper, and they are not on the same side.
The platform's helper. Built by the store to sell you more. It works for the store.
Your agent. Sits on your side and works only for you. It wants your best outcome, not the store's best margin.
Once your agent is doing the choosing, a quiet conflict shows up. The store wants to sell the item that makes the most money. Your agent wants the item that gives you the most value. Those are rarely the same item.
And the old game of nudging human feelings stops working, because the agent has no feelings to nudge. Gartner expects that by 2028, about 15 percent of everyday work decisions will be made by agents on their own, up from zero in 2024. The buyer is handing the wheel to software.
Agents will buy your product. They will not enjoy it.
Draw one line, and the whole future gets clearer.
Some things are about the outcome. Buying a laptop, booking a flight, reordering supplies. You do not enjoy the search. You want the result. Those are easy to hand to an agent.
Other things are the experience itself. Hearing a song, watching a movie, scrolling a feed for fun. An agent can queue the playlist, but it cannot enjoy the music for you. That part cannot be handed off.
Drawing the line
Some things are worth outsourcing. Some are the point.
| Attribute | Tasks you can delegate | Experiences you keep |
|---|---|---|
| Examples | Buying, booking, subscribing | Music, video, news for fun |
| Where the value is | The outcome | The experience |
| Hand it to an agent? | Yes, easily The result is what you wanted. How it got done does not matter. | No Delegating it removes the only part that was worth having. |
The handoff of the first kind has already started, and the money behind it is real. Adobe Analytics found that traffic to U.S. retail sites from AI tools jumped more than 1,200 percent year over year. Salesforce projects that AI and agents will influence about 263 billion dollars in holiday sales in 2025. If you sell an outcome, the agent is coming for that transaction whether you are ready or not.
There is a three-way tug of war now
The old one-on-one between a person and a platform just became a triangle.
Human to platform. The world we know. Platforms use layout, urgency, and limited-time deals to steer human choices. A Princeton study of about 11,000 shopping sites found roughly 1,818 of these steering tricks, on about 11 percent of the sites.
Agent to platform. A new contest. The agent sends a precise request ("find a laptop under 1.2 kilograms with 16 hours of battery"), and the platform tries to guess the bigger picture behind it to keep some advantage. A quiet game played in data, not pixels.
Human to agent. The tricky part, and it is about being clear. Tell an agent to "buy a reliable laptop," and it might chase the cheapest price while ignoring the brand you trust. So it has to ask you a few sharp questions first.
That last gap is real. On WebArena, a test of real web tasks, early agents finished only about 14 percent of jobs correctly, while people finished about 78 percent. Most of the gap came from half-stated goals. Clear input, better output.
The numbers, side by side
Agentic commerce
The traffic is already machines. The competence is not.
| What the data shows | Figure | Source |
|---|---|---|
| Measured | ||
| Automated share of all internet traffic | More than half 2025 | Imperva |
| AI-driven traffic to U.S. retail sites | +1,200% Year over year | Adobe Analytics |
| Projected | ||
| Business software including AI agents by 2028 | 33% Up from under 1% in 2024 | Gartner |
| Everyday work decisions made by agents by 2028 | 15% Up from 0% | Gartner |
| Holiday sales AI and agents will influence | ~$263B 2025 season | Salesforce |
| The counterweight | ||
| Real web tasks early agents finish correctly | ~14% Against roughly 78% for people | WebArena |
| Shopping sites using steering dark patterns | ~11% Of 11,000 sites studied | Princeton |
Projected figures are vendor and analyst forecasts, not observed results. Add publication dates and links to each source.
Your old scoreboard is lying to you
A better score on paper is not the same as a better result.
For years, the recommendation industry graded itself with lab tests on old, frozen data. Neat to run. A poor match for real life.
The clearest proof is the Netflix Prize. In 2009, Netflix paid one million dollars for an algorithm that scored about 10 percent better on its lab test, then never fully used it, because the lab gain did not turn into enough real value to justify the work.
The lesson is sharper now that an agent is the reader. What matters is whether the answer is correct and useful, not whether it wins a tidy metric. If your team is chasing a number that no longer maps to real satisfaction, you are polishing the wrong trophy.
The machine recommends differently now
The engine under the hood is being rebuilt, and it changes what wins.
The old way worked in two steps: pull a big pile of items, then rank them. The new way does it in one step, generating the answer the way your phone predicts your next word.
This is not a lab dream. Meta scaled a recommender like this to 1.5 trillion settings and reported a 12.4 percent gain in live tests. Kuaishou's OneRec, running in production, lifted watch time by 1.6 percent, which is large at that size.
The takeaway is simple. Recommendations are becoming less like a scorecard and more like a conversation, and the brands with clean, honest facts are the ones the conversation is about.
The old job was to be picked. The new job is to be readable by the machine that does the picking.
The real risk, said plainly
A world of agents needs honest data, or it drifts into strange outcomes.
If agents all chase the same few signals, prices and choices can bunch up in ways no one intended. In a study published in the American Economic Review, simple pricing programs taught themselves to hold prices high together, with no communication and no instruction to do so.
So the shift is not automatically good or bad. It rewards whoever gives the agent something true to hold onto, and it punishes the brand that only ever knew how to charm a human. If you build only for eyeballs and engagement, you are building for an audience that is shrinking.
How to start today
You do not need a new department to begin. You need four moves, in order.
Read your pages the way an agent would. Ignore the design for an hour. Are the facts that decide a purchase clear, structured, and easy for software to pull? Fix those gaps first. That is the new shelf space.
Make one transaction agent-ready. Pick a single buying or booking flow and make it something an agent could complete with no human tricks. Learn from that one before scaling.
Add a real-world measure of success. Stop grading on lab metrics alone. Track one honest signal of whether people, and their agents, got what they wanted.
Watch your agent traffic. Count how much of your traffic and how many of your mentions come from AI tools. Most of it arrives unlabeled, and you cannot manage a shift you are not measuring.
The rules of the internet are being rewritten in code, not clicks. That sounds cold, but the heart of it is old and human. Be clear, be honest, and be worth recommending. Do that, and the machine will make the introduction for you.
Camino5 helps marketing leaders stay visible as AI becomes the front door to discovery. If you want to see how an agent reads your brand today, ask us about the AEO Invisibility Score.
Sources
Gartner, agentic AI in enterprise software (33% by 2028): sdxcentral.com
Gartner, 15% of work decisions autonomous by 2028: getathenic.com
Imperva Bad Bot Report (bots pass half of internet traffic): imperva.com
Adobe Analytics, generative-AI retail traffic (+1,200% YoY): blog.adobe.com
Salesforce, AI and agents to influence ~$263B in 2025 holiday sales: salesforce.com
Princeton, "Dark Patterns at Scale": arxiv.org/pdf/1907.07032
WebArena benchmark: arxiv.org/abs/2307.13854
Netflix Prize, model not fully deployed: techdirt.com
Meta, "Actions Speak Louder than Words" (HSTU): arxiv.org/abs/2402.17152
OneRec, Kuaishou: arxiv.org/abs/2502.18965
Calvano et al., American Economic Review (2020): aeaweb.org