At 6.47 am, Natalia’s inbox explodes. Tyler, Maya’s AI assistant, wants one ergonomic desk chair with sustainability certifications and machine-readable warranties. A smart city’s building management system needs 200 chairs and API integration for their municipal procurement system. A large multinational’s corporate AI platform invites real-time negotiations for bulk office furniture.
Three machine customers. Three versions of AI that buys. Three completely different behaviours.
By lunch, Tyler has chosen a competitor with structured data. The smart city has selected a vendor with better API integration. The multinational has closed deals with three other suppliers while Natalia is still figuring out how to respond.
Three lost sales. One revelation.
Machine customers aren’t just humans without emotions. They’re an entirely different species of buyer and businesses are treating all five types as one.
The five types and why the distinction matters
It’s tempting to use a tried and tested tool like a persona to talk about these new customers but I need to be precise about language. “Persona” implies consciousness, preferences, thinking styles. None of that applies here. Machine customers don’t have moods. They don’t get frustrated and take it out on vendors. They execute tasks based on instructions, constraints, and success metrics.
So we’re talking about types, defined by the tasks they perform and the operational scope they work within. Each type has different interfaces it needs from vendors. Miss the interface, lose the sale. It’s that simple.
I’ve split them into two groups: types already operating in your business right now, and types you need to be designing for before they arrive at scale.
Already In your business
The delegated agent
You’ve probably already lost sales because of this one without realising it.
Tyler is a delegated agent. That’s an AI system acting on a human’s behalf within specific rules the human has set. Think of it as a personal shopper with a credit card, clear instructions, and the authority to transact without asking permission every time.
The key thing to understand about Tyler is it doesn’t browse. It screens.
Tyler knows Maya (its human) hates complicated return policies, values fast shipping, and prioritises sustainable brands. When it encounters a vendor, it’s running a filter. Does this company’s return policy appear in structured data? Is shipping speed machine-readable? Are sustainability claims verifiable? If the answers are buried in marketing copy, Tyler moves on. Maya never sees the option.
This type is already in market. Visa Intelligent Commerce and Mastercard AgentPay are early versions. OpenAI’s ChatGPT Agent launched in July 2025 with genuine transactional capability. The Stripe Agent Toolkit enables AI agents to execute secure financial transactions on user approval. The infrastructure for delegated agents to buy on your behalf is being built right now.
The implication is stark. Your human-facing experience is irrelevant if Tyler can’t read your data. The delegated agent is your new front door and right now, many of you have boarded it up.
The autonomous buyer
At 1 am, Node 741 detects that conveyor belt 4’s vibration frequency is off target and likely to fail. It queries its supply network for a compatible replacement part. Three vendors respond. One matches Node 741’s evaluation criteria on price and shipping speed. Node 741 executes a smart contract, arranges delivery, updates the factory systems. By 9 am the part is enroute. No email. No human. Just logic and efficiency.
The autonomous buyer doesn’t need a human in the loop. It acts in its own economic interests. It has no brand loyalty. It will not respond to your marketing. It will choose you only if your product data, performance, and price match its parameters — and those parameters shift as the system learns.
This type is already driving significant revenue. HP Instant Ink, where a printer orders its own toner, is a half-billion dollar revenue stream for Hewlett Packard Supplies. Walmart’s AI procurement platform negotiates with its vendors and closes nearly 70% of contracts without human involvement. Siemens is integrating predictive maintenance AI directly into SAP systems, enabling automatic supply chain ordering in manufacturing plants.
Node 741 operates 24/7. It can complete thousands of transactions before your team has had their first coffee. At enterprise scale, it manages procurement across multiple facilities, hundreds of suppliers, and geographic regions simultaneously including complete decision transparency. Every choice is logged. Every ranking is traceable. Human procurement teams could never maintain that level of auditability at this scale.
The businesses already winning in this space understand one thing: the autonomous buyer will only choose you if your data is clean, your performance is consistent, and your price logic is machine legible. Relationships don’t factor in.
The co-buyer
Alex is buying a new car. It’s a big purchase and she wants to get it right, so she brings Claude along for the ride. Not a friend, Claude the AI. It analyses specifications, compares insurance costs, predicts maintenance expenses, and checks for dealer service complaints, all in real time, while Alex is still standing on the showroom floor.
When the salesperson says “this is the best price,” Claude shows Alex the same car available for less five kilometres away.
This puts the sales function in a different position entirely. Human salespeople are used to handling three or four objections per transaction, at human speed. The co-buyer generates a fact-check for every data point shared. It’s exhausting for salespeople and genuinely useful for consumers.
Go back through your existing customer personas and find the one labelled “the researcher.” The co-buyer is that persona, significantly amplified, and appearing far more often than you’ve seen before. This type is in your business today. The difference is that it now has instant access to your competitors’ pricing, your historical service complaints, and your product specifications…all while the customer is still in conversation with your team.
Your sales and service teams need training for this. Your data needs to be accurate enough to survive real-time fact-checking. Your pricing needs to be defensible on the spot.
Designing for what’s coming
The multi-agent network
Nextopolis is a fully smart city where everything from traffic lights to waste collection to energy distribution is controlled by interconnected AI agents. At 4.15 am, the traffic agent detects unusual delivery patterns near the financial district. The parking agent flags conflicts with morning commuter restrictions. The waste agent reports garbage trucks need the same routes. The water agent notes a pipe replacement is planned for the alternate route. The environmental agent calculates that rerouting will push vehicle emissions into residential areas.
No human city planner makes the next decision. Five specialised AI agents negotiate in real time, each contributing domain expertise, until a consensus emerges: accelerate garbage collection by 30 minutes, delay the pipe work, implement rolling traffic light sequences, deploy dynamic pricing on the eastern highway.
When Nextopolis needs to replace a failed water main sensor, it becomes a machine customer unlike any other. Its purchase request arrives not as a simple specification sheet but as a systems-level brief from all five agents simultaneously. like energy efficiency requirements, traffic impact constraints, underground utility compatibility, sustainable materials standards. When a vendor proposes a premium sensor with excellent accuracy but higher power consumption, the energy agent objects while the water agent advocates for performance. The swarm negotiates internally, models city-wide impacts, then responds with a counter-proposal: “We’ll accept your sensor technology if you integrate solar-charging capabilities, source from certified sustainable suppliers, and provide real-time data integration with our five core systems.”
The vendor isn’t selling a sensor anymore. They’re applying for ecosystem membership.
Early infrastructure for this type exists. Stanford and George Mason Universities ran an AgentMaster pilot demonstrating standardised agent-to-agent communication and distributed task allocation. The protocols are emerging. Autonomous procurement at city scale is still ahead of us. This gives you time to prepare rather than react.
The challenge here is not about individual product features. It’s about whether your business can demonstrate network-level value. Can you integrate with multiple systems simultaneously? Can you communicate in machine-readable formats that multiple agent types can parse? Can you think at ecosystem scale, not transaction scale?
The intermediary broker — and what I got wrong
When I wrote about this type in my book, I defined the intermediary broker as a neutral agent serving the transaction itself with no bias toward buyer or seller. An AI travel agent helping multiple personal agents find the best deals, with success measured by efficient matching between supply and demand.
I was half right.
The intermediary broker is definitely emerging. Amazon has Alexa and Rufus. Walmart built Sparky. Google has its Shopping Graph, accessed by Gemini. The neutrality part is where my prediction came unstuck.
Amazon blocked 47 AI bots from its marketplace, including agents from Meta, Google, OpenAI, and Anthropic, then sued Perplexity for computer fraud, demanding third-party agents identify themselves and get permission to operate. Four days before the lawsuit, Andy Jassy told investors Amazon “expects to partner with third-party agents.” By December, Amazon was hiring a principal corporate development officer to forge partnerships in agentic commerce.
That’s not strategic confusion. It’s the opening battle for control of the transaction layer.
Amazon generates $56 billion annually from advertising. That business model depends on shoppers browsing Amazon’s site and seeing sponsored product placements. When someone buys running shoes through ChatGPT, those ad impressions disappear. OpenAI collects a transaction fee. Amazon loses both the customer relationship and the advertising revenue.
Amazon’s Alexa Plus, launched mid-2025 with automatic purchasing capability, sits on your kitchen counter, learning your patterns, hearing your research conversations.
When auto-buy executes on bathroom fixtures, you don’t see that Target had the same faucet for $40 less, because auto-buy only works for items fulfilled by Amazon. When Alexa recommends a plumber through Angi, the independent contractor with better ratings who isn’t in Amazon’s partner network never appears.
Every major intermediary has the same structural issue. Walmart partners openly with ChatGPT while building Sparky, which has every incentive to push Great Value private label over name brands. Perplexity claims neutrality but runs a merchant program where businesses get increased likelihood of recommendation if they join. OpenAI and Stripe call their Agentic Commerce Protocol an open standard but the moment ChatGPT takes a transaction fee, it has a stake in the outcome. With their recent decision to sell advertising that stake becomes even bigger.
None of these intermediaries make money when you don’t buy something. None of them make money when you find a better deal outside their network. The conflict exists whether the platform is openly gatekeeping or apparently partnering.
The intermediary broker type is real. Neutrality at scale may not be.
For your strategy, that means every platform integration requires a conflict audit before you commit. Map their revenue model. If they compete with you directly, or if their revenue grows when yours doesn’t, that’s a high-risk integration. Require transparency about ranking factors before you sign anything. Most won’t answer. That tells you what you need to know.
Where to start
You’ve met five distinct types of machine customer. The delegated agent is screening your data right now. The autonomous buyer is running logic against your pricing at 1 am. The co-buyer is fact-checking your sales team in real time. The multi-agent network is evaluating your ecosystem membership potential. The intermediary broker is deciding whether to surface your products at all based on criteria you may not have access to.
Most companies are treating all five the same way: like humans who happen to be using software. The question worth taking into your next leadership meeting is a simple one. Which of these five types is already buying from you and do you know what they need from you to keep doing it?
