Direct answer: A machine customer is an AI agent that researches, shortlists, negotiates, or purchases on behalf of a human or an organization. Gartner calls them custobots and describes a three-stage trajectory: bound agents that execute rules, adaptable agents that optimize within constraints, and autonomous agents that choose vendors themselves. For B2B marketers the practical implication is immediate: part of your buying committee is already software, and it evaluates you differently than people do.
If you want to see which accounts, people, and agents are researching you right now, Book a demo of Abmatic AI's first-party visitor intelligence.
Key takeaways
- Machine customers are not a 2030 hypothetical. Bound-mode buying automation (auto-replenishment, programmatic ad buying, procurement bots enforcing rules) has existed for years; what is new is LLM-based agents doing vendor research.
- Trade coverage of Gartner research projects AI agents involved in the large majority of B2B purchase interactions by 2028, with trillions of dollars in purchasing influenced. Treat the exact numbers with skepticism; treat the direction as real.
- AI buying agents reward structured data, transparent pricing, plain-HTML content, current API docs, and llms.txt files. They punish JS-only rendering, gated everything, and blocked crawlers.
- ABM changes shape: the account now includes agents, and agent sessions are intent signals you can attribute back to real companies.
- You do not need to rebuild your site. Most of the work is an 8-point machine-readability pass you can finish in a quarter.
What is a machine customer?
Gartner defines a machine customer as a non-human economic actor that obtains goods or services in exchange for payment. The firm's shorthand is custobot, and its research lays out three generations. First come bound customers: agents that execute a human-defined rule, like a printer reordering toner or a procurement system auto-renewing a contract under a spend threshold.
Second come adaptable customers: agents that optimize within constraints, comparing options and picking the best one against criteria a human set. Third come autonomous customers: agents with enough delegated authority to define some of their own criteria, select vendors, and transact with minimal human review.
The projections attached to this trajectory are aggressive. Digital Commerce 360 reported in late 2025 on Gartner research projecting that AI agents would participate in roughly 90 percent of B2B purchase interactions by 2028, influencing something on the order of 15 trillion dollars in purchasing. Coverage through 2026 from Demand Spring, Mirakl, and others shows the term moving from analyst decks into mainstream marketing conversation.
Here is my honest read after watching agent traffic on B2B websites all year: the bound and adaptable stages are already here, and the autonomous stage is further away than the headlines imply. Nobody is letting an agent sign a 36,000 dollar software contract unsupervised in 2026. But agents are absolutely doing the research, building the shortlists, and drafting the comparison matrix that a human buyer approves. That middle step is where marketers win or lose.
This post is the mirror image of the agentic marketing playbooks we usually publish. Those cover agents you deploy to sell. This one covers agents that buy from you, and what they need from your marketing surface to include you in their answer.
How AI buying agents actually evaluate vendors
A human buyer skims your homepage, watches a demo video, and forms a gut impression. An agent does none of that. It fetches pages, parses text and structured data, and scores what it can verify. Understanding that pipeline tells you exactly what to fix.
They read text, not design
An agent evaluating your pricing page sees the DOM, not the gradient. If your capability claims live inside images, videos, or JavaScript-rendered components that never make it into server HTML, the agent records nothing. Plain semantic HTML with real headings is the single highest-leverage format decision you can make.
They reward verifiable specifics
Agents synthesizing a vendor comparison prefer claims they can quote and attribute: named integrations, published pricing floors, explicit feature lists, documented API endpoints. Vague superlatives get dropped from the synthesis because there is nothing to cite. Transparent pricing matters disproportionately here; an agent asked to shortlist vendors under a budget will simply exclude vendors whose pricing it cannot find.
They use machine-facing files when you provide them
The emerging llms.txt convention gives agents a curated map of your most important pages in markdown, the format LLMs parse most reliably. We covered the implementation details in our guide to llms.txt for B2B websites. Alongside it, schema.org structured data (Product, Service, FAQPage, Offer) gives answer engines typed facts instead of prose they have to interpret.
They check your documentation like an engineer would
For technical products, buying agents increasingly fetch API docs directly. Current, public, versioned documentation reads as a trust signal. Docs behind a login read as a gap, and the agent will note the gap in its summary to the human buyer.
What changes for ABM when the account includes an agent
Account-based marketing has always modeled the account as a committee of people: the champion, the economic buyer, the security reviewer. The machine customer era adds a non-human member to that committee, and it changes three operational assumptions.
Agent sessions are account intent
When an agent operated by someone at a target account fetches your pricing page, your comparison pages, and your security docs in a 40-second burst, that is not bot noise. That is a buying committee member doing assigned research. The problem is that most analytics stacks either miss these sessions entirely or lump them into direct traffic, a blind spot we mapped in the agentic dark funnel.
Detection is a solvable problem. Declared fetchers like ChatGPT-User and Perplexity-User identify themselves in user-agent strings and published IP ranges; undeclared automation shows behavioral fingerprints like datacenter ASNs and headless browser signals. Our guide to detecting AI agent traffic on a B2B website walks through the full three-layer method.
Personalization needs a human-or-agent branch
Showing a personalized banner to GPTBot is wasted effort; feeding clean structured content to a buyer-delegated agent is not. This is where Abmatic AI's first-party visitor intelligence earns its place in the stack: it identifies the companies and individual contacts behind anonymous traffic, and distinguishing human sessions from agent sessions lets you route each one correctly. Humans get web personalization and Agentic Chat; agents get fast, complete, machine-readable pages. Both resolve to the same account record, so the intent signal is unified either way.
Your content is now an API for buying decisions
In classic ABM you wrote content to persuade a person. Now the same content also serves as the data source an agent queries when its principal asks "compare these five vendors for me." That means every claim on your site should be structured enough to survive extraction and honest enough to survive verification, because agents cross-check claims across sources and penalize contradictions.
Skip the manual work
Abmatic AI runs targets, sequences, ads, meetings, and attribution autonomously. One platform replaces 9 tools.
See the demo →The 8-point machine-readability checklist
Run this against your marketing site this quarter. Each item is checkable in under an hour; most fixes ship in days.
- Server-rendered HTML for all revenue-relevant pages. Pricing, product, integrations, security, and comparison pages must return their full content in the initial HTML response. Verify with curl, not a browser.
- AI crawlers allowed in robots.txt. GPTBot, ClaudeBot, PerplexityBot, and Google-Extended should not be disallowed on marketing content. You can still opt out of training via content signals while allowing retrieval.
- An llms.txt file at your root. Curate your 20 to 50 most decision-relevant pages with one-line descriptions in markdown.
- Published pricing, or at least a published floor. "Starts at X per year, enterprise tiers available" is enough for an agent to include you in a budget-filtered shortlist. Total silence gets you excluded.
- Schema.org structured data on key templates. Organization, Product or Service, Offer, and FAQPage markup, validated and kept in sync with visible content.
- Public, current API and integration docs. No login wall on reference material a technical evaluator needs.
- Plain-text capability and integration lists. Every claim you want quoted in an agent-written comparison must exist as extractable text, not only inside a demo video or PDF.
- Agent traffic detection and attribution. Instrument your site to classify agent sessions and resolve them to accounts, so machine research shows up in your intent scoring instead of vanishing.
Anti-patterns that make you invisible to machine customers
The inverse list matters just as much, because several common practices actively delete you from agent-generated shortlists.
- Blocking AI crawlers wholesale. Flipping the "block AI bots" toggle on your CDN protects nothing on a marketing site and removes you from answer-engine retrieval. If an agent cannot fetch you, it recommends someone it can fetch.
- JS-only content. Single-page apps that render capability content client-side serve agents an empty shell. Many fetchers never execute JavaScript.
- Gating everything. Pricing behind "talk to sales," docs behind signup, comparisons behind forms. Each gate is a dead end for an agent and a mark against you in its summary.
- PDF-only technical content. Parseable, but second-class. Mirror the substance in HTML.
- Contradictory claims across pages. Agents cross-reference. A feature listed on one page and absent from another reads as unreliable data.
- Treating all bot traffic as hostile. Rate-limit abuse, yes. But indiscriminate bot-blocking rules written in 2022 are now filtering out members of your buyers' committees.
An honest timeline, and what to do this quarter
Skepticism first: the 90 percent and 15 trillion dollar figures are directional analyst projections, not measurements, and the fully autonomous custobot that signs enterprise contracts is not arriving in 2026 or, in my view, 2027. B2B procurement has approval chains, security reviews, and legal negotiation that resist full delegation, and buyers know it.
But the adaptable middle stage does not need to be fully autonomous to reshape your funnel. If an agent builds the five-vendor shortlist and a human picks from it, then making the shortlist is the new first-page-of-Google. That game is being played right now, on your website, in sessions your analytics probably mislabels.
The preparation is refreshingly concrete: run the 8-point checklist, kill the anti-patterns, and instrument agent detection so machine research becomes account intent instead of noise. Teams running this on Abmatic AI get the last mile built in, because the platform already resolves anonymous sessions to companies and contacts, captures first-party intent across web, LinkedIn, ads, and email, and lets Agentic Workflows act on the signal, for example alerting an AE when a target account's agent burst hits the pricing page.
The vendors that win the machine customer era will not be the ones with the flashiest AI positioning. They will be the ones whose sites answer an agent's questions completely, honestly, and in milliseconds.
FAQ
What is a machine customer in B2B marketing?
A machine customer is an AI agent or automated system that researches, evaluates, negotiates, or purchases products on behalf of a person or organization. Gartner's term is custobot, with three maturity stages: bound (rule-executing), adaptable (optimizing within human-set constraints), and autonomous (self-directed vendor selection). In B2B today, most machine customers operate in the first two stages, doing research and shortlisting that humans approve.
Are AI agents really buying B2B software today?
Fully autonomous purchases of complex B2B software are rare. What is common in 2026 is delegated research: buyers asking agents to compare vendors, summarize pricing, and draft shortlists. Since the human decision increasingly starts from the agent's output, being legible to agents already affects revenue even though the final signature is still human.
How do I know if an AI agent is visiting my website?
Use three layers: match declared user-agent strings and published IP ranges (GPTBot, ClaudeBot, PerplexityBot, ChatGPT-User), fingerprint undeclared automation via datacenter ASNs and headless browser signals, and resolve agent sessions to companies with first-party visitor identification. Abmatic AI handles the account and contact resolution natively, and our agent traffic detection guide covers the full method.
Should I block AI crawlers on my B2B site?
Not on marketing content. Blocking GPTBot, ClaudeBot, or PerplexityBot removes you from AI answer retrieval, which is where a growing share of vendor research happens. If training use concerns you, use content signals to allow search and AI input while opting out of training, rather than blocking retrieval outright.
What is llms.txt and does my company need one?
llms.txt is a plain-markdown file at your site root that gives AI systems a curated index of your most important pages. It is inexpensive to create, carries no downside, and improves the odds that agents fetch your best pages first. For B2B vendors it is a sensible default in 2026.
How does ABM change when target accounts use buying agents?
The account becomes a mixed committee of humans and agents. Practically, that means treating agent sessions from target accounts as intent signals, branching site experiences so humans get personalization while agents get clean structured content, and ensuring every capability claim exists as extractable, verifiable text an agent can quote in its recommendation.
Want to see which of your target accounts are researching you through agents this week? See it live with Abmatic AI.
