The Ghost of Cluetrain

In 1999, the Cluetrain Manifesto famously declared that "markets are conversations." It was an inspiring, romantic notion that promised to democratize commerce, wresting power from faceless corporate monoliths and handing it back to a sovereign consumer. Fast forward to today, and that conversation has been thoroughly co-opted. What was supposed to be a bilateral dialogue has devolved into an automated, highly-optimized monologue. The emergence of agentic AI, which features autonomous software agents supposedly operating on our behalf, promises a return to that original democratic vision. But let us be honest: is this actually a revolutionary shift, or is it just another iteration of vendor-controlled slop designed to monetize our decisions before we even make them?

The dream of conversational commerce was simple: technology enables humans to speak to other humans at scale. Instead, the vendor community realized that humans are expensive, inconsistent, and prone to demanding fair treatment. The corporate response was to replace them with IVR systems, chatbots, and automated messaging. These tools were never designed to foster actual conversations; they were designed to create efficient deflection barriers. Now, we are told that generative AI and agentic systems will change all this by acting as our personal proxies. But will it come true?

TL;DR

If you want to watch the full CRMKonvo, please go ahead here (optimized for smartphones) or here (optimized for tablets/computers).

Else, be my guest and continue to read.

Or do both …

The Illusion of Agentic Agency

During our recent CRMKonvo with Dan Miller, founder of Opus Research, we wrestled with this paradox. We have been apocaloptimists when it comes to conversational AI, marveling at the technology's ability to improve our lives while ignoring its potential as a tool for corporate surveillance. The simple truth is that the economic incentives of surveillance capitalism remain unchanged. When a vendor provides you with an "autonomous assistant" to help you shop, that assistant is not working for you; it is a digital Trojan horse. It is programmed to maximize the vendor's margins, steer you toward high-commission partners, and dynamically adjust prices based on your historical data. They call it serving you better; in reality, it is just more sophisticated extraction.

This is where the asymmetry of power becomes glaringly obvious. The consumer enters the arena with a simple objective: to find a quality product at a fair price. The vendor enters with predictive algorithms, historical CDPs, and agentic bots designed to extract the maximum possible lifetime value from that specific consumer. When these two forces meet, it is not a conversation; it is a “negotiation” where one party has access to the other's entire cognitive blueprint. If your personal shopping agent is hosted, managed, or trained by the same corporate infrastructure it is supposed to negotiate against, your agent is effectively a double agent.

The Guardrail Paradox and the Friction of Safety

One of the most fascinating aspects of our discussion centered on the concept of guardrails. In theory, guardrails are designed to protect users, prevent systemic bias, and ensure compliance. In practice, they are a friction point. If you are a malicious actor, or a vendor looking to maximize short-term profit, you do not want guardrails; building and maintaining them requires computational and human effort. Consequently, the path of least resistance is to deploy systems with minimal oversight and dealing with the possible fallout later. When they put restrictions in place, they often reduce legitimate user choices instead of protecting the user.

This creates a bizarre scenario where the consumer is locked in a digital playpen, restricted by strict guardrails on what their agent can ask or do, while the vendor's algorithms roam free in the wild west of data exploitation. The guardrail paradox is that by trying to make AI safe, we often make it useless for the consumer while doing absolutely nothing to stop the vendors from using unbridled models to leverage market dynamics. It is an asymmetric conflict: the defensive side must comply with every rule, which are set by the offensive side.

The 'Trusted Agent' in a Corporate State

Senator Warner and others have proposed regulatory frameworks that would authorize approved entities, be they banks, credit card issuers, or the vendors themselves, to host "trusted user agents." This is a farce of epic proportions. How can anyone believe that a vendor-hosted agent will prioritize the consumer's interests? The Martech community has spent decades building systems to capture, analyze, and exploit user data. To expect these entities to host an objective, consumer-first agent is akin to asking the fox to protect the chicken coop.

Such proposals do not democratize AI; they institutionalize the power dynamic favoring the vendor, dressed up in the shiny new clothes of trusted agentic technology. The vendor-hosted agent will inevitably suffer from a conflict of interest. It will prioritize the products that yield the highest margin, mask competitive alternatives under the guise of "simplifying choice," and feed our preferences back into the corporate data lake. True consumer agency cannot exist within a closed corporate ecosystem. It requires independent, decentralized, and locally run models that answer to no one but the individual user.

Pay-to-Play Algorithms and the Opacity of LLMs

Let us look at a concrete example of how this plays out in the real world. Generative AI led to the discipline of GEO (generative engine optimization) to ensure being highlighted in search feeds and assistant recommendations. This is the reality of a black box. When you ask a modern LLM for a product recommendation, you have absolutely no way of verifying why it chose a particular vendor. There is no transparent ledger of recommendations. It is entirely possible that the recommendation you receive is the result of an agreement between the LLM provider and a corporation.

As long as these models remain opaque, any promise of objective personal assistance is a marketing myth. The algorithms are trained on data that is already heavily skewed by advertising dollars and SEO manipulation. Therefore, when an agentic bot uses it, it is recycling corporate propaganda, presenting it as unbiased advice. This is not artificial intelligence; it is automated salesmanship. To combat this, we need absolute transparency in how recommendation engines operate, including a public ledger of all corporate sponsorships and algorithmic biases that influence the output. A tall order.

The Scalability Farce of Manual Compliance

Even if we establish clear privacy guidelines, such as the right to be forgotten or standard opt-outs like in the GDPR, e.g., implemented using the IEEE My Terms standard, the enforcement mechanism is broken. If a consumer requests that their data be deleted or excluded from training sets, how do they verify compliance? They cannot. If you send a compliance request to a trillion-dollar tech company, that request likely lands on the desk of an understaffed compliance team using a manual process to scour databases, call transcripts, and unstructured chat histories. This does not scale. It is impossible for these enterprises to manually comply with millions of granular privacy requests.

The vendor's SOP will be to say they complied. Yet, once your data has been ingested into an LLM, it is practically impossible to "un-train" that model on your information. The data becomes an inseparable part of the algorithmic weights. Therefore, any regulatory framework that relies on retroactive compliance is a toothless tiger. We must shift the battleground from retroactive deletion to proactive, systemic prevention.

VCONs and the Architecture of True Data Sovereignty

If we want consumer agency, we must shift the paradigm. This is where technologies like Virtual Conversations (vCon) become critical. A VCON is a standardized, secure digital container that houses the transcript, audio, and metadata of a conversation. Crucially, instead of relying on a vendor's pinky-promise to respect our privacy, the data itself is encapsulated with its own governance rules. This is a step toward true data sovereignty, but it requires a massive cultural and technical shift.

Conclusion: Taking Back the Loop

The term "human-in-the-loop" is frequently used to describe safe AI integration. But as agentic AI evolves, we are moving toward a world where humans are removed from the loop, replaced by autonomous agents transacting with other autonomous agents. If we do not demand models that genuinely operate on our behalf, we will find ourselves shut out of our own decision-making processes. This is time to stop being passive consumers of AI convenience and start being active architects of our digital autonomy.

Pragmatic Playbook for Enterprise CX Buyers

Enterprise buyers are currently being bombarded with vendor pitches promising that agentic AI will magically solve their customer experience woes. If you are a buyer and concerned about ethical AI use, here is your survival guide to avoid making an expensive, possibly brand-damaging mistake:

Prioritize Architectural Integrity Over Hype: Do not be seduced by an agent's ability to generate natural-sounding excuses. Demand to see the integration map. If the agent cannot access your back-office CRM and ERP data securely and deterministically, it is not an agent; it is a glorified chatbot with a larger vocabulary.

Mandate Strict, Verifiable Data Boundaries: Ensure that your customers' data is never used to train a vendor's public LLM. If the vendor cannot guarantee and prove that your proprietary customer interactions are kept in a secure, isolated RAG environment, walk away. Your customer data is your competitive moat; do not give it away to train your competitor's next model.

Implement 'Agent-in-the-Loop' Safeguards: Autonomous agents are highly efficient at going sideways before they go south. Never deploy an agentic system in a customer-facing role without a deterministic routing mechanism that instantly escalates complex, emotional, or high-value interactions to a well-trained human agent, complete with full conversational context.

Insist on Standardized Metadata and vCon Support: Prepare for a future of decentralized data. Your architecture should support standard containers like vCons to ensure that as consumers demand greater control over their conversational data, your systems can comply programmatically rather than relying on manual, unscalable processes.

The e-commerce industry has been pushing the exact same shopping cart down the exact same digital aisle for the better part of two decades. We have endlessly debated the optimal color for a checkout button. We have deployed massive Customer Data Platforms (to track users across the web. We have implemented traditional Customer Relationship Management tools and a full-on MarTech stack to send personalized emails that usually find a direct way into the spam folder. Yes, despite all this expensive digital plumbing, the average conversion rate stubbornly hovers around a meager two percent.

Enter the industry's latest shiny toy: agentic commerce. The vendor pitches are certainly alluring. We are moving away from the tedious "click and wait" era into a frictionless "talk and buy" reality. On the surface, it sounds like a massive leap forward. However, any technology analyst worth a grain of salt must ask the difficult questions. Is this actually revolutionary, or is it just a database update with a new coat of paint? Are we solving a genuine consumer friction point, or is this just a solution looking for a problem to help a vendor's stock price?

TL;DR

If you rather want to watch the CRMKonvo, find the mobile optimized version here and the tablet/laptop version here.

Or feel free to read on. Or do both.

The Death of the Traditional Search Bar

Raj Balasundaram, founder and CEO of agentic Commerce vendor Bayezon AI identifies a fundamental shift in consumer behavior. The traditional search engine model is dying. For years, the standard consumer journey has been a repetitive four-step process. You type a query into a search bar. You scroll through pages of results on a Product Listing Page. You click through to a Product Detail Page. Finally, if you have not abandoned your cart out of sheer frustration, you make a purchase.

First social media and now large language models have trained consumers to expect immediate gratification. ChatGPT, Claude, Perplexity and their siblings collapse those four steps into a single interaction. The consumer asks a question and receives a definitive answer. Balasundaram argues that consumers will no longer tolerate the friction of traditional website navigation. They want the low friction of an autonomous shopping agent to handle the discovery, the comparison, and the checkout.

This is a compelling narrative. The architectural integrity of traditional e-commerce platforms is entirely predicated on catalog browsing. If the consumer refuses to browse, the entire tech stack becomes obsolete overnight. However, while building a chatbot that can mimic human conversation is easy, building an autonomous agent that can reliably execute a financial transaction without hallucinating a non-existent return policy is an entirely different technical challenge.

The Vanishing Brand and the Commodity Trap

Here is where some skepticism becomes necessary. If an AI agent is handling the entire transaction, what happens to the brand? Retailers have spent decades and billions of dollars cultivating unique brand identities. They built their own websites and mobile applications specifically to avoid being commoditized by mega-marketplaces like Amazon, or now by ChatGPT, Claude, or Perplexity.

If a consumer simply asks Google or Claude to find the best power drill for a beginner, the retailer vanishes behind an anonymous digital infrastructure. The brand becomes nothing more than a localized fulfillment center for a massive tech conglomerate. Balasundaram acknowledges this as an existential threat. His solution is that brands must deploy their own proprietary agents to control the narrative and maintain their unique story.

This strategy assumes that consumers actually care about the brand story for every purchase. For luxury goods or highly specialized retail, a bespoke AI concierge probably makes perfect sense. A high-end food retailer just needs an agent that can explain the heritage of a specific organic jam. But for basic commodities, brand affinity is dropping rapidly. If a customer needs a basic black t-shirt or a standard set of screws, they are not looking for a narrative. They are looking for speed and price. Deploying a highly sophisticated agentic platform to sell generic commodities therefore sounds like a gross misallocation of enterprise resources.

Synthesizing Intelligence from Thin Air

The most glaring flaw in the agentic commerce hype cycle is the underlying data foundation. Generative AI is incredibly adept at stringing words together in a plausible manner. It is not, however, inherently factual. To build an agent that can actually sell products, you need a robust Retrieval-Augmented Generation architecture based on consistent data and governance. The LLM must be tightly coupled to a factual database of product information.

When pressed on whether retailers actually have this product data, Balasundaram offered a refreshingly candid admission. They do not.

This is the dirty secret of the AI revolution. Most enterprise data environments are rather messy than clean. Retailers possess basic catalog data like SKU numbers, dimensions, and voltage requirements. They lack the contextual metadata required to answer complex, intent-driven consumer questions. A standard database knows a drill has 18 volts. It does not know if that drill is appropriate for trying to hang a heavy mirror on a concrete wall.

Vendors like Bayezon AI attempt to solve this by synthesizing metadata from instruction manuals and technical data sheets. This is the heavy lifting that marketing brochures conveniently ignore. You cannot build a shiny new AI agent on top of a crumbling data foundation. If you feed garbage data into a sophisticated LLM, you will simply generate highly personalized garbage, and that at unprecedented scale.

Ultimately, agentic commerce has the potential to break the two percent conversion ceiling. Bayezon AI has the data to prove this. The concept of an AI concierge that can negotiate, recommend, and checkout is technically viable. But it requires an architectural overhaul that most retailers are unprepared for. It is not a plug-and-play widget. It is a fundamental rewiring of how a business manages and understands its own product data and its processes.

Reality Check: Navigating the Agentic Minefield

If you are an enterprise buyer sitting in a boardroom listening to a slick pitch about the future of autonomous customer experience, take a deep breath. You are standing on the precipice of making a very expensive mistake. Here are three main learnings and recommendations to keep your strategy grounded in reality rather than generative hype.

Data Quality Dictates AI Success

Generative AI cannot generate facts from nothing. Your new AI shopping agent is only as intelligent as the product data feeding it. If your current inventory database is a spreadsheet of incomplete SKUs and missing descriptions, do not even think about investing in an LLM. You must fix your data plumbing first. Focus your budget on data synthesis and structuring before you buy a conversational interface. Intelligence requires a factual foundation.

Questions to ask the vendor:

• Will your AI solution connect directly to our data? This will clarify if the vendor can use your data where it is today or if data preparation is required.
• How do you mitigate and manage hallucinations? It is critical to know how they prevent the AI agent from sharing misinformation with your customers.

Acknowledge the Integration Realities

Connecting an autonomous agent to your legacy CRM and inventory management systems is not a simple weekend project. It is a grueling integration reality. You will face latency issues, restrictive API rate limits, and constant synchronization bottlenecks. Do not treat agentic commerce as a standalone marketing initiative. It must be a core IT infrastructure project. Plan for a rigorous, lengthy integration phase that goes far beyond a controlled proof of concept.

Questions to ask the vendor:

• How compatible is your AI with our current tech stack? Ensuring compatibility can prevent disruptions and make integration seamless.
• What is the projected timeline for integration? Knowing the timeframe upfront helps align the project with other business goals and resources.

The Human-in-the-Loop is Mandatory

There is a dangerous misconception that AI agents will completely replace human customer service teams. This is a recipe for disaster. No matter how advanced the model becomes, it will inevitably encounter an edge-case query it cannot resolve. Or your customer just wants to talk to a human. You need a seamless, immediate escalation path to a human representative. The goal of this technology is to augment your staff by handling routine transactions. It cannot (yet?) replace a human when a complex customer service issue arises. Keep your humans firmly in the loop.

Questions to ask the vendor:

• When and how does your generative AI agent hand off an interaction to a human agent?
• Can the AI agent ask the human agent for the input it needs to resolve the customer's issue without handing over the interaction to the human?

The digital commerce industry has spent the last twenty-five or so years optimizing a single, unit of measurement: the session. We built cathedrals of conversion rate optimization (CRO), obsessed over pixel-perfect hero images, and deployed armies of "customer success" bots that were little more than glorified FAQ routers. We tracked users from the moment they landed on the homepage, watched them struggle through navigational hierarchies, and celebrated when 3% of them actually bought something.

Anywhere else, a 97% failure rate would be grounds for executive termination. In e-commerce, it was the benchmark for success.

We can safely say that the era of the session comes to an end, thanks to conversational and then agentic commerce, which put the "homepage" on life support. What comes more and more into the foreground is the intent, whichis what the session was supposed to help derive. And crucially, the entity expressing that intent is increasingly likely to be a machine, not a human.

What we are seeing now is the transition from browser-based commerce, where humans operate interfaces, to agentic Commerce, where AI agents operate APIs. This isn't just a channel expansion like conversational commerce; it is a fundamental inversion of the retail power dynamic. In the browser era, the retailer controlled the environment. In the agentic era, the customer (or their proxy) controls the context. This is quite similar to what happened in the 2000s with the advent of social media. And it will likely be countered by vendors as fast as the power shift back then, e.g., using GEO instead of SEO.

The demise of the search box

Since the rise of Google, the search box was the primary interface for intent. According to findings by Bain and Company, it gets increasingly exchanged by generative AI. 30 to 45 per cent of US consumers already use generative AI for product research and comparison with 17 per cent stating that they start their (holiday online) shopping with ChatGPT, Perplexity and co. This basically says that the traditional search box is becoming obsolete and replaced by the prompt. While this doesn’t look dramatically different, the use is different. Instead of punching in some disjointed keywords, we can now use real phrases to express complex intents, like "I need a sustainable gift under fifty dollars for a coworker who loves cooking, feels premium, and arrives by Friday".

A traditional e-commerce engine likely chokes on this request. It sees "cooking" and shows a spatula, or maybe a pot? It misses "sustainable" because that data is buried in a PDF product description, and it misses "feels premium" because that is a sentiment, not an attribute. An AI agent can parse these constraints and orchestrate a query across the catalog and across vendors.

If a commerce architecture cannot serve this kind of "headless" request from an AI agent with the same fidelity as a human visiting a homepage, it is effectively closed for business because it is still optimizing for SEO when it should be optimizing for GEO (Generative Engine Optimization), which is the art of making your data palatable to an LLM.

To make this work, it needs two key ingredients: Discoverability, which needs to be changed from search centric to language centric. For this, it needs protocols, which are emerging, the Agentic Commerce Protocol ACP from Open AI and the Universal Commerce Protocol UCP from Google, the former more checkout centric and the latter mor centering around interacting entities. All this needs to support the gamut from simple agent to site to brokered agent interactions that include a buyer agent, a seller agent and a broker agent.

The other one is far more basic: interoperability, which is the domain of the model context protocol and the agent 2 agent protocol.

Until around the second half of 2025, OpenAI had plugins, Google had Actions, and none of them talked to each other. Recognizing that fragmentation kills adoption, and that it needs a kind of a “USB port” for commerce, the industry got to its senses and moved toward standardization with the formation of the Agentic AI Foundation under the Linux Foundation.

So, one can say that the basics are sorted now.

The discovery portion is still up for grabs.

This is also the portion where it will be decided whether the customer stays in control or not.

However, …

Salesforce, Cimulate, Microsoft, Algolia and the Discovery Imperative

There is lots of movement in this area. Algolia partnered with Microsoft, Salesforce acquires Cimulate, other vendors build on own agents, protocols and strength. All of them rely on strong ecosystems.

Every single one of them, plus probably some more, are likely to play an embrace, extend, extinguish strategy to gain the upper hand in this emerging market. EEE is basically about achieving vendor lock-in by polluting standards: vendors don’t win by building a better product on a level field but by changing the field so rivals can’t interoperate without copying their proprietary stuff. Having said that, not every extension is evil. Extending a standard can be legitimate innovation if it’s standardized back upstream or remains interoperable. HTTP cookies are an example for this. Extending a standard becomes EEE when the extensions are used, especially by a dominant player, to make competitors incompatible and to shift the ecosystem onto proprietary rails.

What does this mean for the enterprise buyer?

As usual, the path forward is fraught with traps, which makes it important that retailers maintain immediate control of their end points and do not rely on intermediaries. Discovery is migrating off your site, into agent surfaces. So, do not buy the "Agentic Suite" just because a vendor bought a startup last week.

Some recommendations:

Stop thinking in channels and become channel agnostic, or headless, if you will. Your commerce logic from pricing, inventory, to catalog, must be decoupled from the presentation layer. If your Add to Cart function is tied to a JavaScript button on an HTML page, an AI agent cannot trigger it. Even your customer that comes via WhatsApp, will not be able to trigger it. You need an API-first architecture where the "channel" is just an implementation detail.

Structure Your data to embrace GEO. Your catalog is probably a mess. An agent doesn't care about your "whimsical fall vibes" marketing; it cares about structured attributes. Implement a PIM that supports vector embeddings and maybe publish an /llms.txt file on your domain. Invite the agents in; don't make them scrape.

Implement agent-ready security. If an agent is buying, who is checking the ID? We are already seeing prompt injection attacks, where malicious inputs hijack autonomous agents. Adopt the Agentic Commerce Protocol (ACP) to ensure you aren't holding the bag for a rogue agent.

Measure the right things. Measuring NPS doesn’t matter for bots. It just doesn't have feelings. Measure goal completion rate (GCR) and similar. If you are optimizing for session time, you are optimizing for a ghost.

Prepare for MyTerms. As I've written previously, the consent banner might have found a successor. We are (hopefully) moving toward IEEE 7012-2025 ("MyTerms"), where the user’s agent negotiates privacy terms with your site automatically. If you don't support this machine-to-machine negotiation, you might be avoided by the privacy-conscious agent.

We are leaving the era of the operator and entering the era of the orchestrator. The new front door is invisible. It is an API call from a user's agent to your agent. For the unprepared retailer, this may very well be an extinction event. For the channel-agnostic retailer, it is the moment where scale can decouple from headcount.

Welcome to 2026. The agents are here. Try not to let them automate your chaos.