Could you have AI sales agents that buyers would actually appreciate?

A couple of years ago, I surveyed marketing leaders about sell-side AI technologies. But I flipped the script. When they were on the receiving end — as prospects themselves — which ones did they think would make their experience as a buyer better or worse?

By a large margin, AI SDRs were viewed the most negatively — 59% to 64% unfavorably, depending on whether the outreach was inbound or outbound. Marketers might have been fine deploying them in pursuit of their own lead generation goals. But they did not relish being the target in the target function of these new AI automatons.

Yet as negatively as these same marketers felt about AI SDRs selling them, they were overwhelmingly enthusiastic about AI that would enable sales reps to better answer their questions — 72% expected that would make their experience as a buyer better.

The interpretation is obvious: buyers want to buy, they don’t want to be sold.

And you know it to be true.

In the software industry though — especially in martech — the Camelot dreams of inbound often take a back seat to outbound, in varying degrees of cold or warm. To be fair, it generates pipeline. But the math is brutal. The vast majority targeted are not in market, not a fit, not interested. Constantly being sold, every week on every channel, by hungry vendor upon hungry vendor, is a numbing experience.

Of course, this was happening before AI. But thanks to the brazen advertising campaigns of a few first-generation AI SDR companies, everyone saw the flood coming. Outbound maxxing.

The irony is that when buyers are interested, they often find the process more frustrating. IDC recently published a blunt list of five practices that push your best buyers out. They can’t get the answers they want, when they want, without being chaperoned. One more MEDDIC, and they’re going to need a medic.

I recently connected with Jonathan Kvarfordt of 1mind. I had always appreciated the company’s positioning that emphasized the buyer’s experience over solely seller-centric objectives such as reducing costs and juicing top-of-pipeline metrics.

It’s not to say that seller objectives don’t matter. But a seller’s real objective is to acquire and grow happy, successful, profitable customers.

I got my start in martech in the field of conversion optimization. How do you drive clicks on an ad? Form fills on a landing page? Responses to an email? The most important lesson I learned was the danger of local optimization over global optimization.

You can play all kinds of tricks to win a click. But if it’s not aligned with the rest of the buyer’s journey, it’s a false win. Worse, blindly boosting that vanity metric is probably working against the ultimate customer outcomes you’re aiming to achieve.

In AI SDR land, a version of this local optimization trap is the meetings booked metric.

You know that nightmare scenario, where an AI tasked with producing paper clips wipes out all of humanity by absorbing all the earth’s resources in mindless pursuit of that goal? It’s nowhere near that dire! But whether inbound or outbound, an AI SDR optimizing for meetings booked will happily fulfill that goal — even if a meeting is not a good fit, even if it’s likely to result in a no-show, even if the buyer would have preferred to get what they were looking for without a meeting.

When I asked Jonathan what he thought an ideal metric was for an AI agent, if you were to optimize for both the buyer and seller, he responded: sales process coverage. The further a buyer can get in driving their journey, faster, with fewer friction points, the better for everyone.

A buyer isn’t prevented from engaging with a human seller. But it’s when the buyer chooses, when a meeting has a purpose that isn’t better fulfilled through self-service.

Not surprisingly, that approach correlates with faster sales cycles and higher win ratios. Without the collateral damage of pelting the uninterested or annoying your best prospects with artificial gates that just slow them down.

Okay, that sounds great. But that’s a more sophisticated AI agent and a more transformational change to sales operations. What does that look like? What’s necessary to make it successful?

I agreed to research and develop the following frameworks to help answer those questions. While I did this as a project for 1mind, my goal was to create something vendor-agnostic that would be useful to GTM teams regardless of the particular products they were considering.

First, let’s distinguish between different kinds of buyer-facing agents. Even the label “AI SDR” has come to have a pretty diverse range of interpretations.

I’ve mapped out five revenue agent archetypes that progress by level in sophistication:

I’m using the term “revenue agent” because it has fewer preexisting associations and because as the capabilities become more sophisticated, it’s more than “sales agent.” Advanced revenue agents help coordinate and orchestrate across internal teams and multiple buyer touchpoints.

As a baseline, I think of the original chatbots on websites as Level 1. Prospects could ask questions, and the chatbots tried to answer them, mostly with structured FAQs. They also tried to capture those visitors as leads, albeit often ham-handedly. Let’s be honest: they kinda sucked. But they were the genesis of the buyer-facing agent.

Level 2 is the modern incarnation of conversational website agents. Powered by conversationally adept LLMs, smarter inference, and access to richer unstructured data — not just FAQs, but full knowledge bases, product documentation, and sales enablement resources — these agents deliver a much more useful buyer experience

However, they’re mostly reactive. They aren’t intelligently managing the interaction as a sales opportunity. They’re not coordinating with the rest of the sales process.

I put today’s stereotypical AI SDR at Level 3. These agents are proactively striving towards a sales objective. They are designed to bring the prospect into the company’s sales process, by capturing intent, doing basic qualification, scheduling meetings or demos as next steps, and routing those leads to the right human rep.

The benefits of the agent’s capabilities at this level largely accrue to the seller. They move faster than human SDRs, which can be a good thing for buyers too. But the buyer experience and ultimate revenue impact depend a lot on the overall sales process and objective function the agent is optimizing for. As noted, meetings booked probably isn’t the right goal on its own.

At Level 4, the agent becomes more functional as a member of the revenue team. It can handle objections, deliver personalized demos, update systems, trigger workflows, and execute a more informed handoff to a human seller. It can also continue to participate in those calls with the human rep, in a “ride-along” fashion, to provide instant answers to deep technical questions.

Remember the 72% of marketers-as-buyers who overwhelmingly said AI that would enable sales reps to better answer their questions would make their experience better? This is the way.

The distinction in Level 4 is not merely that the agent can do more things. It can remain involved across a larger portion of the buying process.

A buyer might begin by asking a product question, explore how the product would work for their use case, see a tailored demonstration, investigate pricing, involve additional stakeholders, and only then connect with a salesperson who inherits the full context of everything that has already happened. Less friction from sell-side organizational amnesia.

That greater capability requires much deeper product knowledge, stronger integrations, smarter escalation paths, and more serious governance.

Level 5 is the ultimate vision: a revenue agent capable of coordinating an autonomous customer experience across the full customer lifecycle and the different teams, systems, and touchpoints inside the GTM organization.

Admittedly, this is aspirational today. The technology isn’t there yet. But the bigger hurdles are that most organizations aren’t ready for it. Their current GTM processes aren’t fully connected and integrated — if they’re even all well-defined, which, in many cases, they’re not.

But the technology is advancing at AI speed, and the work that organizations put into maturing their capabilities with AI SDRs (Level 3) and more sophisticated AI revenue teammates (Level 4) builds the muscles that will enable them to take advantage of the Level 5 generation of revenue agents as they arrive.

The archetype model answers one question: how capable and expansive is the agent intended to be? This second framework addresses a different question: what capabilities must be in place for the agent to deliver that experience reliably?

It is tempting to evaluate an agent primarily by having a conversation with it. Does it sound natural? Does it give an impressive answer? Can it parry a clever objection?

Those things matter. But conversation quality is only one layer of the system. And with today’s LLMs, often the easiest to deliver.

A revenue agent might be articulate while citing inaccurate product information. It might understand a buyer’s intent but be unable to take any useful action. It might take actions while leaving no audit trail of what it did or why. It might dazzle in the pilot and quietly deteriorate as products, policies, pricing, and competitors change.

How do you avoid these failure modes in production?

This capability ladder identifies seven layers to evaluate together:

The foundation is infrastructure. The agent must be fast, available, secure, observable, and economically viable at the volume it will operate.

In a conversational experience, latency isn’t a technical metric. It is the customer experience. So are outages, authentication failures, and unpredictable performance. And if each useful conversation costs more than the value it creates, you don’t have a sustainable capability.

The second layer is knowledge. Does the agent accurately understand the company’s products, policies, pricing, competitors, customer examples, and constraints?

Generative AI makes it remarkably easy to produce an answer. Producing the correct answer, grounded in current and approved sources, with a high degree of reliability, is another matter.

“Confidently wrong” becomes a critical failure mode when the agent is discussing security requirements, contract terms, implementation expectations, or pricing. Those mistakes come with real customer, legal, and reputational consequences.

Knowledge alone, however, does not make an experience personal. That requires context. Does the agent understand the buyer, their account, their likely use case, their stage in the journey, their previous interactions, and the intent behind the current conversation?

Without that context, even an agent with encyclopedic product knowledge treats every visitor like a stranger. Buyers repeat information they have already provided. Existing customers get prospecting pitches. Enterprise accounts get small-business answers. The agent knows everything about the seller and almost nothing about the person it is supposedly helping.

The fourth layer is conversation: the ability to listen, qualify, explain, persuade, and adapt.

This is more than producing pleasant prose. A capable conversational agent recognizes ambiguity, asks useful follow-up questions, explains complex subjects at the right level of detail, acknowledges uncertainty, and adjusts as the buyer reveals new information. Otherwise, you get “scripted chatbot theater” — a generative interface performing a more fluid rendition of a decision tree.

The fifth layer is action. Can the agent schedule, route, update records, trigger workflows, create follow-ups, deliver a demo, and hand off to a person with the relevant history intact?

Action is where an assistant becomes an agent.

But “action” covers an enormous range of sophistication. At one end, an agent that can book a calendar slot is taking action. At the other, an agent that can autonomously run a personalized product demo, adapting on-the-fly to the buyer’s questions and use case, is in a different league entirely.

And that range matters, because the breadth and depth of actions an agent can execute largely determines how much sales process coverage it can deliver. An agent whose action repertoire tops out at scheduling can cover exactly one step of the buyer’s journey: the step where a meeting gets booked. Every other need — a demo, a technical validation, a pricing scenario, an updated record — falls back to a human, with all the delay and friction that implies. An agent that can perform those steps itself keeps the buyer moving.

But action without governance is dangerous.

As agents gain permission to access data, make decisions, communicate externally, and modify systems, organizations need boundaries, logs, approval rules, escalation paths, and performance monitoring.

Governance isn’t a bureaucratic layer bolted on after the interesting work is complete. It is what makes greater autonomy deployable. No organization can confidently hand an agent more authority when nobody can determine what it said, what it did, which information it used, or why.

The final layer is learning. Does the agent improve from customer conversations, rep feedback, win/loss analysis, content changes, and observed performance?

An agent is deployed into a moving environment. Products change. Competitors change. Policies change. The questions buyers ask change. Without an explicit learning loop, the agent doesn’t just stand still. It falls behind — decaying relative to the business around it while its dashboard stays green.

The pattern across everything above is a shift in the question. First-generation AI SDRs asked how AI could scale what sellers already do — more touches, more sequences, more meetings booked.

The better question is how much of the buying journey an agent can genuinely help a buyer through.

Jonathan and I will be digging into both frameworks — with examples, edge cases, and the questions I couldn’t fit in a single article — in a live webinar on August 5 at 1pm EDT. Register here. It’s free, the registration form is short, and no AI SDR will chase you afterward. Promise.

Scott