13 minute read

The $2 Debate: Are We Asking the Wrong Questions About AI Agents?

Home » Blog » The $2 Debate: Are We Asking the Wrong Questions About AI Agents?
Home » Blog » The $2 Debate: Are We Asking the Wrong Questions About AI Agents?

The wrong debate

There is a lot of discussion about whether Agentforce’s cost of $2 per conversation is too expensive. It seems to be a sticking point that people can’t get beyond. It’s being cited as the reason that agents don’t or won’t work.  

$2 is not an issue. It shouldn’t be the focus. We certainly shouldn’t be architecting agents to minimize the number of conversations.

If I haven’t lost you already, stick with me.

The agents are coming

Agents are going to be part of the tech landscape for many companies. They won’t replace every application. They won’t replace every job. But they will have a profound impact on how some work is delivered.

We believe that, in 2025, we may see the first AI agents “join the workforce” and materially change the output of companies

Sam Altman : CEO : OpenAI

We are at the very beginning of the agent journey, which started less than 6 months ago for the non-AI, non-techy folks. At Elements. cloud we were playing with agent frameworks and semantic databases over 2 years ago. But at that time it was research and not development. It was not scalable. It was not maintainable. In short, it was not viable as a mainstream technology.

Now it is. But only just.

We are about to witness a major disruption. A platform shift. Let me take you back to the last time this happened. Back in 1999 cloud-hosted enterprise applications started to be available. Salesforce was the vanguard. Gartner’s view was that the incumbent, Siebel, would prevail and fight off the cloud competition from Salesforce as no large corporation would put valuable customer data in the cloud. It was considered laughable that highly regulated pharma or financial services data would ever be in the cloud.  Look where we are now. Veeva and nCino are two of Salesforce’s most successful ISVs.

(Agents)remind me of the early, early internet. Who began and who succeeded are very different parties.

Kara Swisher : tech journalist and author

So it is misguided to judge the capabilities and potential of agents by what is in the market now. We are at the very beginning of this shift. It is very difficult to predict what is possible in the future with agents. So it is easy to sit on your hands because the early predictions of AI replacing jobs entirely within a few years have proven to be wrong. But history has shown us that those who lean in early when there is a major platform disruption will be winners. 

We overestimate what will be delivered in two years and underestimate what will be possible in ten years. 

Bill Gates : ex CEO : Microsoft

I was on Microsoft’s Worldwide Partner Advisory Board for 4 years when they went from on-premise to the cloud. It was ugly. Nothing worked for the new cloud world; product development, licensing, pricing, sales compensation, support, and the 600,000 partners….  But Microsoft invested heavily ahead of the curve, so now look at them. They were one of the few companies that safely transitioned. Salesforce, in contrast, was “born in the cloud”.

Amazon only existed because of the cloud.  Back in 1999, Jeff Bezos appeared on ‘The Tonight Show with Jay Leno,’ where the audience laughed at Amazon’s lack of profits despite its billion-dollar valuation. Bezos, unfazed, smiled and emphasized his long-term vision. Fast forward to the end of 2024 when Amazon announced a quarterly profit of $15.3 billion.

You have to be willing to be misunderstood if you’re going to innovate. 

Jeff Bezos : CEO : Amazon

Now the major players, Salesforce included, need to manage the transition to the agent world. It started with the clear commitment and huge pivot by CEO, Marc Benioff, to Agentforce. Some may say it is too early, but he recognizes the potential and understands the time it takes to change so much of the business. Microsoft is fast out of the gate with its Copilot offering and huge investment in OpenAI. Bret Taylor, ex-Salesforce, is the CEO of a well-funded agent start-up, Sierra.ai. And there are a slew of “born in AI” startups, many of whom will not survive.

The coming wave of generative AI will be more revolutionary than any technology innovation that’s come before in our lifetime, or maybe any lifetime. It is reshaping our world in ways that we’ve never imagined. This AI revolution is just getting started.

Marc Benioff : CEO : Salesforce

How much is an agent?

The current list price of an agent should be considered “as a stab in the dark”. At the moment the adoption and volumes are too low to be a material cost. Naturally, CIOs and their CFOs are terrified of the costs if AI adoption really kicks in for certain use cases. 

However, the price list costs are negotiable with all the vendors as they want to establish themselves as the new platform standard. Customer stories are more valuable than revenue. What is clear is that there is no established agent charging mechanism yet. Different vendors have set out their starting positions, which are very different.

As every customer wants to minimize cost there is a risk they will architect agents to optimize the metric that is being charged. Whilst this is understandable, it really only makes sense once the charging approach is established. You could build Agentforce agents optimizing per conversation, and then the established approach is by outcome. That could mean completely re-engineering all your agents.

Until the charging model has been established, it is best to optimize agents for the user experience and use this period for experimentation and learning. Early adopters will get preferential treatment and pricing. 

Interestingly Aaron Levie, CEO of Box.com laid out 4 charging mechanisms for the agent platform vendors in a recent blog. They are paraphrased below:

1 – Leverage the relationship between AI Agents (digital labor) and traditional work (human labor). 

Agents are priced like humans but at a discount. An AI Agent performs a certain amount of work, and you pay for the amount of time or units it took to do that work. The discussion is then on how to define and measure the unit. Agentforce is per conversation, and a conversation starts when the user initiates it and ends when the user stops or 24 hours from initiation – whichever comes first. Microsoft Copilot is based on the message, where a message is one back and forth between the agent and the user. 

2 – Look at the outcomes and charge on a results basis. 

Whilst this sounds equitable, it is difficult to match the value of the outcome and the cost. Some outcomes could be hugely valuable in terms of human labor costs saving or improvement in customer experience. Others may play a supporting role for a human. So if you offer different types of outcomes, you need pricing models for everyone. One benefit is as underlying AI costs drop over time service providers can extract more margin for this work. Sierra.AI does not publish its pricing but its website suggests it is tied to outcomes

3 – Price as close to the underlying AI costs as possible. 

This is likely to be the lowest cost for a customer and could drop over time. This can be great for technically savvys customers but makes it difficult for the vendor to generate a ton of value over time. It does not factor in any compensation to the vendor for building the platform. This is how the LLM vendors are pricing  – by token – and it has been a race to the bottom driven by Meta open-sourcing their Llama model. The LLM providers are using the same pricing for their agents.

4 – Maintaining a pure SaaS seat subscription model. This is offering Agents to users that do unlimited work attached to a seat. Depending on the use case – and how many seats the customer would need – this model could be quite disruptive. In areas where there are lots of seats used by end-users, it’s possibly very strategic; in areas where there’s only a small number of seats, you’re likely giving up too much value.

At Elements.cloud our AI capabilities do not have a consumption-based pricing component. We decided that a consumption charge was too difficult to negotiate, measure, and invoice. So we absorb that cost as part of the annual license fee. This has proven to be a smart move as AI token costs have dropped dramatically, so our exposure is reducing. 

It’s the learning, not the price

Over time the cost of running agents will come down as use cases emerge, the results get better, and the market economy kicks in. We’ve already seen this with the cost of AI tokens. At the moment we’ve seen different vendors with wildly different mechanisms. It is easy to focus on Agentforce’s $2 per conversation and the additional per-user license costs and say that agents are not too expensive. But that is missing the point.

If we only focus on cost, then there are clear use cases where there is a strong ROI. But the current pricing models, many use cases are not viable. But that should not stop us from exploring them: 

  • An AI-driven conversation for an employee to complete an opportunity is probably far more innovative and fun than entering data into a screen. It might even provide more accurate results. But it may be difficult to justify the $2 cost vs the time saved. 
  • Using AI for an employee to create a new vacation booking record in a custom object is definitely not worth the $2, when the alternative is a simple Flow.
  • Customer-facing FAQ use cases seem obvious winners but only if AI can deliver accurate results without passing them back to a human. The aim is successful outcomes, not call deflection.

There is something more important than the ROI though.  It’s learning how to exploit the competitive advantage of a new platform paradigm. 

If you believe that agents will dominate the future and it is a platform change, then you need to gain experience in launching, managing, and measuring agents. There are also cultural and operational changes that need to be established. Organizations that take a “wait and see” approach may never catch up, particularly when AI is evolving so quickly.  

At Elements.cloud, we’ve been using ChatGPT since the start. It’s now deeply integrated across every corner of our business—from product management to marketing, business excellence, IT, and even finance. And we’re not stopping there; we’ve embedded AI in multiple areas of our Elements.cloud platform.

What’s been fascinating is the cultural shift that’s emerged. Instead of fearing AI or being dismissive of its capabilities, our teams have embraced it with a mindset of, “Why not?” or “What if we tried?” Every task is now seen through the lens of “can AI assist with this?” The result? A workplace that’s more productive, curious, experimental, and, quite frankly, more open-minded. We’re already experimenting with agents and have established the approach to building reliable, easily maintained agents by taking a process-led approach.

If I look back over the last 2 years we’ve invested in some areas that rapidly became obsolete, or available through the platform. That is a cost of being early. 

I’d rather be early than irrelevant

Ian Gotts : CEO : Elements.cloud

So, $2 is a small price to pay to be able to learn about agents and prepare for the future. On the back of our implementation, we wrote a book that is helping set some guidelines for building agents.  #1 Learning: Agents are digital labor. Human labor needs clear direction – i.e. process. Digital labor is no different. Design the agent by using a process diagram.

It’s the ROI, not the price 

Not everything is a use case for AI. It is tempting to believe the hype of “everything is going to be an agent” and “SaaS is dead” Clearly this is not true. The challenge as agents evolve is working out the best use cases. The most obvious use cases – customer or product support – also have the greatest cost and risk exposure. 

  • Your agent is your brand ambassador. You need the results to be correct and on message. 
  • It exposes the quality of your support and policy documentation. They probably need to be reviewed and revised considering the agent as the audience, not a human.
  • If the agent cannot complete the request without passing it back to a human the user experience could be worse and you still incur an agent cost
  • If adoption soars, then so do the costs. This is not an issue if you are able to scale down the human support team and realise the benefits, or you are using agents to handle peak loading. 

But before we think about every use case, it is useful to think about the different categories of work and how AI can support them – or not. A recent post by Alexandre Kantjas has grabbed a lot of attention as he distinguished between automation, AI workflows and AI agents. Again, this is his opinion, and it is such a new area the universally accepted definitions have not been established. Hence the debate on the post.

However, he takes a very academic definition and claims that only the last category is a true agent. But this is missing one of the core values of the AI and agents when applied to a business process. The debate really centers around deterministic vs non-deterministic. i.e. non-deterministic outcomes are not consistent for the same input. But for many of the initial use cases for agents – support – we want deterministic results. The outcome should be the same for the same input.  

He defined 3 categories

  • Automation: This is not AI, but is the current workflows in organizations: Flow, PBW, Apex. The outcome is predefined by the business logic coded into the workflow. Not every process needs to be AI-enabled. A well-designed workflow, kicked off by a field value change or a button is often the most cost-effective approach. 
  • AI Workflow: This is where AI is layered on top of a clearly defined process that is driving existing workflows, so the result is consistent. AI makes the user experience better because the user doesn’t need to understand the terminology, policies, and business logic. The AI agent guides the user through the process to an outcome. It also ensures better data quality because the agent is designed to work with the user to ensure accurate data. Finally, business logic is easier to develop and maintain as it is natural language. AI Workflow mimics the customer support agent but is not limited to this use case.
  • Agents: These are non-deterministic where the agent is able to plan and come up with a result, but the answer will not always be the same based on the input, but this doesn’t matter because there is not necessarily a right answer. These more sophisticated agent use cases cover tasks that are more likely to be done by someone with “Assistant” in their job title – e.g. Executive Assistant, Marketing Assistant, HR Assistant, RevOps Assistant. But, the agent may only be able to replace part of their duties or can make them more effective. There is structure to the input that is used to brief the agent. 

AI Workflow is the biggest Agent market

We believe that AI-workflow is the most prevalent use case for AI Agents in the short term and 90% of AI agents in the next 2-3 years will be AI workflow. In the US in 2022 there were over 3 million people employed as agents according to DataUSA.   They are all deterministic. Some / many of them will be replaced by AI Agents.

If you look at Salesforce Agentforce, Microsoft Copilot, and CrewAI platforms (and many more), this is what they have been designed for. The recent interview with Microsoft CEO where he talked about agents ending SaaS. This was “AI-workflow” on top of CRUD databases.

So looking at your current processes that require a deterministic result is the best place to start applying agents. You already have the processes in place because you have human agents delivering them. You have the backend workflows in place, because they are being used by human agents today. 

Try not to be influenced by the $2 per conversation. Instead focus on the outcomes for the agent and use it as a learning experience. 

Where to start

Firstly you need to understand what it takes to build an agent from the ground up. Clearly, this is far more than completing the Trailhead Coral Resorts badge or getting the AI Associate and Specialist certifications. These are great prerequisites.

We’ve written an ebook talking about our journey when building agents which should take your thinking on.

What we’ve discovered is that great agents require strong up-front planning, design, and architecture. You need to stop yourself from jumping into Agent Builder and just starting to type instructions. Whilst it feels like progress, you will never get an agent you can trust because the results will appear inconsistent.

Once you’ve decided on the use case you need to understand and reengineer those business processes to optimize them for AI. The great news is that this is not wasted time. What you are actually doing is building the agent.  That process diagram is the “visual design canvas” for agents. It generates the instructions, guardrails, prompts and actions. 

What we’ve discovered is that you need eliminate ambiguity in your process diagramming to build really good agents”

Adrian King : CTO : Elements.cloud

TL; DR in 3mins 33 secs

Back to News