The News

On February. 25, 2026, Salesforce announced a pricing and metrics update. During the company’s Q4 FY2026 earnings call, CEO Marc Benioff, together with CMO Patrick Stokes, unveiled the Agentic Work Unit (AWU). Positioned as a metric to quantify the labor performed by autonomous digital systems, Salesforce defines an AWU as one discrete task accomplished by an AI agent.

According to Salesforce, this discrete task represents the exact moment "raw intelligence is converted into real work". It is not a fixed unit but measured as a processed prompt, a completed reasoning chain, or an invoked tool. Salesforce explicitly designed the AWU to move the industry conversation away from the raw consumption of Large Language Model (LLM) tokens. As Benioff noted, tokens only measure "how much an AI talks," whereas the AWU is intended to measure actual business execution.

The scale of this rollout is massive. Salesforce reported that its platform has already processed over 19 trillion AI tokens, translating them into 2.4 billion Agentic Work Units, with 771 million AWUs delivered in the fourth quarter alone. This new metric serves as the underlying foundation for Salesforce's evolving Agentforce monetization strategy.

The bigger picture

Following a nearly 18-month period of pricing triangulation, which included a $2.00 per conversation model and a $0.10 per action "Flex Credit" model, Salesforce is leveraging the AWU to track system utilization, even as it wraps enterprise purchasing in familiar, unmetered per-user license agreements starting at $125 per user per month.  

To understand the significance of the Agentic Work Unit, one must view it through the lens of a broader industry crisis: the so-called "SaaSpocalypse" and the looming threat of the seat cannibalization trap. For two decades, the Software-as-a-Service business model has been dominated by seat-based licensing. However, as agentic AI systems mature and promise to be capable of executing multi-step workflows autonomously, they inherently reduce the need for human software operators. If an AI agent resolves 84% of tier-one support tickets without human intervention, the enterprise requires fewer human support seats. I have repeatedly written about pricing models, e.g., here, as part of my CustomerThink column.

This dynamic has forced the software industry into a frantic transition toward usage-based and attempts at outcome-based pricing models. Usage-based pricing, popularized by cloud infrastructure providers like AWS and data platforms like Snowflake, charges customers based on system consumption, e.g., compute seconds or data processed, or, these days, tokens consumed. While this protects the vendor's margins and aligns with their variable cloud GPU costs, it shifts the financial risk of system inefficiency entirely onto the buyer. The vendors essentially play the role of Pontius Pilate and wash their hands in innocence.

Conversely, agile AI disruptors and customer service incumbents are aggressively pioneering true outcome-based pricing, where the billable event is delayed until a verified business success is achieved. For instance, Intercom's Fin AI agent charges a strict $0.99 per successful resolution, while Zendesk recently started a $1.50 per automated resolution model. In these models, if the AI fails to resolve the customer's issue, the customer pays nothing. Correspondingly, the vendor has skin in the game and needs to be interested in its software actually delivering value.

Salesforce’s introduction of the AWU represents a kind of a middle ground. Industry analysts like Constellation Research’s Liz Miller observe that the AWU acts as a placeholder for the agentic era, much like clicks and likes functioned in the early days of online and social media. It is still a usage-based consumption metric masquerading as an outcome metric. The industry is currently witnessing a tug-of-war: legacy giants are deploying metrics like the AWU to track utilization and justify high enterprise license costs, while pure-play AI vendors intend to leverage outcome-based pricing as a competitive weapon to steal market share by guaranteeing and demonstrating return on investment.

My point of view and analysis

As someone who has spent years helping organizations unlock their potential through digital transformation initiatives, I look at the Agentic Work Unit highly skeptical. When evaluating generative and agentic AI investments, the critical measure is the ability to deliver measurable business results, not just technological activity. In this context, the AWU represents a fundamental conflation: it equates doing work with achieving outcomes, which simply is not true.

By defining an AWU as a discrete task, such as invoking an API or triggering a workflow, Salesforce has created a metric that measures machine exertion rather than enterprise value. In the realm of autonomous systems, an AI agent can execute thousands of discrete tasks, burn through immense computational resources, and work incredibly hard while achieving absolutely nothing of commercial consequence.

Working hard on the wrong thing still doesn’t deliver results.

To fully grasp why measuring discrete AI tasks is a poor proxy for value, consider the analogy with an unoptimized database query vs. an optimized one in cloud data warehouses. A highly optimized SQL query returns a vital dataset in seconds for pennies. Conversely, a poorly written query forces the database engine into massive data scans. It might be running for hours and consuming plenty of CPU and memory resources. From the vendor's billing perspective, the system successfully performed the discrete scanning tasks it was instructed to execute. However, it results in a massive consumption bill for the customer. The business gains little value, as much of it is harvested by the vendor; even worse, if the result is wrong. Yet the financial penalty is severe.  

The Agentic Work Unit operates exactly on this flawed economic principle. Autonomous AI agents are still highly susceptible to unique failure modes, e.g., the infinite reasoning loop. If an agent encounters an ambiguous prompt or lacks solid memory tracking, it may repeatedly call the same tool or query the same database in an endless cycle due to perfection bias. While engineers desperately build so-called Meta-Reasoners to halt this wasted computation, the AWU metric actively monetizes it. If a confused agent loops fifty times before timing out, it has successfully generated fifty AWUs delivering zero result. The customer is actively billed for the machine's confusion.

Furthermore, agentic workflows suffer from compounding hallucinations, or epistemic debt. If an agent hallucinates a false premise in step one, it will still confidently execute subsequent tools based on that fabrication. By the time the workflow concludes, the agent may have triggered dozens of AWUs across multiple enterprise systems, corrupting data and requiring costly human remediation.

Ultimately, a dashboard celebrating 2.4 billion AWUs gives the illusion of massive productivity, but it is a vanity metric. If those tasks were merely redundant internal data reshuffling or failed reasoning loops, the actual profit multiplier of the organization remains unchanged. An Agentic Work Unit quantifies motion, but motion is not progress. Until AI pricing models mature to align the cost of digital labor with the verified delivery of business outcomes, enterprises must continue to treat effort-based metrics like the AWU with extreme skepticism.

Just my $.02. What do you think?

These days, we are drowning in conflicting information about the value of generative and/or agentic AI. I, myself am researching for good studies that dive into the ROI that is generated by this technology, with limited success. Most information is anecdotal, or comes from success stories, which cannot get used too literally.

Two major 2025 reports from MIT and Wharton, respectively, paint starkly different pictures of AI adoption and adoption success. While the meanwhile often quoted MIT NANDA “report” on the state of AI in business often gets quoted with 95 percent of all businesses not getting any ROI from their gen AI initiatives, a recent study by the Wharton Business School shows a very different result with 74 per cent of enterprises showing a positive ROI. Why is one so pessimistic and the other so optimistic? As I have written before, a closer look at the data reveals the 95% "failure" narrative is a myth, or even a scare, and the real story is probably a different and far more differentiated one, which Wharton names Accountable Acceleration.

Is GenAI really a 1-in-20 lottery ticket or is it rather a core business function? So, let’s have a look.

Methodology matters – debunking the 95% failure rate

In contrast to the NANDA “report” that relies on a fairly small sample of about 150 survey responses and 52 structured interviews, the. Wharton report bases on a large-scale, quantitative and longitudinal study. It surveyed around 800 senior decision-makers at businesses of different sizes and is tracking trends for the third consecutive year. Therefore, its data is built for statistically valid conclusions.

In addition, MIT NANDA’s rather sensational number of a 95 per cent failure rate is the result of a different focus. This “report” solely looks at custom, task specific agentic AI implementations that reach production state and measurable P&L impact. It does not look at the value of widely adopted generative AI tools like ChatGPT, Copilot or Perplexity. Instead, it dismisses their impact by labeling them as as tools that primarily enhance individual productivity, not P&L performance. And this, despite their obvious success that the report also identifies. 40 per cent of general purpose LLM implementations go into production with about 90 per cent of all employees using privately purchased licenses for ChatGPT, Perplexity, et. al. This is an indication that the corporate implementations are not exactly looking at the right pain points.

In contrast to this, the Wharton report identifies this as the primary source of value creation. Its data shows the top use cases are data analysis, content creation and its summarization and presentation (p. 30). This is the ROI. The MIT report is like arguing 95% of companies get no return from email because it only boosts individual productivity. This doesn’t make much sense as an increase in productivity is a return. Wharton, in turn, identified how businesses identify and measure ROI and found that almost 75 per cent of all businesses report a positive return on investments, with smaller businesses (p. 45) and some industry sectors (p. 47) seeing a higher return. But the picture is clear.

The ROI story – mainstream and measured

Given all this, the Wharton study seems to be a far more credible source of information about the current usage and ROI of generative AI. There is only a small fraction of businesses that sees a negative ROI, with some businesses seeing it as neutral or too early to assess.

The numbers vary across industries, but the general picture is clear. Generative AI is not a high-risk gamble. On the contrary, if implementations are managed appropriately, it does deliver results and consequently, Wharton concludes that we have now regular usage in core business operations, with embedded ROI metrics. More than 70 per cent of the surveyed organizations are formally measuring ROI, with a focus on productivity gains and incremental profit.

There is both, an increasing usage of generative AI and a continued significant investment, including into own research and development. Wharton dubs this phase as “accountable acceleration”, moving on from exploration- and experimentation-oriented phases in the previous years. I guess, there needed to be a catchy marketing phrase …

And, as said above, it’s showing results. Of the surveyed businesses, 74 per cent are seeing positive returns. This includes 35 per cent reporting "significantly positive ROI" and 39 per cent "moderately positive ROI". This data flat out contradicts the 95 per cent failure narrative of MIT NANDA. Apparently, we are not looking at a divide with a 95 per cent chasm but rather a far smaller spectrum of adoption speed and maturity.

Buy – or rather build?

Now that we are clear about well-managed generative AI initiatives showing a positive impact, the make or buy question looms again. NANDA quite unequivocally says “buy”, as they find that the failure rate of custom implementations is double the failure rate of strategic partnerships with a vendor and systems integrator.

Wharton roughly shows an equal distribution of efforts in new technology, enhancing existing technology and internal research and development efforts.

Which is not necessarily a contradiction, as it basically says, “buy and adapt”. Some scenarios can be supported by delivered software, some needs training or fine tuning, other models need to be specifically built.

Businesses are adopting a far more sophisticated hybrid strategy than simply make or buy. Firms are not just "buying" off-the-shelf tools; they are investing significant capital to build custom, proprietary solutions that drive competitive advantage.

The MIT report's "buy, don't build" advice, in contrast, is very simplistic. While I stick to the recommendation that I made in my article Beyond the Hype: Unlocking GenAI ROI in the Enterprise, the truth is more nuanced. It is basically the same as for the implementation of enterprise software in general. Stick to best practices where there is no significant or lasting competitive differentiator and invest into custom implementations where a differentiator with a competitive advantage lies.

The real barrier

Which leads us to what prohibits businesses from taking even more advantage from the use of generative AI. As Wharton shows, the main problem is not the technology, at least not only. There are credible studies that show a decreasing likelihood of success with increasing complexity of the modeled scenario, including CRM-Arena Pro and TheAgentCompany, but then these agree with the Wharton findings that success is very well possible.

The technology is mostly ready.

The real barrier is people and organization – essentially corporate culture. Management needs to set an example, people need to be educated and, very importantly, governance needs to be in place that both helps the users and protects the business. This requires change management. According to Wharton, the toughest challenge facing businesses are a lack of skills, employee fear for their jobs and management’s ability to constructively manage the necessary change by convincing the employees that the use of AI is not about technology grabbing their jobs.

From sensationalism to strategy

What we can confidently say is that the generative AI landscape is not as dystopian as the NANDA “report” makes it appear. This narrative is created on a too small data set and a questionable definition of “value”. It is fundamentally biased by an underlying agenda – which comes from the authors pointing a way into their solution of overcoming the technology problem. The Wharton "Accountable Acceleration" study provides the more data-driven, and optimistic view. Generative AI has gone mainstream. The return of investments can get measured, is being measured, and it is proving positive, with few exceptions.

The "GenAI Divide" isn't a chasm between success and failure, caused by poor technology.

It is rather a spectrum of maturity and of connecting investments into generative AI into strategic business KPIs. And, of course, selecting the right problems to solve. The real challenge is not to find a silver bullet that automagically solves business challenges. It's the boring work of business transformation. It is about investing into and aligning people, investing in training, setting smart guardrails, and executing a hybrid build and buy strategy.

This is the real, actionable roadmap to success.

If you want to explore, how this roadmap could look like for you, get in touch with me for an informal conversation.

Now, that we are in the middle of – or hopefully closer to the end of – a general hype that was caused by Open AI’s ChatGPT, it is time to reemphasize on what is possible and what is not, what should be done and what not. It is time to look at business use cases that are beyond the hype and that can be tied to actual business outcomes and business value.

This, especially, in the light of the probably most expensive demo ever, after Google Bard gave a factually wrong answer in its release demo. A factual error wiped more than $100bn US off Google’s valuation.

I say this without any gloating. Still, this incident shows how high the stakes are when it comes to large language models, LLM. It also shows that businesses need to have a good and hard look at what problems they can meaningfully solve with their help. This includes quick wins as well as strategic solutions.

From a business perspective, there are at least two dimensions to look at when assessing the usefulness of solutions that involve large language models, LLM.

One dimension, of course, is the degree of language fluency the system is capable of. Conversational user interfaces, exposed by chatbots or voice bots and digital assistants, smart speakers, etc. are around for a while now. These systems are able to interpret the written or spoken word, and to respond accordingly. This response is either written/spoken or by initiating the action that was asked for. One of the main limitations of these more traditional conversational AI systems is that they are better in understanding than in – lacking a better word – expressing themselves. Relying on well-trained machine learning models, they are also quite regularly able to surface a correct solution for problems in the problem domain that they are trained for. They usually work based on pretrained intents.

And, based on the training data, they usually give quite accurate responses to questions in their domain.

The problem: They are usually limited to a fairly small number of domains.

LLMs, on the other hand, are generally trained “to understand the relationships between words, phrases and sentences in a language. The goal is to have the LLM generate outputs that are semantically meaningful and reflect the context of the input.” This is part of ChatGPTs answer to the question what the purposes of an LLM is. The training set of an LLM is usually a vast amount of “real world” knowledge that usually comes from publicly available sources – aka the Internet. The output itself can be in written, graphical or other formats.

What LLMs excel in is generating responses to questions in a human way. And they can respond to a wide variety of topics. When focusing on text, they are built to generate coherent and meaningful responses.

The problem: They sometimes lack accuracy and give wrong output with full confidence. Even worse, wrong or inaccurate output is not easily identifiable by a user without the requisite knowledge. Again, refer to the Google Bard example that (temporarily, at least) wiped off $100 billion US from Googles valuation. Not picking on Google, there are plenty of examples around that call out ChatGPT or You.com or other tools.

Consequently, the other dimension to look at is accuracy.

The question is whether both dimensions always matter equally or not. In a business sense, one can argue that accuracy matters always. Receiving factual errors in a business conversation is not only a poor customer experience but may in extreme cases even lead to legal issues.

What is also important to understand is that the more accuracy is required the more the necessity of integrating additional systems to augment the LLM increases. An LLM on its own is not much more than some form of entertainment. Even in search engines, LLMs only augment the search by enabling natural language queries and the delivery of results in human language instead of a mere link list.

At least they should do this.

With all this being said, what are business use cases involving a large language model? As said, there needs to be a reasonable accuracy. Obviously, they require fluency as a precondition, as fluency is the core differentiator of an LLM.

Let’s look at some use cases in no particular order of priority.

• I’d start with something that I’d call “storytelling”. This is basically the creation of market-relevant documents that describe the capabilities and differentiating factors of a product, solution, or service. Being somewhat marketing related (no offence intended) and a first point of contact for customers, it needs to be easy to understand without requiring a great deal of technical accuracy. At the same time, it must not be wrong. A stripped-down version of this could be the (improved) generation of social media content, e.g., tweets. Benefits are faster creation of high-level content for general websites but also, more specifically, for ABM scenarios and landing pages. To be able to create this text, an LLM needs to be connected to internal systems holding requirements, specifications as well as communications between the involved persons. This is also a use case that should be implement-able near-term.
• One of the main tasks of people is the writing of, and more so, responding to emails. Especially, in sales scenarios, customer inquiries can get formulated and suggested based upon previous emails and the context given by the CRM system, e.g., about proposals made. This scenario would already require quite a high accuracy to avoid sending out faulty information that might be legally binding. The benefit of this scenario is a significant reduction time needed to send emails, resulting in increased productivity. It is a scenario that Microsoft has already implemented in its Viva Sales solution.
• Generation of documentation is a scenario that somewhat varies in the requirement for fluency. It can be mainly divided into technical and user documentation. While user documentation needs to be extremely readable, the writing style is somewhat less important for technical documentation. Conversely, technical documentation likely needs to have a high degree of technical accuracy that is not needed in user documentation, which means that either different repositories or different parts of source documents need to be used to create the texts and potentially diagrams and images.
• One of the most promising use cases in the short term is customer service, including enterprise search. Here, users want answers to their questions, not just links or something actioned. To achieve this, it is necessary to connect to a conversational AI, business systems and a well-functioning knowledge base that helps in generating accurate answers when searching for something. The actioning of issues is very similar to what conversational AIs do already now. The differences are that the intent detection can be far better as the LLM can create more than enough training sets for this and that the answers given by the system are far more fluent. The same holds true for an inquiry scenario. However, as a word of caution, the accuracy of responses to inquiries depends heavily on the kb content that gets searched by the enterprise search. Therefore, the kb needs continuous and rigorous scrutiny. If this is given, the benefits lie in increased call deflection and customer satisfaction. Properly implemented, benefits include an improved call deflection as more cases can get handled by the system, combined with an increased customer satisfaction as issue handling can become quite easy and efficient for the customer. Cognigy has recently presented some very good examples (here and here) that also include voice in- and output.
• Agent assistance is somewhat easier to implement, as it mostly needs to connect to the customer service application, including the chat history. Having complete access to sales and marketing data, of course is helpful, too. Combined with a sentiment analysis, the LLM can suggest text blocks for the agent to use. The benefits of this are an increased agent efficiency and quite possibly also higher customer satisfaction as the text blocks do exhibit more empathy with the customer’s situation than texts generated without an LLM.

In summary, these five scenarios show use cases involving an LLM that are beyond the hype. They can get implemented in a short time and they can also be easily tied to business outcomes. That way, their benefits can get measured.

Which other use cases do you see? And how would you tie them to business value?