Haris Rahi: AI Pricing Must Be Transparent, Human-Led and Built for Learning
I began my career as a programmer. I loved computers and programming, but after a few years I realized that the pace and depth of traditional software engineering were not the best long term fit for me I was more drawn to the intersection of technology products and commercial problems..
So I moved into product management. I managed enterprise software products, directed their development, worked with clients and helped bring products to market. It was a great fit because it allowed me to remain close to technology while also working on the commercial side: product marketing, client engagement and sales.
That eventually led to a role in technology alliances, where I worked with partners such as Microsoft and Intel to develop joint products and take them to market. Later, a CIO asked me to bring the same mindset to the company’s vendor relationships. At first, I resisted. It sounded a lot like procurement, and I had seen procurement teams that knew the price of everything and the value of nothing.
But the CIO wanted something different: a partner-oriented approach to technology vendors. That experience ultimately led me to Accenture, where I ran a practice that helped clients make technology investments. We were managing roughly $80 billion when ChatGPT arrived. In late 2022, I felt as though I had been bitten by the bug.
In 2023, I went all in on AI. I helped establish Accenture’s AI footprint, built a product that trained nearly 600 consultants, and supported sourcing and contracting work. I came to believe that this capability should not be limited to consultants. Every enterprise should be able to develop it. That was the impetus for going out on my own.
Today, my work is an extension of the journey I began in 2023. I effectively treated myself as a product: learning how AI worked, applying it in practice and moving beyond the technology itself. Very early on, it became clear to me that AI adoption is much more about human psychology than it is about learning a new technology.
You do not need to “learn AI” in the traditional software-training sense. You can start a conversation with a chatbot. The challenge is learning how to partner with the technology in a way that improves real work.
That is how I work with clients. I help teams become capable and confident in using AI, but the training is grounded in their actual jobs rather than generic concepts. General instruction is useful only if people can return to their desks and apply it immediately.
I also help organizations build solutions for their own needs. We are entering a world in which companies do not always need to acquire a massive off-the-shelf software package. They can build targeted solutions for a team, a workflow or a specific operational problem. Finally, I help clients evaluate AI product and service vendors through the same practical, human-led lens.
What buyers should ask AI vendors
When a vendor presents an AI agent—say, one designed to review contracts, produce redlines or support lawyers and paralegals—I begin with the underlying technology.
My inquiry has two dimensions. First, how has the vendor instilled the domain knowledge and expertise the agent needs to operate? Second, what infrastructure supports it?
If a vendor is built on OpenAI, Anthropic or Google models, there is nothing inherently wrong with that. But it means the vendor is reselling tokens. If the underlying provider changes its model, its pricing or its terms, that change directly affects the economics and potentially the performance of the product being sold.
I also want to understand the agent’s design principles. Is it based on the vendor’s own model? Is it an off-the-shelf model that has been adapted or retrained? Each approach can be valid, but buyers should understand how the agent was developed in the first place.
AI is never one-and-done. Models change. Techniques change. Drift occurs. So the next question is: how will the vendor maintain the solution?
Then I ask the question that matters most: Where did it fail?
I do not ask that to stump anyone. I ask because I know how often a solution must fail before it becomes trustworthy and effective for the people who use it. When a vendor can candidly explain where its product struggled—where it produced poor results, what users found difficult, and how the team responded—that is a positive signal. It tells me they have worked with real practitioners and confronted the weaknesses of their product.
If they cannot answer foundational questions about accuracy, trust and failure modes, that is a red flag.
The final question is how they expect customers to use the solution. If the proposition is the wholesale outsourcing of cognition, I become concerned. Professionals—whether attorneys, paralegals or other domain experts—must retain cognitive sovereignty. The technology may be a virtual teammate or digital assistant, but it is not yet a substitute for human accountability, expertise and judgment.
Value depends on mobilization
Once the technology itself has passed initial scrutiny, I want to know how the vendor plans to mobilize users.
In my view, AI adoption is not comparable to legacy software deployment. It is less about learning a tool and more about learning how to partner with it. That requires a different kind of capability-building.
If a vendor says, “You have 50 lawyers, so buy 50 licenses,” that is not enough. The organization must enable those 50 lawyers to work effectively with the agent. They need to understand where it helps, where it does not, how to review its output and how to incorporate it into their workflow.
That is why the rise of forward-deployed engineers is so important. These are the people who work alongside practitioners, helping them apply the technology in context. The value of an AI solution does not materialize at contract signature. It comes when people can use it well enough to achieve better outcomes.
Pricing must reflect real economics
When I ask vendors about pricing, I want transparency—not just in the price a customer sees, but in the economics underneath it.
I recently worked with a client whose long-standing software vendor had introduced an AI assistant into its platform. The implementation itself was strong. The vendor was transparent about the models it was using and how it deployed different models for different tasks.
But its commercial proposal was a five-year fixed annual fee. On the surface, it looked excellent. Let us say the price was $200,000 per year. Given the potential value of the AI capability, that could have been a reasonable investment.
The issue was that the product relied on third-party foundation models. The vendor was effectively reselling token consumption, and it did not control the cost of those tokens. So the question became straightforward: how can a vendor guarantee a customer’s economics when its own economics are not guaranteed?
The vendor could decide not to use a more expensive model if prices rose. That may be reasonable. But it should be explicit. Alternatively, it could pass through token costs while charging separately for the value it creates: the product, the integration, the intellectual property, the research and development, and the service layer.
That kind of arrangement can be more honest. A vendor deserves to be paid for what it has created. But the customer should also understand which portion of its spend reflects a variable external input and which portion reflects the vendor’s own value.
This transparency is missing from many AI products today. I can see the credits I have consumed in various AI tools, but I often cannot tell what activity caused the consumption. A task-level meter would be far more useful: this task used this number of tokens with this model; that task used another amount. Then I could make informed choices about cost, quality and speed.
Think of the way ride-hailing works. You can choose an economy ride, a comfort option or a premium car, and you can see the price before deciding. AI users need more of that agency.
Credits are useful—but not fully predictable
Credits can be a useful commercial mechanism, but AI introduces a deeper challenge: usage is inherently difficult to predict.
AI is unlike conventional software because it is probabilistic. The same action does not always produce the same result. How you communicate with the system affects what it produces. More importantly, before people begin using AI, they often cannot imagine the valuable workflows they will discover.
That makes predictability difficult. It is not merely that customers cannot forecast their costs. They may not even know how they will use the technology six months from now.
This is why I caution clients against treating experimental token use as waste. Tokens spent learning, testing, failing and redesigning workflows are not necessarily wasted. They are part of the discovery process. Teams need the freedom to find out where AI is strong, where it is weak, where it needs human augmentation, and which workflows become newly possible.
I saw this with procurement teams. In 2024, I was teaching teams that they could use AI not simply to respond to a request for proposal, but to design the optimal commercial structure for that RFP before vendors responded. Many had never considered the possibility. Yet they understood their operations well enough to develop a range of potential commercial models that could create a genuine win-win with suppliers.
Transparency helps with credit-based pricing, but it will not entirely solve predictability. Users are still discovering what the technology makes possible.
AI can support more flexible pricing
This may also create a more interesting future for B2B pricing. Rather than accepting a standard commercial model from a vendor, buyers may increasingly use AI to develop and propose models of their own.
A vendor might worry that a different commercial structure for every customer would become impossible to manage. My response is: manage that complexity with AI.
A company could have 15 different commercial models for 15 clients, as long as each is profitable and mutually beneficial. AI agents can help manage the variation that businesses once had to eliminate through standardization.
Organizations have historically pursued standardization because managing variety was too complicated. But now we have technology that can absorb much of that complexity. Over time, I would expect patterns to emerge. Different organizations and industries may appear unique on the surface, but finance is finance and sales is sales. At some level, many enterprises are performing similar tasks in different ways.
The goal should not be commercial chaos. It should be the ability to create more precisely aligned, profitable and durable agreements.
Buy less software, learn more AI
I do not believe buyers should rush into purchasing every AI product or agent that enters the market.
Before buying a specialised AI solution, organizations should use general-purpose AI tools themselves. Get access to an enterprise-grade model, learn how it works and augment your people first. You will gain a much better understanding of where AI genuinely creates value—and you may discover that you do not need another software product at all.
It would be a disservice if AI simply became another SKU in an already bloated software budget. The opportunity is much bigger: to rethink how work gets done, rather than layering more tools on top of existing processes.
A realistic view of the future
I am concerned about the public narrative around AI. Not because people should be uncritical, but because we need realism rather than fear or hype.
The technology is not going away. Its influence will increase. If organizations and countries respond mainly through anxiety, they risk falling behind.
I also think the prediction of an immediate jobs apocalypse is overstated. AI will replace certain tasks, and in some cases it may replace enough tasks to disrupt particular roles. But every job is a collection of tasks. AI itself is uneven: very good at some things, poor at others. I expect adoption and workforce impact to be jagged rather than instantaneous.
The same applies to the discussion of artificial general intelligence. We may eventually reach it, but I do not believe anyone can credibly say exactly when. Current systems can perform many tasks, but task performance is not the same as fully replacing a job or a person. Today’s models still have meaningful failure and incompleteness rates.
We are at the foothills, not the summit.
My advice to anyone beginning this journey is simple: set aside the headlines for a moment and sit down with the technology. Have a conversation with it. Learn its limitations as well as its strengths.
I have trained nearly 700 people since 2023, including 300 practitioners at Accenture during my time there, and I have seen what happens when people use AI well. It amplifies who they are. It can help them do things they had not imagined possible, without stripping away their judgment or their voice.
One nonprofit leader I worked with wanted to submit more funding proposals. With an AI solution designed carefully around her voice and her values, she was able to reach more donors and generate more funding opportunities. The technology did not replace her. It helped her express what mattered to her at a greater scale.
People can recognize generic AI output. They can also recognize when technology has helped someone communicate with more clarity, capacity and authenticity.
That is the right ambition. We should neither be blind optimists nor abject pessimists. We should be realistic, responsible and curious. AI is powerful—but like any powerful technology, it requires skill, safeguards and human judgment.