Main content

Clarity may be emerging in AI capabilities pricing. Here's how

Brian Sommer Profile picture for user brianssommer May 28, 2024
Software vendors have poured on the communications blitz to inform customers and buyers that they have started incorporating artificial intelligence into their products. But, they have simultaneously, conspicuously and quietly avoided discussions about what these new capabilities will cost. Hang onto your checkbooks folks as software vendors might be gearing up to wallet-frack your company’s bank account again.


For most of 2024 (and before), Wall Street and IT analysts have been peppering the executive teams of application software vendors as to how they will price new AI-powered capabilities within their product sets. This is an important question as:

  • Vendors almost never do anything for free
  • Vendors always want to grow their top line (revenue) as it allows top executives to remain employed with the firm, keeps investors happy, etc.
  • Vendors may incur significant fees from their hyperscaler partners to store, train and execute AI utilities. The cost to host all of that data could be great as is the electricity and cooling costs to keep all of those servers and processors humming.
  • Customers need to know whether to acquire these new capabilities or not
  • Customers need to budget for these expenditures
  • Customers need to know what the cost drivers are that will potentially drive outsized bills coming their way
  • Customers hate to be surprised
  • Customers and prospects want transparency into how they will be charged to use these new capabilities
  • A whole ecosystem of procurement, legal and negotiating personnel want time to understand these new pricing models and the economic, business, security, controls, etc. risks that may be present.

State of vendor pricing so far

Let’s stipulate a couple of things first:

  • Most enterprise vendors have been annoyingly and embarrassingly vague, coy, non-responsive, evasive or sketchy in providing answers on these pricing questions. So, if the approaches below seem a bit thin on specifics, it is because that is all the vendors will state.
  • Some vendor executives have already promised Wall Street (or their private equity overlords) that the road to AI is made of gold and that customers will gleefully fork over huge new premiums for the honor of using a vendor’s new AI-powered apps (News flash – they won’t). The executives who made these promises may have shortened their careers.
  • Demand for these tools is not infinite and customers will be the final arbiters as to how elastic their wallets and demand for these capabilities will be.
  • Enterprise customers may be more accepting of these price increases but SMB’s likely will not.

Here are some of the AI-pricing models we have been briefed on:

  • It’s a percent of the customer’s ARR – Talk about simple and easy to calculate, these vendors are employing a pricing surcharge based on what your firm is spending annually on the products it has subscribed. Essentially, the vendor is suggesting that the AI capabilities will cost an additional X% per year to use.
  • AI is part of the platform – This is a rather enlightened approach some vendors are stating. By placing a number of AI capabilities, data and more within the vendor’s platform, the technology capabilities are available in all its applications. In this Utopian pricing model, the AI capabilities are just part of the pre-existing software subscription. This part of the story usually includes AI tools that respond to user inquiries (e.g., chatbots), tools that anticipate workflows, tools that pre-fill fields and other tools that make incremental improvements to existing applications and their usability.
  • AI is part of the platform BUT… - And here is where the limitations come in. We’ve heard the AI is part of platform speech from several vendors but the conversation can segue quickly when you realize they were making a technical point and not a pricing one. The pricing has nuances. Some vendors will let you use the products with most AI-powered capabilities turned off as part of the base subscription. If you want more, you must pay for that. This is akin to how airlines have ‘disaggregated’ the flying purchase experience where every aspect of the ticketing (e.g., preassigned seat, upgrades to premium cabin seating, checked bags, etc.) comes with a separate SKU/price point. One software vendor even used the delineators of Good, Better and Best to describe how different AI components/functionality will be priced.
  • AI is part of the platform EXCEPT… - Other vendors drew a bright line where an application that is predominately a new AI-powered replacement for an existing application will require a new subscription. In contrast, they will not charge an additional fee for adding an AI-search tool or chatbot to a screen in an already subscribed module.
  • There’s the cost of the new functionality and the cost to operate it – Vendor AI-tools that need access to lots of your historical data want to charge more. Why? Every historical database that helps improve sales forecasts, cash flow projections, customer purchases, compensation plans, employee career plans, etc. will need to be trained against the new AI tools. These tools will look for previously unknown relationships, patterns, anomalies, etc. that should, in theory, help the software make better suggestions, predictions, etc.  Customers should really understand what all of this data access means when it comes to driving up the costs of your enterprise software subscriptions. Customers should find out what the vendor is paying for their compute and storage usage from their hyperscaler and only accept a small markup over that.
  • Oh, let’s add a ‘value assessment fee’ – This is a charge the vendor (alone) makes up based on the value they perceive the AI tool will generate for your firm. Buyers should be careful what you disclose to vendors during the sales process as they might play that back to you in a multi-variate pricing algorithm that optimizes the absolute extent of what they think they can charge you. The theory here is that they vendor wants to participate in the upside benefits that their AI functionality might possibly provide.  The key words here being: might possibly.  I have more on this further down the piece.

You Can Trust Us Pricing – From the used car salesperson pitch book came this classic. Customers are supposed to start using the product with no idea how much retroactive pricing will occur later on. Customers should never sign with a vendor whose pricing model is simply “Trust Us – we’ll figure something out later”. You have ZERO negotiating leverage after you become a customer. I’ll let the readers provide their own cautionary guidance re: the wisdom of this pseudo-pricing approach.

Colleague Jon Reed also pointed out that:

Consumption-based models seem to be emerging here, potentially based on buying "credits", For example, two major ERP vendors are using this for some AI, but then they have other AI that is built into products and won't cost more. So even within a vendor's product line/solutions there is pricing variation.


The current AI pricing in application software is so nebulous, still, that customers and prospects must exercise caution. Above all else, customers need assurances, contractually, that these things will work as promised, deliver expected value/benefits and that vendors will stand behind these solutions. That means vendors must indemnify customers against reputational damage (e.g., when an AI tool is caught plagiarizing a competitor’s content or an HR product’s chatbot swears at potential jobseekers). If the vendor simply expects to collect fees but not stand behind its AI tools, prospects should seek true love elsewhere.

Everyone will need remedies if the tools start hallucinating, expose a customer’s private data to third parties, provide obscene or inaccurate results, require more hyperscaler storage or compute power than previously expected, etc. then who pays? What buyers should seek is a pre-nup of sorts so that customers and vendors can exit one of the AI deals if it starts to go sidewise.

Nothing in this world is static. Customers should not agree to be locked in to an AI-capability that could be rendered obsolete by new, third-party products in a few weeks or months. For example, why would a customer today want to pay for an AI job description generator when there are gobs of them available for free? Likewise, why should a customer pay for an AI capability who usefulness is rendered null and void after other products appear (e.g., Because jobseekers can now generate thousands of applications a week, is an ATS (applicant tracking system) still relevant in a resume spam infested world?). There’s a harsh reality to consider here: the lifespan of these AI tools and capabilities may be very limited with some becoming obsolete almost as soon as they are made available. If you are thinking in terms of 10-year lifecycles for these new capabilities, you may need to dramatically readjust your planning horizons.

Money back guarantee? Smart software buyers should insist on contract mechanisms that guarantee that the customer will receive the promised results and that the results will not possess problems (e.g., obscene material, plagiarized content, incorrect answers, hallucinations, etc.). Moreover, if your vendor insists on a value-sharing or value-pricing arrangement as part of its recompense, customers should get a crack attorney to go through this agreement.

Other macro Issues to consider

Why would a buyer agree to these pricing formulas? That’s a fair question.  The models above are incomplete, potentially difficult to use, and potentially short-lived. There may be other issues to consider, too.

Can AI be free?

Given that AI can write code and that hyperscalers continue to get ever more powerful while also achieving lower cost economies of scale, the cost to develop new AI functionality is headed inexorably to zero.  Or is it?

A recent Bloomberg Businessweek article (“The Age of AI Ambiguity” by Brad Stone and Rachel Metz) noted:

“The cost of training and running an AI model is enormous. GPT-4 used an estimated an $78 million worth of computing power as it was being trained by OpenAI, according to Stanford University’s Artificial Intelligence Index Report 2024, released in April. Google’s Gemini Ultra cost $191 million to train. Both models were developed on graphics processors that are expensive and difficult to obtain from Nvidia Corp., which for now is the only company making them.”

Application software vendors will need to recover costs incurred via their use of third-party AI computing resources. And, some AI tools will need frequent training so that their results are as current and accurate as possible. All of the training, electricity, compute hardware, cooling, etc. costs money and someone will need to pay for it.

Value pricing with AI

Vendors and implementers have tried to implement value-based pricing for several decades. It hasn’t caught on as it’s tough to quantify and business conditions can change quickly and materially. Value pricing slows down deals and may require a level of business transparency that customers will not allow vendors to have.

The 800-lb. gorilla in the room could be this issue though: how can a vendor (or a customer) prove, conclusively, that the value the AI-tools delivered was what was projected? I suspect it can’t and that an approximation or guess may be the closest answer any of us will see.  To illustrate, how do you prove that a chatbot gave an answer that actually solved a person’s problem? For the most part, you can’t. Like many of you, I encounter chatbots on a daily basis and often abandon them quickly as they aren’t giving me the answers I need. Worse, these brain-dead tools often lack feedback mechanisms to report their low-value responses, And, the final insult is that even if feedback were possible, would any human actually see it and would they know how to fix the problem? No, I’m not convinced customer service agents are empowered to capture feedback let alone possess the technical skills to improve the training database behind these bots.

Likewise, can a vendor prove that an AI-tool that suggested a new process flow would, in practice, actually deliver value? No again. In this case, maybe a user has a string of anomalous transactions that trigger the development of new workflow processes. But, if the anomalous transactions cease, then the new workflows don’t get used nor do they create value. So, value could be fleeting, transitory and/or unpredictable while the billing of the AI utility could be definitive and long-lasting. This misalignment WILL cause problems between vendors and customers.

And then we get to the monetization of an AI tool. I know why vendors struggle with this as it’s a complex, multivariate issue. Ironically, that’s exactly the kind of problem an AI tool can solve for if it only had the right database to train with.   

My take

Short answer: Caveat Emptor (let the buyer beware)

Longer answer: With great uncertainty comes great risk. Smart customers will bulletproof their contracts to minimize surprises.

A grey colored placeholder image