the economics of software are changing
an ai feature can become more popular every month and make the business less profitable at the same time.
imagine launching an ai assistant that helps customers complete tasks faster. adoption grows, customers use it more frequently, and the product starts becoming part of their daily workflow. by most product metrics, things look healthy.
then the infrastructure bill arrives.
every interaction consumes resources. more complex requests require more reasoning. some tasks involve multiple model calls, external tools, retrieval, retries, and human intervention. the product is creating more value, but the cost of delivering that value is increasing too.
this is becoming an important challenge for software businesses.
in its august 2026 research, gartner predicted that inference costs per agentic ai workflow could increase more than fivefold through 2028. its research points to an interesting tension: model inference is becoming more efficient, while increasingly sophisticated ai products are consuming more tokens and performing more complex work.
that creates a new set of questions for product teams. how much does it cost to deliver a successful outcome? does the revenue from a customer justify the resources they consume? and does the pricing model account for the way customers actually use the product?
these questions connect product decisions directly to business sustainability. understanding them requires looking beyond adoption, engagement, and feature usage, and getting closer to the economics of delivering software.
why ai changes the economics of software
traditional saas businesses generally operate with a relatively predictable cost structure. building the software requires significant investment, but once the product is developed, serving an additional customer usually adds a comparatively small amount to the cost of delivery.
this is one of the reasons subscription and per-seat pricing became so popular. a company can charge customers a recurring fee, forecast revenue based on subscriptions, and serve a growing customer base without infrastructure costs increasing at the same rate as usage.
ai introduces a different cost structure.
every interaction can trigger additional computation. a simple request might require a single model call. a more complex task could involve retrieving information, sending multiple prompts to a model, calling external tools, processing the results, and retrying when something fails.
the cost of delivering the feature therefore depends on how customers use it, how much work each request requires, and how efficiently the underlying system performs.
consider an ai assistant that helps a user research a market.
a basic interaction might involve one prompt and one response. a more involved request could require searching multiple sources, summarising documents, comparing findings, checking the results, and generating a final report. the second task delivers considerably more value, but it also consumes more resources.
this becomes even more pronounced with agentic systems. an agent can perform multiple steps before completing a task, with each step potentially involving another model call or external tool.
gartner’s research on the inference paradox highlights this growing tension. improvements in model efficiency can reduce the cost of individual operations, while increasingly complex workflows increase the total amount of inference required.
there are several cost drivers product teams need to understand:
model usage: input tokens, output tokens, reasoning tokens, and the model selected for each task.
context and retrieval: processing long conversations, retrieving relevant information, generating embeddings, and reranking results.
tool execution: external api calls, database queries, browser operations, and integrations.
orchestration: coordinating multiple model calls, managing agent workflows, and handling retries.
human intervention: reviewing outputs, correcting mistakes, and taking over tasks the system cannot complete reliably.
the actual cost varies significantly across products and workflows. a customer support assistant, an ai coding tool, and an autonomous research agent can have very different consumption patterns even when they use similar underlying models.
there is another important distinction here: the price of an individual model call is only one part of the equation.
openai’s api pricing and anthropic’s pricing documentation show how costs vary by model, input and output tokens, caching, and other capabilities. these differences matter when choosing models and designing workflows, but they don’t tell us the full cost of completing a customer task.
that requires looking at the entire system.
for product managers, this changes the way feature decisions need to be evaluated. adding another step to an ai workflow might improve the quality of its output, but it could also increase latency and operating costs. using a more capable model might improve completion rates, but the additional expense needs to be justified by the value it creates.
these trade-offs make product architecture, user experience, and business economics closely connected. as we explored in our guide to ai product reliability, the way an ai system is designed has a direct influence on how it performs in production.
the more capable an ai product becomes, the more important it is to understand what each successful interaction costs the business.
understanding the true cost of an ai product
if you ask an engineering team how much an ai feature costs, the first number you’ll probably hear is the model’s api bill.
it’s a useful starting point, but it doesn’t tell you how much the product actually costs to operate.
consider an ai research assistant. a user asks it to prepare a competitor analysis. the system retrieves documents, sends context to a model, searches for additional information, processes the results, and generates a report. if the first attempt misses important information, it might run another search or ask the model to try again.
each operation adds to the cost of completing that task.
to understand the economics of an ai product, we need to account for the entire workflow.
1. model inference
this is the cost of sending information to a model and receiving its output. it depends on the model selected, the number of input and output tokens, and whether features such as extended reasoning are involved.
a longer conversation or a document-heavy task can consume significantly more tokens than a simple question. using a more capable model can also increase the cost of processing a request.
2. retrieval and context
many ai products need information beyond what the user provides in a prompt. they retrieve documents, search databases, generate embeddings, and sometimes rerank results before passing relevant context to a model.
these operations introduce additional infrastructure and processing costs.
for products using retrieval-augmented generation (rag), the quality and efficiency of retrieval also matter. retrieving too much irrelevant information can increase token consumption without improving the answer.
3. tools and external services
ai systems increasingly interact with other software. they might search the web, query a database, access a CRM, execute code, or call an external api.
some of these operations have their own pricing. others introduce infrastructure, latency, and operational overhead.
an agent that makes ten tool calls to complete a task can have a very different cost profile from one that completes the same task in two calls.
4. orchestration and retries
complex ai workflows often involve multiple model calls. one model might classify a request, another might generate an answer, and another might evaluate or refine the result.
failed operations can trigger retries. agents can also take unnecessary steps, repeat tool calls, or get stuck in loops.
these costs are easy to overlook when measuring only the initial request. they become particularly important as workflows grow more complex.
5. human review and intervention
not every ai task can be completed without human involvement.
a customer support agent might draft a response that needs approval. a research assistant might produce findings that require verification. an automated workflow might escalate an exception to a human operator.
that review has a cost, whether it comes from employee time, customer support operations, or a dedicated quality team.
research from ibm on agents in production highlights how much human involvement remains in deployed agentic systems. this is an important reminder that the cost of an ai product doesn’t necessarily end when the model generates an answer.
the difference between cost per request and cost per outcome
let’s say an ai product processes a request for $0.10.
that sounds inexpensive. but what happens if the system needs three attempts to complete the task? what if a human has to spend two minutes correcting the result? and what if the user abandons the workflow because the output isn’t useful?
the initial request cost tells us very little about the actual economics.
we need to distinguish between three measurements:
cost per request: the cost of processing an individual interaction.
cost per task: the total cost of completing a defined workflow, including all the operations involved.
cost per successful outcome: the total cost divided by the number of tasks that achieve an accepted result.
the last measurement is particularly useful because it connects the cost of operating the system to the value it is supposed to deliver.
[
\text{cost per successful outcome} =
\frac{\text{total delivery cost}}{\text{number of successful outcomes}}
]this changes how we evaluate model and architecture decisions.
a cheaper model that frequently produces unusable results might turn out to be more expensive than a capable model that completes the task reliably. fewer retries, less human correction, and higher completion rates can offset a higher cost per individual call.
the goal is to understand what it costs to deliver useful work, not simply what it costs to generate an answer.
and once we understand that number, we can start asking a more important business question: how much revenue does the product generate relative to the cost of serving each customer?
understanding ai unit economics
knowing how much an ai workflow costs is useful. knowing whether the business makes money from running that workflow is what makes the number meaningful.
this is where unit economics comes in.
unit economics looks at the revenue and costs associated with a specific unit of a business. depending on the product, that unit could be a customer, a subscription, a transaction, or a completed task.
for an ai product, the choice of unit matters. two customers paying the same subscription fee can generate very different infrastructure costs. two workflows that look identical to the user can also have different costs depending on the amount of reasoning, retrieval, and human intervention involved.
product teams need to understand both sides of this equation: how much revenue a customer generates and how much it costs to serve them.
start with the cost of serving a customer
imagine an ai customer support product that helps businesses resolve incoming support tickets. it reads a ticket, retrieves relevant account information, drafts a response, and either resolves the issue or escalates it to a human.
assume the average cost of processing a ticket looks like this:
the product costs $0.30 to process each ticket, including the expected cost of human intervention.
now imagine two customers paying the same $100 monthly subscription.
customer A processes 100 tickets per month. the delivery cost is $30, leaving $70 before other business expenses.
customer B processes 500 tickets. the delivery cost is $150, which means the company spends $50 more than it earns from that subscription before accounting for sales, marketing, research, and other operating expenses.
the pricing is identical. the economics are very different.
this is why average subscription revenue alone can be misleading when evaluating ai products. usage patterns can have a significant impact on the cost of serving individual customers.
understand gross margin
gross margin tells us how much revenue remains after accounting for the direct costs of delivering a product or service.
[
\text{gross margin} =
\frac{\text{revenue} - \text{cost of revenue}}{\text{revenue}}
\times 100
]in our example, customer A generates $100 in monthly revenue and incurs $30 in delivery costs.
that gives the company a gross margin of 70%.
customer B generates the same $100 but incurs $150 in delivery costs, resulting in a negative gross margin of 50%.
these numbers are illustrative, but the underlying problem is real: a pricing model that works for moderate usage can become unsustainable when customers consume significantly more resources.
it’s also worth distinguishing gross margin from contribution margin. gross margin accounts for the cost of revenue, while contribution margin can include additional variable costs associated with acquiring, supporting, or serving a customer, depending on how a business defines it.
both are useful, but they answer different questions.
measure the economics of successful work
there is another way to look at the same product.
instead of charging a flat subscription, imagine the company charges $1 for every successfully resolved ticket.
customer A processes 100 tickets and successfully resolves 80. customer B processes 500 tickets and resolves 350.
the economics look different now. revenue grows with successful work, giving the business a closer relationship between the value delivered and the amount it earns.
however, outcome-based pricing introduces its own challenges. what counts as a successful resolution? does the customer agree with that definition? how do we account for tickets that are reopened or require additional human intervention?
these are product and commercial decisions that need to be worked out before choosing the pricing model.
the metrics product managers should understand
a useful starting point is to track a small set of metrics together:
average revenue per customer: how much revenue each customer generates over a given period.
cost to serve: the direct cost of delivering the product to that customer.
gross margin: the share of revenue remaining after the cost of revenue.
cost per successful outcome: how much it costs to deliver an accepted result.
usage distribution: how consumption varies across customers and segments.
contribution margin: how much remains after relevant variable costs beyond the direct cost of revenue.
these metrics become more useful when viewed together. high usage might indicate strong adoption, but it can also expose expensive workflows. high completion rates might improve customer satisfaction, but the business still needs to understand whether the cost of achieving those outcomes is sustainable.
the objective is to connect product performance with business performance.
a product manager doesn’t need to become the company’s finance team. but understanding these numbers makes it easier to prioritize features, evaluate trade-offs, challenge pricing assumptions, and work with engineering and finance on decisions that affect the business.
once we understand how usage affects profitability, the next question becomes unavoidable: does the traditional saas pricing model still make sense for products with variable delivery costs?
why traditional saas pricing starts breaking down
per-seat pricing became popular because it offered a simple way to package and sell software.
a company could charge $30 per user per month, forecast recurring revenue, and expand its contract as more employees adopted the product. customers understood what they were paying for, and software companies had a relatively predictable revenue model.
this worked well for products where the number of users was a reasonable indicator of usage and value.
ai products complicate that relationship.
imagine two companies using the same ai assistant. both have 50 employees and pay for 50 seats. one company uses the assistant for occasional summaries and basic questions. the other uses it to analyse documents, generate reports, and automate complex workflows throughout the day.
the number of seats is identical. the amount of work performed, the value delivered, and the cost of serving each customer can be dramatically different.
this creates a mismatch between how software companies charge and how their products are consumed.
several pricing models are emerging to address this.
1. per-seat pricing
customers pay a recurring fee for every user who has access to the product.
why it works: revenue is predictable, customers understand the model, and procurement is relatively straightforward.
where it struggles: usage can vary significantly between customers. a handful of heavy users can generate substantial infrastructure costs without generating additional revenue.
per-seat pricing can still work for ai products, particularly when the value comes from providing access to a capability rather than charging for every interaction. the challenge is ensuring that the economics remain sustainable as usage grows.
2. usage-based pricing
customers pay according to how much of the product they consume. this could mean tokens, API calls, documents processed, minutes generated, or tasks executed.
why it works: revenue can grow alongside consumption, creating a closer relationship between usage and infrastructure costs.
where it struggles: customers may find their bills unpredictable. usage-based pricing can also discourage experimentation when customers worry that every interaction will increase their bill.
the pricing metric matters here. charging customers per token might reflect infrastructure consumption, but it rarely communicates the value they receive. customers generally care more about what the product helps them accomplish.
3. credit-based pricing
customers receive a predefined allowance of credits that can be consumed by different features or operations.
for example, a product might include 1,000 monthly credits, with simple tasks consuming fewer credits than complex workflows.
why it works: companies can package different levels of usage, introduce spending limits, and offer customers more predictable subscriptions.
where it struggles: credits can become difficult to understand when different features consume different amounts. customers may struggle to estimate how much usage remains or what a particular task will cost.
credits are useful when the underlying consumption varies, but the product needs to make the relationship between credits and customer value clear.
4. outcome-based pricing
customers pay for a defined result rather than access or consumption.
an ai customer support product, for example, might charge for successfully resolved tickets instead of the number of messages processed.
why it works: the pricing metric is connected more closely to the value delivered. customers pay for an outcome they care about, and the company has an incentive to improve completion rates.
where it struggles: defining and verifying successful outcomes can be difficult. the provider also takes on more performance risk because unsuccessful attempts may still incur infrastructure costs.
outcome-based pricing is attractive when results can be measured reliably and attributed to the product.
why hybrid pricing is gaining traction
in practice, companies don’t always need to choose one model.
a hybrid approach combines a recurring subscription with a usage allowance, credits, or additional consumption charges.
a product might charge $100 per month for access and include 500 ai tasks. customers who need more can purchase additional usage.
this gives the business a predictable revenue baseline while creating a way to recover additional delivery costs from heavy users. customers also get a clearer starting price and some control over their consumption.
research from openmeter’s 2026 pricing survey found that hybrid pricing was the most common primary model among the 230 software and ai companies surveyed, at 37%. respondents also expected adoption to increase over the following three years.
the shift is worth paying attention to. software companies are experimenting with pricing structures that reflect both access to the product and the amount of work it performs.
however, hybrid pricing introduces its own product decisions. how much usage should a subscription include? when should customers be notified that they’re approaching their limit? should additional usage be charged automatically or require an upgrade?
these decisions affect the customer experience as much as the company’s revenue.
ultimately, the right pricing model depends on how customers receive value, how predictable their usage is, and how much it costs the business to deliver that value. the next challenge is figuring out which of those things should determine what a customer actually pays.
designing pricing around customer value
choosing a pricing model starts with understanding what customers are paying for.
an ai product might consume tokens, execute api calls, process documents, or run hundreds of steps in the background. these activities determine the cost of delivering the product, but they don’t necessarily represent the value customers receive.
a customer using an ai research assistant doesn’t particularly care how many tokens were consumed while preparing a report. they care whether the report is accurate, whether it saves them hours of work, and whether it helps them make a better decision.
this distinction matters when choosing a pricing metric.
consider an ai customer support product. the company could charge customers in several ways:
per agent using the platform
per support ticket processed
per ai interaction
per successfully resolved ticket
each metric creates a different relationship between price, usage, and value.
charging per agent makes the product easy to budget for, but it doesn’t account for differences in workload. charging per interaction reflects consumption, but customers might pay more even when the interactions don’t produce useful results. charging per successful resolution connects the price more closely to the outcome, but requires a reliable way to define and verify success.
there isn’t one universally correct answer. the right metric depends on what customers value, how that value can be measured, and how closely usage correlates with the outcome.
start with the value delivered
before deciding how to charge, product teams should understand the economic value the product creates.
a few questions help establish this:
does the product save employees time?
does it reduce operational costs?
does it help customers generate more revenue?
does it improve the speed or quality of decisions?
does it reduce errors, risk, or manual work?
take an ai support assistant that resolves customer tickets.
if it saves a support team 500 hours every month, that time has an economic value. if it also improves resolution speed and reduces the backlog, the product creates additional operational benefits.
the pricing model should reflect a reasonable share of that value while leaving enough benefit for the customer to justify adopting the product.
this is the foundation of value-based pricing.
usage and value aren’t always the same thing
one of the biggest mistakes product teams can make is assuming that more usage automatically means more value.
a customer generating 10,000 ai interactions might be getting less value than another customer completing 1,000 meaningful workflows.
more interactions could indicate higher adoption, but they could also indicate repeated corrections, unnecessary prompts, or a product that requires too much effort to get the desired result.
this is why pricing around raw consumption can create unintended incentives.
if a product charges for every interaction, the business may benefit financially when customers use it more, even if those interactions don’t improve the outcome. customers, meanwhile, may become reluctant to experiment because every attempt has a cost.
a better pricing metric should reward the behaviour that represents genuine customer value.
make the price predictable
there is another part of pricing that deserves attention: the customer’s ability to understand and control their spending.
ai usage can be difficult to predict, particularly when tasks vary in complexity. a customer might know how many employees will use a product but have little idea how many model calls an automated workflow will trigger.
unexpected bills can quickly damage trust.
this makes usage visibility an important part of the product experience. customers should be able to understand their allowances, monitor consumption, receive notifications before reaching limits, and know what additional usage will cost.
stripe’s guide to usage-based pricing discusses the importance of choosing a pricing metric customers can understand and connecting that metric to the value they receive.
for product managers, this means pricing cannot be treated as a billing decision made after the product is built. it influences onboarding, feature discovery, usage limits, upgrade paths, and even how customers interact with the product.
the goal is to create a pricing model that makes economic sense for the business while remaining understandable and fair to customers.
that balance becomes one of the most important product decisions when building ai software.
the product manager’s role in ai unit economics
product managers have always had to balance customer value, engineering effort, and business priorities. ai adds another dimension to those decisions: the cost of delivering the experience.
a feature might solve a genuine customer problem and generate strong adoption, but if every interaction costs more than the revenue it generates, the business has a problem.
this doesn’t mean product managers should optimize every feature for the lowest possible inference cost. reducing costs at the expense of quality can make a product less useful. the responsibility is to understand the trade-offs and make deliberate decisions about where the business should spend its resources.
there are a few areas where this becomes particularly important.
1. prioritize features based on value and delivery cost
not every ai feature deserves to be built simply because it is technically possible.
a product manager should understand the customer problem, the expected value of solving it, and the resources required to deliver the solution.
consider two potential features for an ai assistant. one generates short summaries of customer conversations. another analyses large collections of documents and produces detailed research reports.
the second feature might create more value for a specific customer segment, but it could also require significantly more computation, retrieval, and verification.
the decision should account for both sides. how many customers need the feature? how frequently will they use it? what business outcome does it improve? and does the expected value justify the cost of building and operating it?
this is where prioritization becomes an economic decision, not simply a comparison of feature requests.
2. work with engineering on model selection
the most capable model isn’t necessarily the right model for every task.
a simple classification task might work perfectly well with a smaller, less expensive model. a complex reasoning task might justify a more capable model because the quality of the result has a greater impact on the customer.
product managers should work with engineering teams to understand where different models are appropriate.
this could involve routing simple requests to smaller models, reserving more capable models for complex tasks, using cached context where possible, and reducing unnecessary calls.
openai’s model optimization guidance provides examples of techniques that can improve efficiency while managing performance and latency.
the important point is to evaluate these decisions against product outcomes. a cheaper model that produces more errors and requires additional human correction might cost more overall.
3. design workflows that avoid unnecessary consumption
product design has a direct influence on operating costs.
poorly designed workflows can create unnecessary model calls, repeated requests, excessive context, and avoidable retries.
for example, an ai assistant that repeatedly asks users to provide information it already has is creating friction and consuming additional resources.
a better workflow might preserve relevant context, retrieve only the information needed, and break complex tasks into clearly defined steps.
these decisions can improve the customer experience while reducing the resources required to complete a task.
4. build usage visibility and controls into the product
when customers pay for ai usage, they need to understand what they’re consuming.
product managers should think about usage dashboards, spending notifications, monthly allowances, usage limits, and upgrade paths.
these controls help customers manage their costs and help businesses reduce unexpected infrastructure exposure.
the challenge is finding the right balance. overly restrictive limits can interrupt valuable workflows, while unlimited usage can create unpredictable costs.
the product should make consumption understandable without making customers think twice before using the features they pay for.
5. connect product metrics with financial metrics
traditional product dashboards might focus on activation, engagement, retention, and conversion.
ai products need additional measurements to understand whether usage is economically sustainable.
alongside traditional product metrics, teams should track:
cost per successful task
model and tool consumption per workflow
completion rates and retry rates
human intervention and correction rates
cost to serve each customer segment
gross margin by feature or customer cohort
these metrics help product managers understand whether an improvement in product performance is also improving the business.
the earlier article on measuring success in ai products explores why measuring outputs alone isn’t enough. the same principle applies to economics: the number of interactions tells us how much the product is being used, but not whether that usage creates sustainable value.
6. bring pricing into product strategy
pricing decisions influence how customers discover, adopt, and use a product.
a feature that consumes significant resources might need a usage allowance, a separate paid tier, or a different pricing metric. a feature that creates substantial customer value might justify a premium offering.
product managers should work with finance, sales, and engineering to understand these trade-offs.
this also means thinking about pricing early in product development. if a feature’s economics only work under a particular usage pattern, the product experience and packaging should account for that from the beginning.
ultimately, product managers need to understand the relationship between what a product promises, what it costs to deliver, and what customers are willing to pay.
the goal is to build products that customers want to use, businesses can afford to operate, and teams can continue improving as adoption grows.
a practical framework for evaluating the economics of an ai feature
understanding ai product economics is useful, but product teams need a way to apply that thinking before a feature reaches production.
a feature can have a strong value proposition, promising early adoption, and impressive model performance. none of those guarantees that the feature will be economically sustainable at scale.
here’s a six-step framework product managers can use to evaluate the economics of an ai feature.
step 1: define the customer outcome
start with the job the feature is expected to complete.
avoid defining success as generating a response, processing a document, or making a model call. those describe system activity, not necessarily customer value.
instead, identify the outcome the customer expects.
for an ai customer support assistant, that might mean resolving a ticket without human intervention. for a research assistant, it could mean producing a report that meets the user’s requirements and needs minimal correction.
define what counts as a successful outcome and how the team will verify it.
step 2: estimate the full delivery cost
map the workflow from the initial request to the final outcome.
include model inference, retrieval, external tools, orchestration, retries, infrastructure, and expected human intervention.
don’t stop at the cost of the first model call. complex workflows can involve several operations before the customer receives something useful.
work with engineering to estimate the average cost, but also consider how much that cost varies across simple and complex tasks.
step 3: measure completion and quality
a low-cost workflow is of little value if it regularly produces unusable results.
measure how frequently the system completes the task successfully, how often users accept its output, and how much correction or human intervention is required.
this is where evaluation becomes important. teams need reliable ways to determine whether the system is producing the intended outcome.
the guide on how to evaluate ai products covers the release-readiness side of this problem.
step 4: calculate unit economics
connect the cost of delivery to the revenue generated.
look at cost per successful outcome, revenue per customer, cost to serve, and gross margin.
segment the numbers where possible. average costs can hide important differences between customer types, usage patterns, and workflows.
a feature might be profitable for occasional users and unprofitable for customers running thousands of complex tasks every month.
step 5: choose a pricing model
once the team understands the value delivered and the cost of delivering it, evaluate the pricing model.
should customers pay per seat, per task, through credits, or based on a successful outcome?
consider whether the pricing metric reflects customer value, whether usage is predictable, and whether the business can recover its delivery costs.
the model should make sense to customers and remain sustainable as usage grows.
step 6: monitor economics after launch
unit economics are not fixed at launch.
customers change how they use products. models improve, workflows become more complex, and infrastructure costs fluctuate. a feature that looks profitable during a pilot can have very different economics once adoption increases.
monitor cost per successful outcome, customer-level usage, completion rates, human intervention, and gross margin over time.
set thresholds that trigger investigation when costs increase or performance deteriorates. product teams should be able to identify whether the change comes from heavier usage, a more expensive workflow, lower completion rates, or an underlying infrastructure change.
putting the framework into practice
consider the customer support assistant from our earlier example.
the team knows that each ticket costs an average of $0.30 to process. it also knows that not every ticket gets resolved successfully.
before launching a new pricing plan, the pm can use the framework to answer six questions:
the framework helps bring product, engineering, and finance into the same conversation.
it also changes the launch decision. instead of asking only whether the feature works, the team can evaluate whether it delivers enough value, at a cost the business can sustain, under a pricing model customers are willing to accept.
that’s a much stronger basis for deciding what to build and ship.
what happens as ai costs continue to fall?
model costs have been declining, and improvements in hardware, infrastructure, and model efficiency are making increasingly capable ai systems more accessible.
this creates opportunities for software companies. features that were previously too expensive to operate can become commercially viable. products can handle more complex tasks, serve larger customer bases, and introduce capabilities that would have been difficult to justify a few years ago.
but lower inference costs don’t automatically translate into higher margins.
when the cost of a model call falls, companies often find new ways to use that additional capacity. a simple assistant becomes an agent that can research, plan, retrieve information, and execute actions. workflows that previously required one model call might now involve several.
the product becomes more capable, but the amount of work it performs can increase too.
there is also a competitive dimension. as access to capable models becomes easier, more companies can build similar features. capabilities that once differentiated a product can quickly become standard expectations.
this puts pressure on companies to find other sources of differentiation.
the value moves closer to the workflow
when underlying model capabilities become widely accessible, the surrounding product experience becomes increasingly important.
customers still need software that understands their context, connects to their existing tools, handles exceptions, and completes work reliably.
consider two ai products built on similar models. one generates recommendations. the other understands the customer’s business data, connects to their existing systems, executes approved actions, and fits into an established workflow.
the second product may create substantially more value because it solves a larger part of the customer’s problem.
this is where integrations, proprietary context, distribution, and workflow ownership become important.
we explored a related idea in the article on ecosystem-led product growth in B2B saas. integrations and ecosystem relationships can make a product more useful while creating advantages that are harder to replicate than an individual feature.
efficiency becomes a competitive advantage
as model costs fall, companies have an opportunity to improve their margins. but capturing that opportunity requires deliberate product and engineering decisions.
teams that understand their cost structure can route tasks to appropriate models, reduce unnecessary computation, improve retrieval, and design more efficient workflows.
they can also decide where spending more makes sense.
a more expensive model might be justified for a high-value task where accuracy matters. a simpler model might be sufficient for routine classification or summarisation.
the objective is to allocate resources according to the value each task creates.
the next generation of ai businesses will need stronger economics
ai capabilities will continue to improve, and the cost of accessing them will likely keep changing.
some companies will pass efficiency gains to customers through lower prices. others will invest those savings in more capable workflows. many will need to balance both.
the businesses that build sustainable advantages will understand their customers, own important workflows, deliver measurable outcomes, and maintain a cost structure that supports their pricing model.
lower model costs create more possibilities. building a sustainable business still requires good product decisions.
building ai products that make business sense
building an ai product requires more than choosing a capable model and creating a useful experience. the business also needs to understand what it costs to deliver that experience, how much value customers receive, and whether the economics hold up as usage grows.
for product managers, this means bringing commercial thinking into product decisions much earlier.
feature prioritization, model selection, workflow design, reliability, usage limits, and pricing all influence the economics of an ai product. treating these as separate concerns can lead to products that perform well in demos but become difficult to operate profitably in production.
the strongest product teams will understand the relationship between three things: the value they create for customers, the cost of delivering that value, and the revenue they can capture from it.
because ultimately, the success of an ai product depends on more than how intelligent it appears. it depends on whether it solves a meaningful problem and whether the business can keep delivering that solution sustainably.
frequently asked questions
1. what are the unit economics of an ai product?
unit economics measure the revenue and costs associated with serving a customer, completing a task, or delivering a specific outcome. for ai products, this includes inference, retrieval, tool execution, infrastructure, and human intervention. understanding these costs helps businesses determine whether their products can remain profitable as usage grows.
2. how do you calculate the cost of an ai feature?
start by mapping the complete workflow, from the initial user request to the final result. include model inference, input and output tokens, retrieval, external api calls, orchestration, retries, infrastructure, and human review. divide the total delivery cost by the number of successful outcomes to understand the cost of delivering useful work.
3. why do ai products have lower gross margins than traditional saas?
ai products often incur variable costs every time customers use them. model inference, data retrieval, tool execution, and other operations can increase with usage. traditional saas products generally have lower marginal delivery costs, although their actual margins also depend on hosting, support, infrastructure, and other expenses.
4. what is the best pricing model for ai products?
there is no universally best model. per-seat pricing works when access to the product represents the primary value. usage-based pricing can work when consumption correlates with value and delivery costs. credit-based and hybrid models provide more flexibility, while outcome-based pricing can align revenue with measurable customer results.
5. should ai products charge per user, usage, or outcome?
the decision depends on how customers receive value and how predictable their usage is. per-user pricing offers predictable subscriptions, usage pricing connects revenue to consumption, and outcome pricing connects revenue to completed work. product teams should choose a metric customers understand and that supports sustainable delivery economics.
6. how can product managers reduce ai inference costs?
product managers can work with engineering teams to select appropriate models for different tasks, reduce unnecessary model calls, optimize retrieval, reuse cached context, limit unnecessary retries, and design more efficient workflows. cost reductions should always be evaluated alongside quality, reliability, latency, and customer outcomes.

