AI functionality is now live in most B2B SaaS products. The P&L impact is not. In our recent survey of 128 software CEOs, CFOs, CPOs, and CMOs, a quarter of companies still bundle AI into existing subscriptions at no charge. Among those that do charge, 79% have attach rates under 50%, and for 82% of companies with paid AI tiers, AI-specific pricing represents less than 10% of total ARR.
The gap between “we have a pricing model” and “AI is showing up in the P&L” is the story most boards are struggling to absorb. It happened for an understandable reason: product shipped because the market demanded it, pricing was set quickly because launch timelines were moving, and the harder work was deferred. Only 27% of companies have formal, enforced margin guardrails on AI pricing. 72% priced AI with informal research or none at all. Just 8% said willingness-to-pay research primarily drove the decision.
Executives know this is temporary. 96% expect to change their AI pricing model, and nearly a third want to but have not yet determined how. Even the largest vendors are still searching: Salesforce has cycled through four AI pricing revisions in twenty-four months, and one of its senior executives said in May 2025 that any business claiming to have figured out AI pricing is kidding itself. The difference is that most companies do not have Salesforce’s room to experiment in public.
Why the window is closing
Three pressures make waiting expensive.
First, cost. AI features look like software to the customer but behave like a variable-cost service to the vendor. Every query and workflow creates inference cost that scales with usage. Yet only 23% of companies have linked AI feature costs to customer-level delivery, which means most cannot say which accounts are margin-accretive and which are being quietly subsidized. This is no longer a private problem: 90% of executives say AI-driven gross margin compression has already surfaced, or will within six months, with their board, investors, or in diligence.
Second, renewals. Pre-AI contracts contain no provisions for AI, so every renewal that passes without a strategy locks in bad economics for another two to three years. The friction is already here: 37% of vendors have faced pushback on AI-tied price increases, and 24% have had customers argue that AI should reduce their bill. Combined, 60% have hit meaningful renewal friction tied to AI.
Third, non-human identities. AI agents, bots, and service accounts already outnumber human users 50 to 1 in many enterprise environments, and 57% of vendors know or believe their customers have non-human users generating value from their software. Of those, only 22% have explicit pricing for it. Seat-based models were built for a world where value tracked with people. Procurement teams have noticed, and are arguing that automation should shrink seat counts and bills.
Companies that move now can still shape buyer expectations. Companies that wait will reprice under duress: at renewal after the customer has anchored to the old model, after an AI-native competitor has redefined what the category should cost, or in diligence when buyers set the terms.
Based on our work with hundreds of software companies, five requirements determine whether AI pricing shows up in the P&L or gets left on the table. They function as a system, not a checklist.
Key 1: Quantify credible ROI
Buyers stopped accepting generic AI claims after last year’s wave of inflated ones. They now expect concrete, dollar-quantified outcomes, an expectation set by headline cases like IBM attributing roughly $3.5 billion in productivity gains to AI agents. A credible ROI claim, one the buyer can stress-test, justifies higher prices and accelerates deals. An overstated one gets you disqualified.
The strongest path is partnering with early customers to validate realized ROI and securing them as references. Where that data does not exist yet, an independent third-party ROI study, confirmed through buyer research, works nearly as well. Either way, capture the product telemetry needed to substantiate the claims over time. The same data supports customer success, retention, and expansion at renewal.
Key 2: Price to value, with cost as the floor
Inference cost is the first real variable cost in modern software, and the reflexive CFO response is usage-based pricing. That instinct fails in B2B SaaS for two reasons: customers cannot forecast their own usage, which suppresses willingness to pay, and usage frequently diverges from value. A customer burning twice the tokens does not necessarily get twice the benefit.
Usage pricing works for the frontier model providers because at the API layer, usage is the value and buyers are technical. Neither condition holds when the buyer is a business user and value arrives through workflows. In one of our buyer surveys, exactly 0% of B2B software buyers preferred a token-based pricing metric.
The better construct is a value-based price that includes a meaningful usage baseline in each tier, with overage charges beyond it, structured like a mobile data plan. COGS belongs in the analysis as a floor and a risk input for sizing bundles, not as the basis price is built up from. Cost-plus pricing caps you at reseller economics and hands the value premium back. And define usage in terms buyers understand: tickets resolved, documents generated, contracts reviewed. Never tokens.
Key 3: Package AI for adoption
Today, Microsoft, Salesforce, HubSpot, and Atlassian all sell AI as a separate premium SKU. That logic is sound while AI is still perceived as new. It will not stay sound. As AI becomes embedded across categories, a non-AI bundle becomes a structurally inferior product, and the standalone SKU turns into an awkward forced upsell that drags on win rates.
The durable answer is integrating AI into core Good/Better/Best tiers, with placement driven by relative value: commoditized capabilities anchor the entry tier, differentiated high-ROI capabilities earn premium placement. Two cautions. Packaging cascades into unit economics, so inference-heavy features should never sit in an entry tier without usage caps. And visible AI (copilots, generative tools) works as a tiering lever, while embedded AI (routing, classification, anomaly detection) is better treated as a product-wide quality uplift.
Key 4: Choose the right metric and structure
More companies will revisit their pricing metric in the next 12–24 months than did during the entire decade-long shifts to per-seat or consumption pricing. Seat-based models fail the value test when an agent replaces five users; the bill should not drop 80% while value holds or rises.
The best-fit metric satisfies five criteria: it correlates with value, trends upward over time, is easy for customers to forecast, is auditable by both sides, and is simple to explain. Correlation with willingness to pay is the decisive test. When it is weak, the same metric prices some customers far too high and others far too low, which erodes credibility and win rates alike.
Outcome-based pricing sounds ideal and usually underperforms: buyers and vendors rarely agree on the outcome’s definition, buyers withhold full attribution, and the outcome often does not track willingness to pay. Pressure-test all three before committing. For most B2B SaaS companies, the answer is a hybrid: a fixed platform fee for predictability, plus a value-correlated variable component that captures upside and protects margin.
Key 5: Set prices through willingness-to-pay research
The first four keys design the model. This one calibrates it. Pricing set through internal debate, stale benchmarks, or sales intuition is dangerously imprecise when buyer expectations and competitive offers move quarterly. The standard is a blinded buyer survey, in which respondents do not know who sponsored the research, analyzed with quantitative methods such as conjoint and Gabor-Granger. Done well, it yields optimal list prices by segment, discount guardrails below which deal economics break, validation of the metric and packaging choices, and cohorts that warrant different prices or motions. In the current environment, refresh foundational research annually rather than every two to three years.
The money is in the base
Here is the finding that should reorder most companies’ plans: only 37% of pricing-program impact comes from new-logo price increases. Over half comes from renewal upsell and cross-sell into existing customers. Capturing it requires the infrastructure to be in place before the renewal conversation starts: ROI evidence, research-derived renewal guardrails, pre-mapped expansion offers per cohort, and sequenced fallback plays so outcomes are determined by the program rather than by individual judgment under pressure. It also requires staffing renewals for commercial negotiation. Procurement has professionalized its opening demands; most customer success teams were not built for that conversation.
The decisions software companies make on AI pricing in the next twelve months will shape unit economics, valuation multiples, and exit outcomes for years afterward. The framework is knowable. The window is not indefinite.
