.png)
Frameworks, core principles and top case studies for SaaS pricing, learnt and refined over 28+ years of SaaS-monetization experience.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
When AI Agents Negotiate the Deal: Rethinking B2B Pricing
Procurement is becoming software-driven
For decades, B2B pricing has been designed around human behavior. Vendors publish incomplete rate cards, sales representatives qualify buyers before revealing prices, procurement teams collect a manageable number of bids, and both sides negotiate through calls, emails and spreadsheets. Discounts depend partly on economics—and partly on timing, relationships, persistence and information asymmetry.
That system begins to break when the buyer is software.
AI agents are moving beyond summarizing contracts or recommending suppliers. They are beginning to identify vendors, construct sourcing events, compare proposals, check compliance, negotiate terms, execute purchases and monitor renewals. Oracle now offers an Autonomous Sourcing Agent for competitive bidding on lower-value, high-volume purchases. SAP is introducing sourcing and negotiation agents in Ariba. Ramp has launched agents that can triage requests, research vendors and review commercial terms. In February 2026, project44 introduced an agent that continuously benchmarks freight rates and negotiates with carriers.
The adoption curve is still early. Gartner estimates that only a small minority of agentic projects are mature and predicts that more than 40% will be canceled by the end of 2027. Yet it also expects 33% of enterprise applications to include agentic capabilities and at least 15% of routine work decisions to be made autonomously by 2028. Its advice is telling: companies should pursue agents where they improve enterprise productivity, “rather than just individual task augmentation.”
The central pricing implication is not simply that agents will negotiate harder. Agent-led procurement will shift pricing power away from negotiated opacity and toward machine-verifiable value. Prices, contract terms, service levels and performance evidence will increasingly need to be structured so that software can discover, compare and act on them. Conventional discounting will lose effectiveness where it depends on buyer fatigue or imperfect information. At the same time, suppliers that can prove differentiated outcomes will gain new opportunities to charge for results rather than access.
Agentic procurement may therefore increase price pressure on interchangeable offerings while expanding pricing power for vendors whose value can be measured and defended.
The old B2B pricing playbook
Traditional enterprise pricing evolved around constrained human attention.
A procurement team cannot evaluate every possible supplier. A buyer may collect three or five bids, negotiate with the finalists and accept substantial complexity in the resulting contracts. Large organizations frequently concentrate their negotiation resources on strategic suppliers, leaving tail-spend agreements on standard or lightly negotiated terms.
Walmart’s experience illustrates the constraint. With more than 100,000 suppliers, the company could not justify focused human negotiations with every vendor. Approximately 20% of suppliers were historically placed on standardized terms because the cost of adding human negotiators would have exceeded the expected benefit. Walmart subsequently used Pactum’s automated negotiation technology for parts of its tail spend. Its initial program reportedly reached agreements with 64% of participating suppliers, produced average savings of 1.5% and extended payment terms by an average of 35 days.
This human capacity constraint supports several familiar pricing practices:
- List prices that serve mainly as negotiation anchors
- Complex packaging that makes direct comparison difficult
- Large discretionary discount ranges
- Expiring offers designed to create urgency
- Auto-renewal clauses that benefit from inattention
- Bundles that obscure the economics of individual products
- Seat commitments that customers rarely revisit in detail
- Different prices for similar customers based on negotiating skill
These approaches do not always reflect bad intent. Enterprise products can be genuinely difficult to configure and price. Implementation requirements, risk allocation and service levels differ by customer. But complexity also creates room for price dispersion. Two similar customers can pay materially different effective rates because one had better benchmarks, more time or a stronger procurement function.
Agent-led procurement attacks that capacity constraint. An agent can potentially inspect every renewal, evaluate hundreds of suppliers and run negotiations simultaneously. Procurement software provider Vertice, for example, introduced an autonomous software negotiation agent in September 2026 after reporting that one in seven software contracts in its dataset auto-renewed without oversight. Whether its advertised results generalize across the market remains to be demonstrated, but the product direction is clear: previously ignored spend is becoming economically negotiable.
A human buyer and a procurement agent do not necessarily optimize the same way.
Human buyers operate under time pressure. They may prefer an incumbent because switching is disruptive, favor a well-known brand because it reduces career risk, or accept a suboptimal bundle because analyzing every component would take too long. They can also value trust, responsiveness and the confidence inspired by a strong account team—factors that may be difficult to represent numerically.
An agent can be instructed to assess a much broader dataset: unit price, total expected consumption, historical utilization, implementation costs, payment terms, security requirements, contractual liability, supplier performance, switching costs and the probability that a promised outcome will actually occur. It can model scenarios continuously rather than only when a sourcing event or renewal is scheduled.
That does not make the agent objectively rational. It means its behavior depends on the objective function, information and limits it receives. An agent told to minimize annual expenditure may choose differently from one told to maximize expected value over three years. A system penalized heavily for compliance risk may reject the cheapest offer. A poorly configured agent may optimize a measurable proxy while ignoring an important but less quantifiable consideration.
Recent research reinforces this point. An August 2026 study of LLM agents negotiating supply-chain contracts found that model identity affected how negotiating surplus was divided, while the agent’s prompted level of strategic patience was the strongest driver of the outcome. Google researchers have similarly found that LLMs can exhibit meaningful strategic behavior in bilateral bargaining, but negotiation remains an imperfect-information problem involving beliefs, incentives and reservation values—not merely a comparison of list prices.
Procurement policy therefore becomes executable strategy. Finance and procurement leaders must specify what the agent is authorized to optimize:
- Lowest immediate cost or lowest total cost of ownership
- Maximum discount or maximum expected ROI
- Short-term flexibility or long-term price protection
- Best technical performance or acceptable performance at lower cost
- Supplier diversification or vendor consolidation
- Fast deployment or lower implementation risk
- Standard terms or strategically negotiated exceptions
Human procurement encodes these trade-offs informally through experience and judgment. Agentic procurement requires organizations to encode them explicitly.
Before software can negotiate effectively, it must be able to understand what is being sold.
Most B2B pricing is not ready for that. Important details sit across proposal documents, online calculators, order forms, contracts and sales emails. Usage units may have ambiguous definitions. Discounts may apply only to certain products or years. Service limits and overages may be buried in legal schedules.
A human buyer can call the sales representative for clarification. An autonomous agent needs structured information.
The early infrastructure is already emerging. Google’s Agent2Agent protocol was designed to let agents discover capabilities and communicate across different systems. Its Agent Payments Protocol provides a common framework for transmitting authority, intent and accountability in agent-led transactions. Stripe’s Agentic Commerce Suite allows businesses to expose near-real-time catalog, price and availability data through hosted endpoints. Visa and Mastercard have introduced infrastructure for agent identity, transaction controls and authorized agent payments.
These initiatives currently emphasize commerce and payments, but the architectural direction has direct B2B relevance. A supplier prepared for procurement agents will eventually need to expose more than a webpage and a PDF rate card. It may need to provide:
- Product and package definitions
- Metering rules and billable-event definitions
- Volume tiers and commitment thresholds
- Contract duration and renewal rules
- Overage, rollover and credit-expiration policies
- Implementation and support requirements
- Security certifications and data-residency options
- Service-level commitments
- Approved discount logic
- Performance and reliability evidence
- Machine-readable contract clauses
- APIs through which agents can request quotes or negotiate
This does not mean every vendor must reveal its absolute price floor. It means ambiguity itself will carry a commercial cost. If an agent cannot determine the expected bill, compare contractual risk or verify a product’s fit, the vendor may not make the shortlist.
Jack Forestell, Visa’s chief product and strategy officer, summarized the trust requirement succinctly: agents must be trusted with payments “not only by users, but by banks and sellers as well.” In B2B markets, that trust will extend to authority, auditability and the ability to explain why an agent selected or rejected a supplier.
Much of enterprise discounting is built around scarcity: limited procurement bandwidth, limited pricing information and limited time before a purchasing deadline. Agents weaken all three constraints.
Traditional tactics such as a “20% discount if signed by Friday” will be less persuasive when the buyer agent can immediately compare the net present value of the offer with hundreds of alternatives. A large percentage discount from an inflated list price may carry little weight if the agent benchmarks the resulting net price against market data. Bundling an unwanted module to preserve the headline discount may also fail when utilization data shows that the customer is unlikely to use it.
But agents will not necessarily eliminate discounts. They will make discounts more conditional and computational.
Instead of a sales representative selecting a concession from a broad discretionary range, a seller agent could calculate a concession based on expected volume, contract duration, payment timing, retention probability, implementation burden and cost-to-serve. A buyer agent could respond with different combinations of price, commitment and risk allocation. Negotiation may become a rapid search for the mutually acceptable combination rather than a sequence of theatrical concessions.
The Walmart program demonstrates the potential for multi-variable bargaining. The system negotiated not only prices but also payment terms, allowing suppliers to exchange concessions according to their own priorities. An experimental study of procurement chatbots likewise found that competitive negotiation strategies could produce larger discounts, better payment terms and faster agreements for appropriate categories.
This suggests that machine-to-machine negotiation will be most effective where:
- Requirements can be stated clearly.
- Several suppliers are genuinely substitutable.
- Performance can be measured consistently.
- Acceptable terms and escalation thresholds are predefined.
- The deal has enough value to negotiate but not enough strategic complexity to require executive judgment.
Tail spend, standardized software, logistics lanes, routine professional services and repeatable sourcing events are natural early candidates. Complex transformations, strategically important partnerships and novel risk-sharing agreements will retain substantial human involvement.
More transparency does not automatically mean lower prices, however. Pricing algorithms can increase price dispersion or respond rapidly to competitors in ways that soften competition. The OECD has also warned that shared algorithms and highly transparent market information can create risks of coordination or personalized pricing. The likely outcome is therefore not one universal market price. It is faster, more granular price discovery—potentially accompanied by more sophisticated price discrimination on both sides.
Our Agentic Monetization Spectrum evaluates pricing using three dimensions: Zero-Human Ability, or how independently the agent works; Operational Domain, or how broad a workflow it controls; and the Output/Cost Curve, or how rapidly customer value grows relative to delivery cost. As autonomy, scope and value-to-cost leverage increase, the appropriate unit of value generally moves from human access toward actions, outputs and outcomes.
Agent-led procurement makes this metric choice even more consequential because the buyer can test whether the metric tracks value.
Per-seat pricing
Per-seat pricing remains viable when an AI product primarily assists a person. If every employee uses an AI copilot, access is still meaningfully connected to value.
It becomes weaker when one agent performs work previously completed by many employees. The customer may eliminate three analyst seats while increasing the work produced by the system. A vendor that continues charging only by headcount loses revenue precisely when its product becomes more effective.
PwC has argued that the traditional SaaS playbook must change when an agent can perform work previously requiring several users. Strategy& similarly distinguishes between AI that enhances an existing job—where seats and add-ons may survive—and AI that transforms the job, where usage or outcome models are more appropriate.
Usage-based pricing
Usage pricing works when activity correlates reasonably with value and vendor cost. Examples include API calls, agent actions, workflow runs, generated assets or compute units.
Salesforce currently sells Agentforce Flex Credits at $500 per 100,000 credits, with a standard agent action consuming 20 credits, or $0.10. Cognition’s 2026 Devin plans combine subscriptions, included quotas and usage charges, while its enterprise contracts continue to use Agent Compute Units.
The weakness of usage pricing is budget uncertainty. A capable procurement agent will model not only the rate but expected activity, variance, overage exposure and the possibility that a vendor’s product design encourages unnecessary consumption. Technical metrics such as tokens are particularly difficult because buyers do not intrinsically value them.
Transaction-based pricing
Transaction pricing fits agents that complete a discrete commercial event: an invoice processed, shipment booked, payment collected or purchase completed. The metric is understandable and auditable.
However, the vendor must prevent low-value transactions from consuming the same resources as high-complexity ones. Tiered transaction rates, complexity multipliers or minimum platform fees may be required.
Outcome pricing is the strongest value-alignment model when the outcome is clearly defined, attributable and economically meaningful.
Intercom currently charges from $0.99 per Fin outcome. HubSpot’s credit system translates its Customer Agent activity into an effective charge of $0.50 for a resolved conversation. Sierra markets its customer-service agents on the principle that customers should “pay for a job well done.”
An agentic buyer is likely to favor outcome pricing because the vendor assumes more performance risk. But it will also scrutinize the definition. What constitutes a resolution? How are reopened cases treated? Who receives credit when AI and employees both contribute? Can the customer independently audit the event?
Outcome pricing should not be treated as the inevitable destination for every AI company. As our POV emphasizes, the pricing metric has to survive the product’s reliability. If successful outcomes are difficult to attribute or the agent still requires significant review, usage or output pricing is more defensible. Aspiration belongs in the product roadmap, not the rate card.
Hybrid pricing
For many vendors, the most resilient model will combine:
Hybrid pricing provides the vendor with recurring revenue and cost coverage while allowing price to scale with delivered work. It also gives the buyer a predictable baseline and explicit marginal economics.
The seat does not always disappear. It becomes a cover charge rather than the primary value engine.
When agents can compare more suppliers, weak differentiation becomes easier to expose.
Generic claims such as “enterprise-grade AI” or “industry-leading automation” carry little weight when a procurement agent can examine reliability, integration requirements, customer retention, security posture, implementation time and realized outcomes. Vendors will need evidence that can be evaluated computationally.
Defensible differentiation may include:
- Higher task-completion or resolution rates
- Lower human-review requirements
- Faster implementation
- Better performance in a defined industry or workflow
- Lower error, hallucination or exception rates
- Proprietary workflow data
- Stronger permissions, controls and audit trails
- Lower total compute or integration costs
- Contractual service guarantees
- Easier switching, migration and interoperability
This is where a vendor’s agentic harness becomes commercially important. Models will continue to improve and may become easier to substitute. The orchestration, memory, permissions, domain logic, integrations and governance surrounding the model determine whether the system can perform valuable work reliably. A stronger harness supports a higher-value pricing metric because it makes outcomes more repeatable and defensible.
In commoditized categories, agents will intensify price competition. In differentiated categories, they may do the opposite: identify that a higher-priced supplier produces a better risk-adjusted outcome. Value-based pricing does not disappear when software buys from software. It becomes more evidence-dependent.
Autonomous purchasing without governance would be financially dangerous. The appropriate model is bounded autonomy.
The agent should have explicit authority by category, supplier, transaction size, contract term and risk level. Spending thresholds, pricing floors, approved clauses and escalation triggers should be encoded before the agent begins negotiating. Every action should produce an audit record showing the information considered, rules applied, offers exchanged and reason for the final recommendation.
Payment infrastructure is already moving in this direction. Visa allows spending conditions to be attached to agent credentials. Mastercard emphasizes registered agents, verifiable intent and traceability. Google’s AP2 protocol was designed around authorization, authenticity and accountability.
Usage is another risk. The more successful an autonomous workflow becomes, the more frequently it may run—and the larger the bill may grow. Customers will demand budgets, real-time consumption dashboards, anomaly detection, hard limits and the ability to pause agents. Salesforce and HubSpot have both introduced digital-wallet or credit-management controls around agent consumption.
Outcome models introduce attribution disputes, while autonomous negotiation introduces security and competition risks. Agents may be manipulated by instructions hidden in documents, disclose information they were not authorized to share or converge on pricing behavior that attracts regulatory attention. Human oversight therefore remains essential for strategic deals, unusual contractual language, sensitive data and decisions that could materially affect the business.
What pricing and procurement leaders should do now
CFOs should treat agent consumption as a new variable cost category. Forecasts should connect agent activity to business volumes, expected outcomes and underlying compute. Finance teams need unit economics at the action or outcome level—not merely an annual software budget.
CROs should redesign discount governance before seller agents inherit it. Broad discretionary discount bands should be replaced with explicit rules tied to commitments, economics, risk and strategic value. Otherwise, automation will scale inconsistent pricing rather than fix it.
Pricing leaders should identify the real unit of value. Ask what remains after human seats decline. It may be a completed workflow, qualified opportunity, shipment, resolved case, approved invoice or percentage of a verified financial benefit. The best metric is not the most fashionable one; it is the closest-to-value metric that customers understand, the product can meter and its current reliability can support.
SaaS companies should create agent-readable commercial infrastructure. Package definitions, rate cards, entitlement rules, usage events, thresholds and contract terms must become structured data. Monetization engineering—the connection between entitlements, metering, rating, billing, revenue recognition and ERP—must become a first-class product capability.
Procurement leaders should begin with bounded, repeatable categories. Deloitte emphasizes that governance, interoperability, data quality and human oversight are prerequisites for multiagent procurement. Its 2025 CPO survey found that leading “Digital Masters” were allocating up to 24% of their procurement budgets to technology and reporting approximately 3.2 times the GenAI investment returns of less mature peers.
Finally, both sides should prepare for agent-to-agent negotiation. McKinsey’s 2025 survey of more than 400 pricing executives found that only around 5% to 10% had fully scaled an agentic pricing use case, but 65% to 85% expected to adopt generative or agentic pricing capabilities within one to three years. Buying agents will increasingly encounter seller agents that can generate quotes, enforce discount policy and produce contract terms autonomously.
The strategic question is no longer whether a company will deploy agents. It is which decisions they will be allowed to make—and whether the organization’s pricing and procurement architecture is ready for those decisions.
The future of pricing when software buys from software
AI agents will not make all B2B products commodities. They will make unsupported differentiation harder to sustain.
When buyers can evaluate more alternatives, continuously monitor performance and reopen negotiations at lower cost, vendors will have less room to rely on opacity, inertia and unused commitments. Machine-readable pricing, auditable outcomes and transparent contract logic will become competitive requirements.
Yet the same development can strengthen value-based pricing. An agent capable of detecting a cheaper offer is also capable of calculating why the more expensive supplier delivers better uptime, lower risk or greater economic impact. The winners will not necessarily be the vendors with the lowest prices. They will be the vendors whose value survives structured comparison.
That is the deeper shift. In the human-led procurement era, a persuasive commercial story could compensate for incomplete evidence. In the agentic era, the story must become data.
When software buys from software, price will increasingly be negotiated as a function of verified work, risk and results. The unit of value will move from who logs in to what gets done.