Underwriting AI Disruption in Software: Five Tests for Differentiation, Retention, and Pricing Power

AI disruption in software should be underwritten as a change in who controls the customer’s work.

The central question is not whether a target has launched an assistant, added generative features, or published an AI roadmap. It is whether AI allows another company to mediate, complete, or reshape the workflow from which the target derives its differentiation, retention, and pricing power.

An incumbent can sell AI features into its installed base without becoming the long-term controller of workflow state, execution, and decision-to-outcome data. Jefferies, drawing on more than 100 conversations with senior technology investors, an analysis of recent control transactions, and dialogue with buyers and bidders, reports that investors increasingly want tangible evidence of defensibility against AI disruption rather than theoretical incumbent advantages. ^[1]

For an investor, this work should produce explicit underwriting conclusions:

• Which revenue streams face substitution, consolidation, or seat compression?
• Which switching costs survive if the user interface changes?
• What additional product investment and AI delivery cost will be required?
• Can the company capture more of the customer outcome and improve its economics?
• Does the evidence support investing, repricing, partnering, acquiring, repositioning, harvesting cash flow, or changing exit timing?

Testing the target’s own AI claims is related but different. That exercise asks whether its AI is real, scalable, compliant, and economically useful, as covered in EES’s guide to AI due diligence for private equity. Disruption underwriting asks whether AI changes the competitive and economic basis of the asset, including parts of the business that do not carry an AI label.

Five tests for underwriting AI disruption

Test 1: Where does the product create and control value?

A system of record can remain important without controlling the workflow around it. Underwriting should distinguish among six positions:

1. Recordkeeping: Storing authoritative customer, transaction, operational, or compliance data.
2. Decision support: Helping a person interpret information or choose an action.
3. User interface: Providing the primary environment in which users interact with the workflow.
4. Orchestration: Coordinating data, approvals, systems, people, and exceptions.
5. Execution: Completing actions rather than merely recommending them.
6. Proof of outcome: Recording what happened and connecting the decision to measurable results.

The more of this chain the target controls, the more options it may have to defend or expand its position. But the analysis must be specific. A product may own the underlying record while an adjacent platform or horizontal agent becomes the interface, makes recommendations, and initiates actions. The incumbent remains installed but loses engagement, influence, and eventually pricing power.

Conversely, software that integrates systems, applies enterprise controls, manages exceptions, and proves outcomes may become more important as customers deploy AI. In its 2025 Form 10-K, ServiceNow argues that turning AI-generated information into operational outcomes requires infrastructure for orchestration, systems integration, governance, enterprise reliability, and an integrated data layer. ^[3] That is a useful diligence hypothesis, but a target still needs evidence that customers depend on those capabilities in practice.

Build a workflow map for the target’s most important use cases. Mark where data originates, where decisions occur, which system executes each action, how exceptions are handled, and where outcomes are recorded. Then assess what would happen if an AI layer became the customer’s primary interface.

A product architecture diagram will not answer this question by itself. The diligence team needs technical evidence, customer evidence, and evidence of actual product use.

Test 2: Which alternatives can deliver the same customer outcome?

The relevant competitive set may extend well beyond the companies listed in the target’s sales materials.

Adobe’s fiscal 2025 Form 10-K, for example, identifies general productivity platforms, applications with built-in editing capabilities, AI-first tools, specialized point solutions, and customers’ internally developed applications among the competitive alternatives facing different parts of its business. ^[2] The underwriting implication is that pressure can come from products that look different but deliver enough of the same customer outcome.

For each major workflow, assess at least five possible substitutes:

• A horizontal AI agent working across multiple applications
• An adjacent suite or platform extending into the workflow
• An AI-native specialist focused on the highest-value task
• A services provider combining people, software, and AI
• A customer-built application assembled from models, internal data, and existing systems

Do not ask only whether these alternatives can reproduce every feature. A substitute can change buying behavior without providing a perfect functional replica. It may remove the need for some users, absorb the most valuable part of the workflow, or make the incumbent product less visible.

Win-loss analysis should therefore include partial substitution and consolidation. Look for reductions in purchased seats, narrower module adoption, delayed expansions, demands to bundle AI at no charge, and customer plans to build an internal interface over existing systems. These signals can appear before outright logo churn.

Test 3: Which switching costs are durable?

Software underwriting often treats switching costs as a single variable. AI requires a more precise decomposition.

Potentially durable switching costs include:

• Proprietary or deeply structured customer data
• Complex integrations with operational systems
• Embedded approvals, permissions, controls, and audit trails
• Regulatory or contractual dependencies
• High-risk migration and validation requirements
• Cross-functional processes involving many stakeholders
• Historical outcome data that improves decisions or execution

More vulnerable forms of stickiness include:

• Interface familiarity
• Training on routine navigation
• Manual data-entry habits
• Reports that an agent can reproduce elsewhere
• Historical reluctance to undertake a replacement project
• A large seat footprint unsupported by equivalent workflow dependence

ServiceNow’s filing illustrates why enterprise adoption and migration friction can be real. It reports that larger, more complex deals can involve extended product evaluation and testing, multiple levels of stakeholder approval, configuration, integration services, implementation support, and longer sales cycles, particularly when customers are switching from legacy on-premises systems. ^[3] But investors should not infer that every element of an incumbent relationship is equally protected.

An AI interface may bypass part of an application while leaving the underlying system in place. That can weaken engagement and seat demand without triggering a full migration. A target can retain the customer logo while losing economic control.

Customer interviews should test what buyers would preserve if they redesigned the workflow today. Ask which data must remain in the platform, which integrations are difficult to recreate, which controls are non-negotiable, and which users could stop entering the application if a reliable agent acted for them. Compare those answers with telemetry and renewal behavior rather than relying on stated satisfaction.

Test 4: How could AI change pricing, margins, and development economics?

Historical seat expansion is no longer sufficient as the default pricing case for an AI-exposed software company.

Current pricing structures demonstrate that AI monetization can be tied to actions, conversations, usage, or completed outcomes rather than only to human access. Salesforce offers Agentforce consumption pricing based on actions and conversations, alongside per-user options. ^[7] Intercom combines per-seat help-desk plans with usage-based Fin AI Agent pricing tied to successful outcomes. For customers using Fin with an existing help desk, Intercom states that it charges per outcome and does not impose teammate seat costs. ^[8]

This does not mean every category will adopt outcome pricing. It means the underwriting model should test more than one monetization unit.

At minimum, run scenarios for:

• Lower seat counts as AI handles work previously performed by users
• Higher usage but lower revenue per human user
• AI features bundled into existing subscriptions
• Separate charges for actions, conversations, usage, or completed outcomes
• Greater revenue volatility under consumption pricing
• Stronger expansion if the product completes more valuable work

AI can also change the customer’s labor economics. A study of 5,179 customer-support agents found that access to a generative AI assistant increased issues resolved per hour by 14% on average. The measured improvement was 34% for novice and lower-skilled workers, with minimal impact on experienced and highly skilled workers. ^[6] The implication is not a universal productivity assumption. It is that effects can vary materially by user group, changing required seat counts, staffing models, product value, and purchasing criteria.

Costs require the same scrutiny. Adobe states that developing, testing, deploying, and scaling its own and third-party AI systems can increase solution costs and pressure margins without assurance of customer adoption. Its filing also identifies the possibility that AI could decrease demand for its solutions. ^[2] Investors should request a unit-cost view covering inference, model providers, retrieval infrastructure, evaluation, observability, human review, security, support, and maintenance.

Do not assume that coding assistance creates a fixed development-capacity gain. A METR randomized trial conducted from February through June 2025 found that experienced open-source developers working on mature repositories took 19% longer to complete assigned tasks when AI tools were allowed. ^[9] METR’s later experiment produced only weak evidence about speedups because participant-selection effects and time-measurement problems made the estimates difficult to interpret. ^[5] Anthropic’s analysis of roughly 400,000 Claude Code sessions found that usage shifted toward more end-to-end agentic work during the observed period, while greater domain expertise remained associated with higher session success. ^[4]

A prudent model separates possible engineering capacity gains from the cost of reviewing, testing, governing, operating, and supporting AI-enabled software. Both can be true at once.

Test 5: What ownership response does the evidence require?

A credible disruption assessment ends with decisions, owners, funding, and timing.

The appropriate response depends on the evidence:

• Accelerate product investment when the company has trusted data, workflow access, and customer permission but lacks execution capacity.
• Strengthen integrations and data control when defensibility depends on connecting operational systems and outcomes.
• Change packaging and pricing when customer value is moving away from seats toward volume, actions, or completed work.
• Partner when a third-party capability can close the gap faster than internal development without surrendering strategic control.
• Acquire a capability when speed matters and the target has a credible integration path.
• Narrow the market position when the company can defend a specific regulated, complex, or operationally demanding workflow better than a broad horizontal position.
• Harvest cash flow or alter exit timing when the required investment exceeds the likely return or the window for repositioning is too short.

Assign each action to the deal thesis, 100-day plan, annual operating plan, or exit plan. Initiatives material to the investment case need estimated cost, technical dependencies, accountable leadership, milestones, and evidence thresholds.

EES’s Private Capital Technology Decision Playbook provides a broader structure for translating technical findings into build, buy, fix, integrate, and defer decisions. For post-close prioritization, apply the same discipline to a small number of AI initiatives with measurable workflow value, as described in Private Equity AI Value Creation.

What should change in the investment model?

The final diligence output should connect evidence to model variables rather than assign the asset a generic AI risk rating.

Consider explicit adjustments to:

• Gross and net retention by customer segment
• Seat growth and module expansion
• Discounting and contract duration
• Revenue mix across subscriptions, usage, and outcomes
• Gross margin after AI delivery and support costs
• Product and engineering investment
• Sales efficiency if the competitive set broadens
• Capital required for acquisitions or integration work
• Exit multiple and the evidence a future buyer will demand
• Hold period if repositioning requires more time

The upside case also needs proof. Require evidence that the company can control or complete more of the customer workflow, not merely attach an AI feature to its current product. Relevant proof may include customer adoption, measurable workflow completion, access to differentiated data, repeatable delivery economics, stronger retention, and willingness to pay under the new model.

AI disruption underwriting is a question of economic control. Determine who owns the record, who mediates the user, who orchestrates the systems, who executes the work, and who can prove the outcome. Then price the asset and design the ownership plan accordingly.