For most of the software era, the transaction was simple: companies bought tools, then paid people to know what to do with them. AI is starting to collapse those two purchases into one.
A new generation of products does not merely store information or speed up a workflow. It interprets inputs, proposes decisions, produces finished work, and increasingly takes action. The customer may still see a software subscription on the invoice, but part of what they are buying looks remarkably similar to the judgment once supplied by an analyst, developer, marketer, researcher, or consultant.
That shift changes more than productivity. It alters software pricing, product design, professional services, hiring decisions, and even the question buyers need to ask before adopting new technology.
The Transaction Has Changed
Traditional business software was valuable because it gave skilled people better machinery. An accountant used financial software. A designer used creative software. A marketer operated an advertising platform. A developer worked inside an integrated development environment.
The software rarely pretended to know the objective better than its operator. AI products are being designed around a different proposition. Instead of helping a user execute every step, they increasingly ask for the desired result and work backward from it.
A sales platform can analyze calls, extract objections, score opportunities, draft follow-ups, and recommend which account deserves attention. A development assistant can inspect a repository, identify the files affected by a requested change, generate code, run tests, and revise its work. A support system can read a customer request, retrieve relevant account information, formulate an answer, and resolve certain cases without an agent writing the response.
The product has moved closer to the decision itself. This explains why raw AI adoption numbers only tell part of the story. Stanford’s 2026 AI Index reports that organizational AI adoption reached 88%, while generative AI reached roughly 53% population-level adoption within three years. The important business question is no longer whether organizations have access to AI. It is how much of the workflow they are prepared to hand over to it.
Software used to sit underneath expertise. Increasingly, software is attempting to contain a piece of it.
Expertise Has Several Layers
Calling AI an “expert” is usually too broad to be useful. Professional expertise is not one thing that can either be automated or protected from automation. It is a stack of different activities with different technical requirements.
At the bottom are mechanical tasks: retrieving information, formatting data, moving material between systems, checking standard conditions, and preparing routine outputs. Conventional automation has handled many of these jobs for years.
Above that sits pattern recognition. A specialist learns that certain combinations of signals tend to produce certain outcomes. A security analyst recognizes suspicious behavior. A salesperson notices buying signals. A performance marketer learns that an apparently successful campaign is attracting the wrong type of lead.
AI has become increasingly capable at this middle layer because modern models can process large quantities of structured and unstructured information and compare new situations with patterns represented in their training or supplied context.
Above pattern recognition sits judgment. This is where the system must weigh incomplete information, conflicting objectives, unusual circumstances, risk, timing, and consequences.
Finally comes responsibility. Someone has to decide whether the result should actually be trusted and acted upon.
A useful way to understand the current shift is therefore not “AI replacing expertise.” It is AI productizing selected layers of expertise.
| Layer of work | Traditional ownership | AI’s growing role |
| Information handling | Administrative staff and software users | Retrieval, extraction and organization |
| Pattern recognition | Analysts and specialists | Classification, prediction and anomaly detection |
| Recommendation | Experienced professionals | Suggested actions based on available context |
| Execution | Employees or service providers | Automated multi-step actions and agent workflows |
| Accountability | Managers and professionals | Still primarily human and organizational |
This is why two AI products using similarly capable models can deliver very different business value. The model is only one component. The real product is the workflow built around it: what information it receives, what tools it can access, which decisions it is allowed to make, and when a person must intervene.
The Product Is Moving Toward Outcomes
The old SaaS model sold access. AI software increasingly sells completed work. That distinction creates a surprisingly large business-model problem.
Seat-based pricing works when employees are the primary units of production. If a company has 50 employees using a CRM, charging for 50 seats has an obvious logic. But suppose an AI system allows ten people to perform work that previously required 20. The software has become more economically valuable while simultaneously reducing the number of seats required.
AI vendors are responding by moving toward consumption, credits, tasks, compute, and other usage-linked pricing models. The shift makes sense because usage starts to resemble the actual unit of value: a document processed, a minute generated, a workflow completed, a customer interaction handled, or an analysis performed.
The more interesting stage comes when vendors start tying price to results. If a product simply provides analytics, buyers evaluate its features. If it claims to reduce fraud, resolve support requests, qualify leads, produce software, or improve campaign performance, buyers start judging it more like a service provider.
The questions change accordingly. Customers want to know not only what the product can do, but how often it succeeds, how failures are handled, whether unusual cases receive human review, and how much responsibility the vendor accepts for the final result.
Software companies then inherit a problem professional-services businesses have always faced: a customer paying for an outcome is much less tolerant of technically correct activity that produces a commercially poor result.
The Interface Is Becoming a Brief
One of the clearest signs of this shift can be seen in interface design. Conventional software forces the user to translate an objective into operations. Someone using analytics software selects dimensions, filters, metrics, date ranges, and visualization settings. A person configuring an advertising campaign chooses audiences, bids, budgets, creative, keywords, and conversion events. Developers specify commands and manipulate files individually.
The expertise required to operate the software sits outside the product. Generative interfaces reverse some of that relationship. Users increasingly describe the desired state rather than every intermediate action. They ask a system to investigate a sales decline, prepare a financial model, refactor a feature, examine customer feedback, or build a campaign for a particular audience.
The user’s instruction starts looking less like a software command and more like a professional brief.
That puts significant pressure on the layers underneath the conversational interface. A useful agent must determine what information it needs, retrieve that information, decide which tools to use, sequence actions correctly, evaluate intermediate results, and recognize when it cannot safely continue.
McKinsey’s 2025 global AI survey found that 88% of respondents reported regular AI use in at least one business function, while organizations were also moving into experimentation with agentic systems. Yet only 39% reported enterprise-level EBIT impact attributable to AI.
The gap matters. Giving employees an AI interface is comparatively easy. Designing a system that reliably owns part of an operational process is much harder.
Feedback Determines What Can Be Automated
The best environments for autonomous software tend to have one thing in common: the system can observe whether its decisions worked.
Digital advertising is a good technical example. Platforms receive rapid feedback from impressions, clicks, conversion events, audience behavior, bids, costs, locations, devices, and campaign history. An algorithm can change an action, observe the result, and update future decisions.
Recommendation systems have similar advantages. A platform recommends something and immediately receives signals about whether the user ignored it, clicked it, watched it, purchased it, or returned later. Fraud models can compare predictions with confirmed fraudulent and legitimate transactions. Logistics systems can compare proposed routes against delivery time, fuel consumption, and cost.
These are rich feedback environments. Professional work becomes more difficult to encode when the feedback is delayed, subjective, incomplete, or contaminated by other variables. A branding decision may take months to evaluate. A hiring decision can appear successful initially and look very different two years later. A strategic recommendation can produce several consequences that cannot be traced cleanly to a single action.
The technical limit is therefore often not whether AI can generate a convincing recommendation. The harder question is whether the system has a reliable mechanism for learning what a good recommendation actually looks like.
Context Sits Outside the Platform
Automation is strongest when the objective is already clear. It becomes less reliable when defining the objective requires knowledge the system cannot infer from platform data alone. A bidding algorithm may become exceptionally efficient at generating conversions while knowing very little about which conversions create profitable customers, which inquiries consume excessive staff time, or which outcomes carry greater strategic value.
Specialized markets make the limitation easier to see. Legal advertising combines expensive search demand, geographic competition, practice-specific economics, differing case values, landing-page performance, and lead qualification. Law Firm PPC Services therefore offer a useful example of the boundary between platform intelligence and domain judgment: increasingly sophisticated ad systems can decide how to pursue a defined conversion goal, while specialist knowledge is still needed to determine whether the goal accurately represents valuable business.
The same distinction appears in healthcare, cybersecurity, financial services, procurement, and other high-context environments. Better optimization does not correct a poorly chosen objective. In some situations, it simply reaches the wrong objective faster.
AI Compresses Expert Time
The most commercially important effect of AI may not be complete replacement. It may be expertise compression, where less specialist time is required to produce each unit of work.
Consider a research workflow. An analyst might previously have spent hours finding documents, extracting relevant material, comparing sources, arranging evidence, and producing a first draft before beginning higher-value interpretation. If AI handles much of the collection and preparation layer, the analyst’s expertise is applied to a narrower but more consequential part of the process.
The same pattern is appearing elsewhere. Developers can spend less time producing routine code and more time reviewing architecture and edge cases. Customer-support specialists can supervise escalations rather than manually answering every standard question. Designers can generate many initial directions before concentrating effort on selection and refinement.
OpenAI’s 2025 enterprise survey reported that users estimated saving roughly 40 to 60 minutes per day with AI, while 75% said AI improved either the speed or quality of their output. Because these figures come from surveyed enterprise users rather than controlled productivity experiments, they should not be treated as a universal productivity rate. They do, however, illustrate the direction of the workflow change.
The economic effect can be significant even when the professional remains essential. If an expert can supervise twice as many processes, examine more alternatives, or spend more time on the cases where judgment actually changes the outcome, the structure of professional work changes without the profession disappearing.
This is also why simple job-replacement forecasts often miss what happens inside organizations. Tasks disappear and recombine before entire occupations do.
Not All Expertise Productizes Equally
AI vendors naturally have an incentive to push their products toward more valuable work, but some forms of expertise are far easier to turn into software than others.
Several characteristics make a task particularly suitable:
- The work produces many comparable examples. Repetition creates the historical evidence needed to identify patterns and evaluate whether the system behaves consistently.
- The objective can be translated into measurable signals. Conversion rate, defect frequency, delivery time, response rate, and fraud loss are easier optimization targets than reputation, employee morale, or strategic positioning.
- The process has clear intermediate stages. AI can be checked after research, classification, generation, or validation rather than being trusted with one large opaque task.
- Mistakes can be discovered before they become expensive. Automation is easier to justify when a human can review the result or the action can be reversed.
- Relevant context can be supplied to the system. Access to company data, policies, transaction history, customer records, product information, and workflow state often matters more than adding another layer of general model capability.
The opposite combination is difficult: uncommon cases, incomplete information, uncertain objectives, irreversible consequences, weak feedback, and a requirement for someone to accept responsibility.
AI will still enter those environments, but usually in a different role. Instead of replacing the final decision maker, it may retrieve evidence, compare scenarios, detect anomalies, prepare recommendations, or maintain continuous monitoring.
The distinction is important because a product that automates 70% of a workflow can be enormously valuable without being capable of owning the remaining 30%.
Software and Services Are Converging
As AI absorbs more work, software companies are starting to encounter the messy realities normally associated with professional services.
Customers need implementation help because their data is inconsistent. Workflows differ between organizations. Internal policies have exceptions. Integrations fail. Employees disagree about what the automated system should optimize. High-value cases need escalation.
Pure self-service software struggles with these problems because the problem is not always software configuration. Sometimes it requires understanding how the organization actually operates.
At the same time, service businesses are becoming more product-like. Agencies build internal automation systems. Consulting firms encode repeatable methodologies into tools. Accounting, legal, research, engineering, and creative businesses can use AI to standardize portions of delivery while reserving specialist attention for unusual or high-value work.
The two models begin meeting in the middle. A software company can automate most of a process and sell expert assistance around the edges. A services firm can automate routine delivery and make human expertise the premium layer.
Buyers may eventually care much less about whether the supplier calls itself SaaS, consulting, managed services, or an AI platform. They will evaluate a more practical combination: how much work is completed, how reliably it is completed, how deeply the system understands their environment, and what happens when automation reaches something it cannot handle.
The Scarce Skill Moves Upstream
As execution becomes cheaper, deciding what deserves execution becomes more valuable.
AI can generate ten positioning ideas quickly. It does not automatically know which market a company should enter. A coding agent can produce several implementations, but choosing the correct architecture still depends on requirements the model may not fully understand. An advertising system can optimize a campaign aggressively, but it cannot rescue a business that has defined the wrong customer as the target.
This shifts competitive advantage toward capabilities that sit before and around automation rather than inside individual tasks.
Organizations with proprietary data can provide systems with better context. Companies with strong measurement can distinguish genuine improvement from attractive-looking output. Teams with mature processes can automate stable workflows instead of automating confusion. Specialists who understand a domain can identify where an AI recommendation conflicts with realities missing from the model’s view.
The 88% organizational adoption figure reported by Stanford therefore should not be confused with 88% organizational transformation. Access to capable models is rapidly becoming common. Building a company that knows what to delegate, what to measure, what to keep under human control, and when to override an automated decision remains much less standardized.
As models improve, that organizational skill may become more valuable rather than less.
Where the Boundary Lands
Software used to sell capability while professional services sold judgment. AI is making that distinction increasingly difficult to maintain.
Modern products can interpret information, recommend actions, create finished material, and execute multi-step workflows. Service businesses can use those same capabilities to reduce the amount of human labor required for routine delivery. Software companies, meanwhile, are adding implementation, supervision, and specialized support as their products move closer to consequential business outcomes.
The useful question is no longer whether AI will turn every expert into software. Many forms of expertise depend too heavily on context, conflicting objectives, accountability, and unusual situations for that to be a realistic description.
A better question is which parts of expertise can become product features and which parts become more valuable precisely because they cannot.
That is where the economics of AI become more interesting than the technology alone. When software can supply analysis and execution at very low marginal cost, human expertise shifts toward defining the problem, setting the objective, recognizing exceptions, and accepting responsibility for the result.
Companies are still buying software. Increasingly, however, the real purchase is a carefully chosen division of labor between machine capability and human judgment.
