AI, Development and AutomationGrowth Transformation
Artificial intelligence is redefining how companies compete: what to automate, what to transform and where to start
Adopting artificial intelligence doesn’t guarantee better results. The difference lies in understanding which capabilities become more accessible, which processes need to be redesigned and where a sustainable competitive advantage can be built.
A company can have artificial intelligence licenses for its entire team, an assistant that drafts proposals, another that summarizes meetings and a chatbot that answers questions, and still close the year with the same cost structure, the same delivery times and the same margin. Each person works somewhat faster. The company, in essence, works the same way.
This is no minor paradox. Artificial intelligence is making capabilities that until recently were expensive much cheaper: analyzing documents, writing, programming, classifying, estimating, answering. But when a capability becomes cheap for one company, it also becomes cheap for its competitors. The question that matters isn’t how much artificial intelligence an organization uses, but what changed in the economics of its business as a result of using it.
The advantage doesn’t come from accessing new capabilities, but from reorganizing the company to turn them into economic value.
Key takeaways
- AI adoption spreads much faster than its financial impact: almost every organization uses it, and a small minority attributes a meaningful share of its results to it.
- From an economic standpoint, AI lowers the cost of certain capabilities. That changes the value of their complements: data, judgment, the ability to act and processes.
- Doing a task faster doesn’t make the company work better if the rest of the system doesn’t change.
- The greatest potential appears when systems of decisions are redesigned, not when isolated activities are improved.
- AI agents expand what can be automated, but require assessing reliability, oversight and cost in each specific workflow.
- If models are accessible to everyone, advantage tends to shift toward what is hard to replicate: proprietary data, workflows, integration, distribution and organizational learning.
AI adoption isn’t the same as business transformation
Stanford University’s 2026 AI Index, drawing on McKinsey’s global survey, reports that 88% of organizations surveyed used AI in at least one business function in 2025, up from 78% the year before. The same report shows that scaled deployment of AI agents remained in the single digits in almost every function.7 The technology is spreading fast; the transformation of work, much more slowly.
McKinsey’s 2026 survey, with 1,719 participants in 97 countries, shows that gap more precisely. Eighty percent said AI had improved their individual productivity. Thirty-seven percent attributed some EBIT impact at their organization to AI use. And only 6% qualified as AI high performers: organizations that attribute at least 5% of their EBIT to AI and describe its impact as significant.8 These are self-reported figures measuring different things, so they shouldn’t be compared as if they were one scale. But the pattern is clear: between a person working better and the company earning more there is a distance that technology, on its own, doesn’t cover.
To map that distance we propose distinguishing six levels, in what we call The AI Transformation Ladder by Advisent®. It isn’t a scientific scale or a mandatory sequence; it is a way to identify where each initiative stands.
Advisent® conceptual ladder, not a validated scale. Figures come from different surveys (McKinsey 2025 via the Stanford AI Index 2026, and McKinsey 2026), are self-reported and are placed only as a reference for the level they illustrate. Source: Advisent®.
When a capability gets cheaper, the economics of the business change
Ajay Agrawal, Joshua Gans and Avi Goldfarb, economists at the University of Toronto, proposed a way of thinking about artificial intelligence without mysticism: as a technology that lowers the cost of prediction.1 Prediction, in its broad sense, means using available information to estimate missing information, and it is valuable because it helps make decisions under uncertainty. When the price of something falls, it is used more, applied to new problems, and the value of its complements and substitutes changes.
The authors distinguish between prediction and judgment. Prediction estimates what is likely to happen; judgment weighs the consequences and decides what matters most. A machine can greatly improve the first without resolving the second. A practical consequence follows: when prediction gets cheaper, the relative value of its complements rises, such as the data to feed it, the judgment to interpret it and the ability to act on it. A more accurate demand forecast doesn’t reduce inventory if the company can’t change its purchasing.
Today’s generative AI does more than predict in the classic sense: it writes, programs, generates images, uses tools and takes part in complete workflows. The prediction framework doesn’t describe all those capabilities, but its economic logic remains useful. For any activity AI makes cheaper, the strategic question isn’t only which task the machine can perform, but what other capabilities the company needs to take advantage of it and who else will be able to do the same at the same cost.
Doing a task faster doesn’t mean the company works better
Evidence on task-level productivity is solid. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the introduction of a generative assistant in a company’s customer support center and found an average increase of about 15% in issues resolved per hour, with larger gains among less experienced agents.4 In an experiment with 453 college-educated professionals, Shakked Noy and Whitney Zhang observed that AI reduced the time spent on professional writing tasks by about 40% and increased their assessed quality by 18%.5 The AI Index summarizes the available evidence with a useful observation: gains are larger in structured, measurable work, and smaller in tasks that require deeper reasoning.7
None of these results amounts to an increase of the same magnitude in a company’s productivity, let alone its profitability. A hypothetical example shows why. A sales team takes two hours to prepare a proposal; with AI, it takes thirty minutes. But approval still takes three days, operations can’t absorb more clients, contracts still require manual review and clients don’t buy more because they receive proposals faster. Task productivity improved by 75%. System performance, perhaps not at all.
Time saved isn’t a cost saving either. It only becomes economic value if it translates into more capacity that is actually used, lower costs that are actually eliminated or better quality the customer values. That is why the unit of analysis can’t be only the task: it has to include the workflow, the process and the economic result. It is the same constraint problem we examine in our article on operations and scalability: improving a stage that isn’t the constraint doesn’t improve the system.
Time saved isn’t a cost saving. It is potential capacity, and it is only worth something if someone turns it into results.
Automating isn’t redesigning
In a second book, the same authors distinguish three levels of impact.2 A point solution improves an existing procedure without changing anything else. An application solution enables a new procedure, but within the same system. A system solution changes procedures that depend on one another and, with them, the organization. Their argument is that the greatest economic value tends to appear at the third level, which is also the hardest, because it requires coordinating changes that affect many areas at once.
The history of electrification illustrates the difference. Economist Paul David showed that when factories adopted electric motors, many initially used them as a replacement for the central steam engine, without changing the plant layout, and that the large productivity gains came decades later, when factories were redesigned around individual motors on each machine.6 The technology was the same; what changed was the design of the system.
Categories from Agrawal, Gans and Goldfarb (2022); hypothetical example by Advisent®.
This leads to the most useful question for any leadership team: if we designed this company from scratch, with today’s artificial intelligence capabilities, would we organize it the same way? Applied to a specific process, the question shifts the focus. Instead of automating every existing step (the client sends information, someone reviews it, another asks for clarifications, another prepares a document, another approves, another enters the result), it forces the question of which steps should exist. Information can be captured in structured form from the start, a system can detect missing data, AI can assist with classification and preparation, exceptions can be escalated to specialists and critical decisions can retain human oversight. Transformation isn’t automating the previous process, but reconsidering why it had that shape.
That redesign usually also changes decision rights: who has the information, who recommends, who decides and who executes. McKinsey’s survey shows an association consistent with this idea: nearly three quarters of the top-performing organizations reported fundamentally redesigning their workflows, compared with about a quarter of the rest.8 It doesn’t prove that redesign causes better results, but it does show that those who get more value rarely limit themselves to adding tools.
What to automate, what to augment and what to leave in human hands
Not every activity should be automated, and not every automation requires artificial intelligence. Moving a structured piece of data between two systems is better handled with an API integration or a deterministic rule than with an agent; interpreting heterogeneous documents, by contrast, can justify an AI model. Using generative AI where a rule solves the problem adds cost, latency and error risk without adding value. Technology should be chosen according to the problem and its economics, not the other way around.
To decide what to do with each activity we propose The AI Decision Model by Advisent®, with five options that can be combined within a single process: eliminate the activity if it shouldn’t exist; keep it human when the cost of error, ambiguity, accountability or the customer relationship justifies it; augment it, when AI assists a person who keeps the decision; automate it under rules and controls; or redesign the system it belongs to. The choice depends on observable variables: how structured the task is, how much variability it has, how much an error costs, what data exists, what autonomy it requires and what economic value is at stake. Keeping an activity human doesn’t mean human decisions are always better; it means that, given current evidence and risk, it is the right option.
| Activity | Problem | Possible technology | Decision | Metric |
|---|---|---|---|---|
| Document classification | High volume, heterogeneous formats | AI for extraction and classification | Automate with controls | Error rate and escalated cases |
| Data transfer between systems | Duplicate manual entry | API integration or rules | Automate without AI | Entry time and errors |
| Sales analysis | Slow information preparation | Generative AI with CRM data | Augment | Time to proposal and win rate |
| Customer service | Repeat queries and complex cases | AI with human escalation | Automate the simple, augment the complex | First contact resolution |
| Contract approval | Legal risk and accountability | AI for preliminary review | Keep human, assisted | Approval time and errors |
| Demand forecasting | Excess or shortage of inventory | Predictive machine learning | Redesign purchasing and replenishment | Turnover and stockouts |
| Software development | Limited engineering capacity | Coding assistants and agents | Augment with review | Quality, defects and lead time |
AI agents expand the possibilities, but also the demands for control
It helps to distinguish terms that are used as synonyms. An assistant helps carry out tasks. A copilot is an assistant built into a work environment. An AI agent is a system that can use tools and take actions to meet a goal within defined limits. And an agentic workflow incorporates components able to choose or execute actions with varying degrees of autonomy. Not all agents are autonomous, and autonomy isn’t a goal in itself.
The relevant question isn’t whether an agent can solve a task, but whether it can do so with the reliability the process requires, with acceptable human oversight and at a tolerable cost. A Scale Labs paper published in September 2026 shows this clearly: in a clinical audit case, two systems with almost identical autonomous accuracy (72.8% and 72.5%) needed human review on very different shares of cases (39.2% versus 29.6%) to reach the same reliability level.12 A benchmark score doesn’t tell you how much it costs to run the system safely.
Oversight design matters. In its study of 51 successful deployments, Stanford’s Digital Economy Lab found that models in which AI does most of the work and people handle exceptions showed larger median productivity gains than models in which a person approves every output, although the authors caution that task types may differ.9 Governance, in this context, isn’t a written policy. It is an architecture: what permissions each agent has, what actions it can take, what data it can see, what gets logged, how errors are handled, who can intervene and who is accountable. NIST’s AI risk management framework, organized around the functions govern, map, measure and manage, and its specific profile for generative AI, offer a basis for structuring those decisions across the system’s entire life cycle.13
AI is also changing which software is worth building and which is worth buying
Programming is one of the areas where AI is advancing fastest, and that changes a decision that seemed stable: buying software or building it. In McKinsey’s 2026 survey, 32% of participants said their organization had decided not to buy at least one software product or feature because it could build it in-house with agentic coding tools.8 It is a relevant signal for any company that buys or sells software.
But the evidence on development productivity is heterogeneous. In a randomized METR experiment with 16 experienced developers working on their own repositories, early-2025 AI tools made tasks take about 19% longer, even though the developers believed AI was speeding them up.10 In February 2026, METR reported that its later data suggested likely gains with newer tools, but that selection problems prevented estimating their size with confidence.11 The conclusion isn’t that AI always slows down or speeds up development, but that the result depends on the type of task, experience, context and tool maturity.
Generating code isn’t the same as having production-ready software, either. A proprietary system requires architecture, security, testing, integrations, documentation, maintenance and technical debt management, and those costs don’t disappear because writing the first version is cheaper. The build-or-buy decision still depends on the functionality’s strategic importance, on whether it differentiates the company, on total cost over time, on available talent and on vendor dependence. AI makes building cheaper; it doesn’t make ownership cheaper in the same way.
Services and products: the same technology, different economics
In service companies (consultancies, agencies, law firms, custom software firms), AI changes the economics of execution and, at the same time, the client’s alternatives, since the client can now research, write, analyze or program on its own. That doesn’t mean it stops hiring services, but it may be less willing to pay for tasks it can already do itself. Advisent®’s hypothesis is that, when basic execution becomes more accessible, the relative importance of diagnosis, judgment, specialization, implementation and accountability for results increases. We develop this in our article on customer relationship profitability.
A hypothetical case makes it concrete. A 20-person consultancy uses AI to research, prepare drafts and generate proposal alternatives, with human validation at the end. Preparation time drops noticeably. But that doesn’t by itself produce more clients, more revenue or better margins. To achieve that, the firm needs to review how it qualifies opportunities, how it sets price and scope, who approves, how it plans team capacity and how it kicks off delivery. Automating the proposal is a point improvement; redesigning how it sells and delivers is a transformation.
In product companies, it is worth separating effects that are often confused: AI that reduces costs, AI that improves decisions, AI that increases capacity, AI that enables new products and AI that transforms the business model. A distributor with excess inventory in certain categories can improve its demand forecast, but a more accurate prediction doesn’t reduce costs on its own. It needs to be able to change purchasing, replenishment, distribution and commercial decisions based on that prediction. The sequence that creates value is prediction, action and operating model. Without the last two, the first is just a more accurate report.
If everyone can access AI, where does competitive advantage lie?
Marco Iansiti and Karim Lakhani, of Harvard, described how organizations built around data, software and algorithms develop a different operating architecture, which they called the AI factory: a cycle in which interactions generate data, data feeds algorithms, algorithms produce predictions and decisions, and those decisions generate new interactions.3 That architecture can reduce some traditional constraints on scale, broaden scope into new activities and enable cumulative learning. But only when the data is relevant, there are rights to use it, it is of sufficient quality, results can be evaluated and feedback is actually integrated into the system. Models don’t learn on their own from every interaction.
Recent evidence reinforces the idea that the model, in isolation, offers little differentiation. In Stanford’s study of successful deployments, the foundation model was fully interchangeable in 42% of cases, and the authors conclude that lasting advantage lies in the orchestration layer: how data, processes, systems and people are connected.9 The same study found that only 6% of cases had fully AI-ready data and that 77% of the hardest challenges were less visible costs, such as change management, data quality and process redesign.
Advisent® conceptual representation. No layer is a sustainable barrier on its own; advantage depends on the combination and on how hard it is to replicate in each market. Source: Advisent®.
Advisent®’s thesis is that the advantage isn’t necessarily in the model, but in the system the company builds around it. But not every system creates a barrier. Three questions are useful: which elements of that combination are hard to replicate, which create value for the customer and which can sustain that value when competitors access the same tools. A redesigned workflow a competitor can copy in three months is an improvement, not an advantage.
Nor is it clear that AI automatically favors small or large companies. Small companies can access capabilities that used to require entire teams and move faster. Large ones keep advantages in capital, data, distribution and investment capacity: according to McKinsey, 40% of organizations with revenue above USD 1 billion were scaling AI agents in 2026, compared with a flat 22% among smaller ones.8 The outcome depends on which capabilities get cheaper in each industry and on who is best placed to reorganize around them.
AI Value Architecture: how to decide where to invest
To avoid technology implementations without business logic, we propose The AI Value Architecture by Advisent®, a conceptual tool (not a validated academic model) that organizes any initiative into eight questions, from strategy to value capture.
Advisent® conceptual framework, not a validated model. Source: Advisent®.
The third and eighth questions are the ones most often skipped. Without a baseline (what the process costs, how long it takes and what errors it produces today) there is no way to know whether the initiative created value. And value isn’t measured in hours saved, but in net incremental value: additional economic benefits minus all additional costs. These include licenses, token consumption, infrastructure, implementation, integrations, data preparation, security, oversight, verification, rework, maintenance, training and change management. About 20% of McKinsey’s survey participants said AI operating costs, including tokens, had constrained their use of AI.8
An illustrative example shows how much the result changes depending on what is counted. A company invests USD 12,000 to implement an AI automation, with additional operating costs of USD 6,000 a year. If it obtains verified incremental benefits of USD 30,000 in the first year, the simple net benefit is USD 12,000. But if those USD 30,000 were the theoretical value of hours saved and only USD 10,000 actually turned into lower costs or more sales, the first-year result is a loss of USD 8,000. The calculation is deliberately simplified (it excludes taxes, the time value of money and later benefits) and isn’t a benchmark. Its purpose is to show that the difference between success and failure often lies in converting time saved into results.
Where to start: a roadmap for leadership
The sequence we propose has six stages, although not every company needs to follow them rigidly. Diagnose, to identify where the constraints and economic opportunities are, not where there is most technological enthusiasm. Map the processes, tasks and decisions that make them up. Prioritize by value, feasibility and risk. Design, choosing the right technology and redesigning the process when appropriate. Validate, measuring results against the baseline. And scale, integrating and governing, only when there is enough evidence.
That roadmap has to coexist with legitimate counterarguments. Not every company needs to become an organization designed around AI. Many tasks are better solved with conventional software. Redesign can be costly and isn’t always justified. AI can increase rework if its outputs require intensive verification, agents can fail and operating costs can erode the expected benefit. Early adoption doesn’t guarantee advantage, because technology also commoditizes capabilities that used to differentiate. And Stanford’s experience suggests failure is part of the path: 61% of the successful projects they studied had at least one previous failed attempt.9 None of these points invalidates the thesis; they define the conditions under which it holds.
Adding tools or changing the economics of the business
Artificial intelligence isn’t simply a new productivity tool. It can change the cost of a company’s core capabilities, and that can change decisions, processes and operating models. But access to the technology is increasingly even among competitors, which is why it doesn’t guarantee capturing value. What separates organizations that get results from those that accumulate licenses isn’t how much AI they use, but how willing they are to reorganize around what AI makes possible.
That reorganization has costs, risks and limits, and it isn’t always worthwhile. But deciding not to do it is also a decision, and its consequences accumulate quietly while others redesign.
Which part of our company should work differently now that artificial intelligence makes possible capabilities that used to be expensive or out of reach?
Anyone can use artificial intelligence. Reorganizing the company to compete with it is another matter.
Frequently asked questions about artificial intelligence in business
Which processes should be automated with AI?
A good place to start is frequent, high-volume activities with available data, limited cost of error and identifiable economic value, such as document classification, information extraction or draft preparation. Before using AI, check whether a rule or a system integration solves the problem better, and whether the activity sits at a point in the process where saving time actually increases the system’s capacity.
What is the difference between automating and transforming a company with AI?
Automating means executing an existing activity with technology, usually faster or cheaper. Transforming means changing processes, decisions, roles, products or the operating model based on what the technology makes possible. Automating proposal drafting is a point improvement; redesigning how opportunities are qualified, terms are approved and delivery begins is a transformation.
How do you start implementing artificial intelligence in a company?
A reasonable sequence is to diagnose where the constraints and economic opportunities are, map the processes and decisions involved, prioritize by value, feasibility and risk, design the solution choosing the right technology, validate results against a baseline and scale only with evidence. Not every company needs to follow it rigidly, but skipping the initial measurement makes it impossible to know whether the initiative created value.
How do you calculate the return on an AI implementation?
The starting point is net incremental value: verified additional economic benefits minus all additional costs, including licenses, token consumption, integrations, data, oversight, verification, rework, maintenance and change management. Hours saved aren’t savings: they only count if they turn into lower costs, more sales or better quality. A full ROI calculation also requires defining the investment, the time horizon and the accounting criteria.
What is the difference between a chatbot and an AI agent?
A chatbot or assistant answers and helps carry out tasks, usually within a conversation. An AI agent can use tools and take actions in other systems to meet a goal within defined limits. That is why agents require more controls: permissions, action logging, error handling, reliability evaluation in each workflow and the ability for humans to intervene.
When should a company build its own software with AI instead of buying it?
AI makes writing code cheaper, but it doesn’t remove the costs of architecture, security, testing, integration, maintenance and technical debt. Building usually makes more sense when the functionality is strategic or differentiates the company, and buying when it is standard, when the vendor takes on maintenance and security, or when internal talent is limited. The decision should be based on total cost over time.
Does AI replace workers or increase their productivity?
The strongest evidence shows productivity gains in specific tasks, such as 15% in customer support with AI assistance or roughly 40% time reductions in professional writing, with effects that vary by experience and type of task. In software development, results are mixed. The impact on employment depends on how each company redesigns work, not just on the technology.
How can a company build competitive advantage with AI?
If models are accessible to all competitors, advantage tends to depend on what is hard to replicate: relevant proprietary data with usage rights, redesigned workflows, system integration, distribution and customer relationships, and the organizational ability to measure, learn and scale what works. None of these elements guarantees a sustainable advantage on its own; the combination is what matters.
Sources and further reading
Academic foundations
- Agrawal, A., Gans, J., & Goldfarb, A. (2022). Prediction machines: The simple economics of artificial intelligence (Updated and expanded ed.). Harvard Business Review Press.
- Agrawal, A., Gans, J., & Goldfarb, A. (2022). Power and prediction: The disruptive economics of artificial intelligence. Harvard Business Review Press.
- Iansiti, M., & Lakhani, K. R. (2020). Competing in the age of AI: Strategy and leadership when algorithms and networks run the world. Harvard Business Review Press.
- Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942. doi.org/10.1093/qje/qjae044
- Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. doi.org/10.1126/science.adh2586
- David, P. A. (1990). The dynamo and the computer: An historical perspective on the modern productivity paradox. American Economic Review, 80(2), 355–361.
Business and institutional research
- Stanford Institute for Human-Centered Artificial Intelligence. (2026). The 2026 AI Index Report, Chapter 4: Economy. hai.stanford.edu
- Tinkoff, D., Van der Veken, L., Chui, M., & Balakrishnan, T. (2026, August 25). The state of AI in 2026: On the road to ROI. McKinsey & Company. mckinsey.com
- Pereira, E., Graylin, A. W., & Brynjolfsson, E. (2026). The enterprise AI playbook: Lessons from 51 successful deployments. Stanford Digital Economy Lab. digitaleconomy.stanford.edu
- Becker, J., Rush, N., Barnes, E., & Rein, D. (2025). Measuring the impact of early-2025 AI on experienced open-source developer productivity. METR. metr.org
- METR. (2026, February 24). We are changing our developer productivity experiment design. metr.org
- Chatrath, V., Zhu, B., Fan, J., et al. (2026). READY or not: Reliable enterprise agent deployment. Scale Labs. labs.scale.com
- National Institute of Standards and Technology. (2024). Artificial intelligence risk management framework: Generative artificial intelligence profile (NIST AI 600-1). doi.org/10.6028/NIST.AI.600-1
The AI Transformation Ladder by Advisent®, The AI Decision Model by Advisent® and The AI Value Architecture by Advisent® are proprietary conceptual developments. Integrating the three books into a sequence of economics, decisions, system design and competitive advantage is an Advisent® editorial interpretation. The concepts of AI as a reduction in the cost of prediction, the distinction between prediction and judgment, and point, application and system solutions (Agrawal, Gans and Goldfarb), and the AI factory (Iansiti and Lakhani) belong to their authors. The surveys cited reflect stated perceptions and correlations, not causal relationships; the Stanford study is based on cases selected for their success and doesn’t allow estimating overall success rates.
Last updated:
