The New Competitive Landscape

The principles of business strategy have not changed. Differentiation still matters, cost position still matters, and network effects still compound. What has changed is how quickly a defensible position stops being defensible.

When AI writes production code, generates design variants, analyzes markets, and handles customer interaction, several of the things that used to take a company years to build take months. Small teams achieve output that recently required departments. Competitors emerge from adjacent industries because the capability barrier that kept them out has fallen. Planning horizons compress accordingly.

This chapter is a framework rather than a manual—industry, scale, and position matter too much for anything more specific to be honest. The starting admission is that nobody has demonstrated what works, because the period has been too short and the survivorship bias in current advice is severe.

What follows are the principles that appear robust across most plausible futures, and the reasoning behind them.

The Moat Problem

Warren Buffett's framing of competitive advantage as a moat around an economic castle has been the dominant strategic metaphor for decades.¹ It is worth reexamining which moats AI actually erodes, because the popular answer—all of them—is wrong and leads to bad decisions.

What AI Erodes

Scale advantages in knowledge work. The economics that justified large professional organizations rested on the leverage of many junior people supervised by a few senior ones. When the junior work is automated, that leverage inverts, and a five-person firm with good tools competes for work that previously required fifty.

Software differentiation at the feature level. Features that took a quarter to build take a week, which means feature parity arrives faster and the window during which any capability is distinctive has narrowed.

Some switching costs. Data migration, integration work, and retraining were real frictions. Each has become cheaper.

Trade secrecy in observable products. Anything a competitor can inspect can be reverse-engineered faster than before.

What AI Does Not Erode

This is the more useful list and receives less attention.

Regulatory position. A banking license, an FDA clearance, or an operating certificate is not reproducible by a capable model. If anything, compliance burden favors incumbents, since a large firm absorbs it more easily than a startup.

Physical assets and distribution. Manufacturing capacity, logistics networks, retail footprint, and installed equipment are unaffected by cognitive automation. A company whose advantage is a port terminal still has it.

Proprietary data that is genuinely proprietary. Not "we have a lot of data"—most companies do, and most of it is worthless—but data generated by operations nobody else runs, capturing outcomes nobody else observes.

Trust and switching costs in high-consequence contexts. Where being wrong is expensive, buyers pay for accountability. A cheaper provider that cannot be held responsible is not a substitute.

Genuine network effects. A marketplace's value to each participant still comes from the other participants, and AI does not conjure them.

The strategic error to avoid is assuming your moat is in the eroding category when it is not, or the reverse. A professional services firm whose advantage was leverage should be alarmed. A utility whose advantage is a regulated franchise should not be, and should spend its attention on operations rather than on existential repositioning.

AI-Native Operations

The distinction that matters most in practice: adding AI to an existing process captures a fraction of the available gain; redesigning the process around what AI does well captures most of it.²

Customer service illustrates it cleanly. Adding a chatbot to a call center reduces call volume somewhat. Designing the function on the assumption that most contacts resolve without a human, that the human queue exists for genuine exceptions, and that the people in that queue are more senior rather than less produces a different organization with different economics.

The reason most firms do the first thing is not ignorance. It is that the second requires changing headcount, reporting lines, and metrics simultaneously, which is a political act rather than a technical one. This is why AI adoption correlates more with organizational willingness to restructure than with technical sophistication.

Four operational principles hold up.

Experiment velocity is the compounding variable. The advantage goes to whoever learns fastest, and learning rate is a function of how many hypotheses can be tested per quarter.

Data is an asset only if it is captured deliberately. Most operational data is generated and discarded. Instrumenting processes to retain outcomes—not just transactions, but what happened afterward—is what creates the proprietary data described above.

Design for augmentation, not replacement or autonomy. Both extremes fail: full automation breaks on edge cases, and unassisted humans are simply slower. The productive configuration is a human accountable for outcomes, working with tools that handle volume.

Close the feedback loop. Systems that produce output nobody evaluates do not improve. This sounds obvious and is the most commonly skipped step.

Strategic Planning Under Uncertainty

Traditional planning—three-to-five-year plans with detailed forecasts—is not merely less accurate now; it is actively misleading, because it produces commitment to a specific future rather than readiness for several.

Three adjustments help.

Scenario planning substitutes a set of coherent futures for a single forecast.⁴ The technique developed at Shell in the 1970s and its value was never prediction—Shell did not forecast the oil shock—but that having thought through the scenario meant recognizing it faster than competitors when it arrived. That is the actual benefit: reduced reaction time, not foresight.

Options thinking favors investments that create future choices over investments that commit to a path. A small position in an uncertain capability is worth more than its expected return suggests, because it purchases the ability to scale if conditions favor it.

Horizon-appropriate specificity. Six to eighteen months admits real planning. Eighteen to thirty-six months admits direction and priorities but not detail. Beyond three years, scenarios and options are the only honest instruments, and pretending otherwise wastes the planning function's credibility.

Talent Strategy

The talent market is tight and expensive—senior machine learning engineers commanded $300,000 to $500,000 and above in 2024, with retention harder than recruitment.⁹

Three observations complicate the obvious response of hiring aggressively.

The scarce skill is not the one being advertised. Model-building expertise is genuinely scarce and genuinely unnecessary for most firms, which will consume capability through APIs rather than train anything.³ What most organizations actually need is people who can identify which processes are worth changing, evaluate whether output is correct, and manage the organizational consequences. That skill set is domain expertise plus judgment, and it is more likely to be developed internally than hired.

Evaluation capability is the binding constraint. The risk in deploying these systems is not that they fail obviously; it is that they produce plausible output that is wrong in ways only a domain expert notices.⁶ An organization without people who can tell the difference cannot safely deploy at all, regardless of what it buys.

The junior pipeline is being consumed. Automating entry-level work removes the mechanism by which junior people become senior ones. Every firm doing this is drawing down a stock of expertise it is no longer replacing, and the bill arrives in about a decade. Someone has to decide deliberately how their organization produces its next generation of senior people, because the default answer is now "it doesn't."

The cultural requirements—psychological safety sufficient for experiments to fail publicly, support for continuous learning, and tolerance for cross-functional work—are the standard list and are standard because they are correct.

Technology Strategy

Build where the capability is your differentiation. Buy where it is not. The common error is building commodity infrastructure for reasons of pride and buying differentiated capability for reasons of speed—exactly inverted.

Avoid deep lock-in to any single model provider. Capability rankings among frontier providers have changed repeatedly, pricing has fallen unevenly, and an architecture that permits substitution has real option value. This argues for an abstraction layer between application logic and model APIs, which costs little to build early and considerably more to retrofit.

Treat security as a design constraint. These systems introduce genuinely new attack surfaces—prompt injection, training data poisoning, extraction of information the model should not disclose. Retrofitting security here is harder than usual because the failure modes are not intuitive.

Design for compliance you do not yet face. Documentation obligations are arriving, and building records as a byproduct of operation is far cheaper than reconstructing them.⁷

Risk Management

The AI-specific risks are model failure producing confident wrong output, novel security exposure, regulatory change, reputational damage from visible failures, and dependency on systems whose behavior can shift underneath you when a provider updates a model.

The management approaches are conventional: adversarial testing before deployment, continuous monitoring for performance drift, human oversight proportional to stakes, audit trails, and an incident response plan that exists before the incident.

Documentation deserves specific attention because it is becoming a legal requirement rather than a best practice. The EU AI Act mandates technical documentation, risk assessment, and human oversight records for high-risk systems. Sectoral requirements compound: financial institutions document model risk under supervisory guidance, healthcare AI requires clearance documentation, and automated hiring tools face audit obligations under laws including New York City's Local Law 144.

Even where regulation has not arrived, records of system design, training data provenance, performance measurement, and decision logs are what allow a firm to defend a decision after the fact. The organizations that will struggle are those that deployed systems nobody documented and cannot now explain.

Industry Considerations

Software. AI raises development productivity and compresses feature differentiation simultaneously. The strategic response is moving up the stack toward workflow and integration, where switching costs still exist.

Professional services. The most exposed sector in this chapter, because the leverage model is precisely what automates. Firms that redesign delivery around judgment and relationship survive; firms that defend billable hours do not. The pyramid is the business model, and the pyramid is what breaks.

Manufacturing. Genuinely advantaged. Physical assets remain moats, and AI improves yield, maintenance, and quality control against a base that is hard to replicate. Data from operations nobody else runs is the textbook case of a defensible data advantage.

Retail and consumer. Personalization and inventory optimization deliver real gains, and differentiation erodes as every competitor obtains the same capabilities. Advantage concentrates in direct customer relationships and in brand meaning that cannot be synthesized.

Financial services. Regulatory position remains protective, and the sector's exposure is operational: underwriting, compliance, and service functions automate heavily, and headcount follows.

Healthcare. Slow adoption from regulation and liability, which frustrates technologists and is substantially correct given the stakes. Advantage goes to organizations that build clinical evidence rather than to those that deploy fastest.

Second-Order Impacts

Capability becomes cheap and integration stays expensive. As model access commoditizes, the differentiation moves to proprietary data, workflow embedding, and the organizational ability to act on output. Firms buying capability and expecting advantage will be disappointed; the capability is available to everyone at the same price.

Speed of imitation compresses the payoff window. Investments recover over shorter horizons, which biases firms toward incremental improvements and against the long-payback investments that build durable position. This is a real strategic trap and it favors whoever can resist it.

Vendor dependency is a governance issue, not a procurement one. A firm whose core process depends on a model that a provider may deprecate, reprice, or change has accepted an operational risk that belongs on the risk register rather than in an IT budget line.

Talent concentration compounds returns. Firms that attract people who can evaluate and deploy these systems pull further ahead, because the capability is complementary to itself.

The Path Forward

Near-Term (2026–2028)

Assess honestly where AI actually affects your economics rather than where it is fashionable. Pursue the applications with clear value and contained risk. Build evaluation capability before deployment capability. Put governance in place early, because retrofitting it is expensive and the deployments happening now are the ones that will need defending.

Medium-Term (2028–2032)

Scale what worked and abandon what did not, which requires having measured. Move from AI-added to AI-native in the functions where the gain justifies the disruption. Invest in the moats that AI strengthens rather than erodes—proprietary data, integration depth, trust. Decide deliberately how your organization will produce senior expertise now that the junior pipeline has thinned.

Long-Term (2032+)

Ask whether the business model survives, and answer it honestly rather than defensively. Maintain options across multiple industry futures. And accept that the answer for some firms is that it does not survive, in which case the useful strategic act is redeploying capital rather than defending a position.

Competitive Dynamics

Incumbents hold capital, customer relationships, regulatory standing, distribution, and operational data. They are constrained by legacy systems, cultural inertia, quarterly pressure, and an unwillingness to cannibalize existing revenue.⁸

Startups hold speed, focus, and the ability to build without legacy. They are constrained by capital, credibility, distribution, and the fact that most of them will not survive long enough for their advantages to compound.

The pattern that has held across previous technology transitions is that neither category wins categorically. Incumbents that overcome inertia usually beat startups, because the underlying advantages are real and hard to replicate. Startups win where incumbents will not cannibalize themselves, which is a predictable and therefore exploitable weakness.

The position that reliably loses is the middle: an organization with neither the incumbent's structural advantages nor the startup's speed, attempting incremental adaptation to a change that is not incremental.

Governance and Ethics

Regulatory requirements are arriving, reputational exposure is real, ethical failures are operational risks, and capable people increasingly select employers on these grounds. These are the instrumental arguments, and they are sufficient.

Structurally: board-level attention to AI strategy and risk, clear executive ownership, a review process for consequential applications with articulated red lines, and periodic audit against actual behavior rather than stated policy.⁵

Substantively: test for discriminatory outcomes rather than assuming their absence, disclose automated decision-making to the people subject to it, minimize data collection to what is used, and maintain human involvement proportional to consequence.¹⁰

The most useful governance question is not whether a use is legal but whether the organization would defend it publicly. That test catches most of what regulation eventually prohibits, several years earlier.

The Deeper Questions

What is your business actually for? If a capable competitor with good tools can do what you do faster and cheaper, the answer has to be something other than doing it. Most firms have never had to articulate this, because the barrier to entry answered it for them.

What do you owe the people who built the company? If roles are eliminated, the obligation to the people in them is a real question rather than a public relations one, and how it is answered shapes what the remaining organization believes about itself.

How do you compete against firms that will do things you will not? If manipulation, surveillance, and exploitation are available and competitors adopt them, refusing carries a cost. This is the oldest question in business ethics and AI has made it concrete again.

Conclusion

Strategy has not been repealed. Understand the environment, build distinctive capability, serve customers, adapt. Those still apply.

What has changed is tempo and the specific location of defensibility. Advantages built on knowledge-work leverage are eroding quickly. Advantages built on regulation, physical assets, genuine data, and accountable trust are not. Firms that misidentify which they hold will either panic unnecessarily or fail to act while they still can.

Three things are worth carrying out of this chapter.

The organizations that capture the gain will be those willing to redesign processes rather than layer tools onto them, and that willingness is organizational rather than technical—which is why adoption tracks management courage more closely than it tracks engineering capability.

The binding constraint on safe deployment is the ability to tell when output is wrong, and that capability is built rather than purchased.

And the junior pipeline problem is the slow-moving decision most firms are making by default. Automating entry-level work is individually rational and collectively produces an industry with no way to make senior people. Whoever solves that deliberately will have an advantage in about a decade that will look, at that point, like luck.

The businesses that survive the next decade will be those that took it seriously while it was still the future. That is now.

Endnotes — Chapter 62

1. Competitive moats: the framing was popularized by Warren Buffett. Traditional moats include brand, scale, network effects, switching costs, intellectual property, and regulatory position—which AI affects very unevenly.
2. AI-native operations means designing processes around AI capabilities rather than adding AI to existing processes; the distinction determines what fraction of available productivity gain is actually captured.
3. Foundation models (GPT, Claude, Gemini, and open-weight alternatives) provide base capability that most enterprises consume through APIs rather than train, making the build-buy-partner decision primarily one about application layer and data.
4. Scenario planning was developed at Royal Dutch Shell in the 1970s. Its value has consistently been reduced reaction time when a scenario materializes, rather than accurate prediction of which one will.
5. AI governance is an emerging corporate function encompassing board oversight, ethics review, audit, and compliance; available frameworks include the NIST AI Risk Management Framework and the requirements of the EU AI Act.
6. Model failure modes include hallucination, bias, brittleness under distribution shift, and adversarial vulnerability. The operationally dangerous case is confident output that is wrong in ways only a domain expert detects.
7. Compliance by design—building regulatory requirements into systems from the outset rather than retrofitting—substantially reduces long-term cost, particularly for documentation obligations that must be generated contemporaneously.
8. Technical debt from expedient earlier decisions is a primary impediment to AI adoption in established organizations, frequently a larger obstacle than either capability or capital.
9. AI talent market: senior machine learning engineer compensation commonly reached $300,000–$500,000 and above in 2024, with retention presenting greater difficulty than recruitment.
10. Responsible AI frameworks include those from IEEE, the Partnership on AI, and constitutional approaches to model training; adoption varies widely and audit against actual behavior is rare.