Today's Winners and Tomorrow's Winners

Success can make yesterday's operating model look permanent.

Protect what compounds

AI does not repeal commercial fundamentals. A weak product does not become compelling because its catalogue is machine-readable. Poor service does not become good because it is automated. Unreliable fulfilment does not become reliable because a new interface promises it faster.

Today's winners earned their position through some combination of product quality, a clear proposition, brand trust, reliable execution, customer knowledge and disciplined economics. Those strengths remain the durable core.

The additional requirement is adaptive capacity: the ability to detect a meaningful change, redesign the affected work, set clear decision boundaries and measure whether the new model is genuinely better.

Earned advantage

Product, proposition, brand trust, reliable execution, customer knowledge and disciplined economics.

Adaptive capacity

The ability to detect change, redesign work, define decision boundaries and learn from operating evidence.

Tomorrow's winners need both. Adaptation without a durable core creates activity without value. A durable core without adaptation allows yesterday's strengths to conceal today's drift.

What is already changing

AI-mediated discovery is commercially observable now. Adobe Digital Insights reported that referrals from AI sources to US retail sites rose 393% year on year in the first quarter of 2026 and converted 42% better than non-AI referrals in March. That does not make AI a primary channel in every category: Adobe measures referred visits, not every recommendation, market or zero-click journey. It does establish that the behaviour is real enough to measure.

Product research and comparison inside generative interfaces are also live, while the commerce layer remains unsettled. OpenAI's current commerce direction emphasises product discovery and merchant-owned checkout after its first Instant Checkout design proved too inflexible. The capability is advancing, but there is no single mature "agent channel" with uniform reach, data or measurement. Influence Humans and Agents examines that discovery problem in detail.

Service and internal work are changing too, at different speeds. Zendesk documents how service agents can resolve or route some enquiries and why complex, urgent or sensitive cases need designed escalation paths. A peer-reviewed study of one customer-support operation found that AI assistance increased productivity by about 15%, with larger benefits for less experienced workers, but its authors explicitly warn against generalising from one firm and tool. The study is evidence that workflow-level gains are possible, not proof of universal automation.

What remains emerging

The strategic mistake is to treat a changing customer interface as the whole transformation. Discovery is one surface. The harder question is whether the operation behind it can absorb new capabilities without creating contradictory data, unmanaged decisions or new forms of cost.

Keep the strength. Upgrade the system.

Tool adoption is not the same as operating-model improvement. An OECD review of AI adoption in firms identifies data readiness, use-case and return uncertainty, implementation cost and change management as recurring barriers. Research from DORA reaches a related conclusion in software-development teams: AI tends to amplify the strengths and weaknesses of the system around it. That evidence comes from technology work, not e-commerce generally, but the operating lesson is useful.

Adding an AI service agent to an inconsistent returns process does not modernise the operation. It accelerates inconsistency. The upgrade begins with one current policy source, defined low-risk decisions, explicit escalation triggers and measurement of resolution quality, repeat contact and cost together.

The same test can be applied across the business:

The economic test runs through every lane. Benefits must survive beyond a demonstration or isolated task. Quality, speed, cost and cash need to be read together; improving one while quietly exporting work to another team is not an operating gain.

Drift appears before decline

Operating drift rarely announces itself first in the P&L. It appears as disagreement between systems, unresolved exceptions and work that becomes harder to explain. By the time the commercial result is obvious, the company may have been compensating manually for months.

These are signals worth investigating, not predictions that every business will follow the same path:

Measurement is improving, but it remains partial. Google began rolling out dedicated generative-AI Search Console reporting to only a subset of sites in June 2026. The practical response is not to wait for perfect attribution. It is to state what can be observed, record the limits and avoid turning directional evidence into certainty.

Run the next-winner review

The NIST AI Risk Management Framework recommends making roles, knowledge limits, human oversight, expected benefits and costs, benchmarks and production monitoring explicit. It is a voluntary, cross-sector framework rather than an e-commerce playbook, but it provides a useful discipline for an operating review.

For one commercially important workflow, ask:

1. What existing advantage must this change not degrade? Name the customer promise, economic discipline or operating strength that remains non-negotiable.
2. What behaviour has changed enough to justify redesign? Separate observed customer or operating evidence from a forecast and from vendor claims.
3. Which workflow is constrained, and who owns its end-to-end outcome? A faster task inside a broken flow is still a broken flow.
4. What may the system decide, what should it recommend and when must it escalate? Change the boundary according to consequence, reversibility and context.
5. What evidence within the next 90 days would show that the new model is better? Choose measures of quality, speed, cost or cash before the pilot begins.

A deployment is not an operating result. If the answer is "we deployed the tool," the operating question has not yet been answered.

Build a shorter learning loop

Adjacent peers can help expose what internal familiarity hides. Their value is not a universal playbook, but a clearer view of whether what feels normal inside one company is already drift outside it. Comparison is useful only when the businesses, definitions and context are genuinely comparable.

Staying a winner is less about predicting every new interface than building an operation that can detect change early, act deliberately and learn without eroding the brand it exists to serve.

The durable advantage is a shorter distance between signal, decision and correction.