Short answer: For most companies, no. The leading models are converging in quality and falling in price, and your competitors can rent the same ones. The advantage now sits in the layer after the model: which workflow you redesign, where a person approves, whether the team adopts it, and whether you measure the result.
The leading models have converged
In early 2023, one lab had a clear lead on the Arena leaderboard, where people vote on model answers head to head. Access to the best model was a real edge.
That edge has thinned. The 2026 AI Index from Stanford HAI reports that the top four models on that leaderboard are now separated by fewer than 25 points. A year earlier, they spanned roughly 97. In professional evaluations for tax, mortgage processing, corporate finance and legal reasoning, the top 15 models sit as little as 3 percentage points apart. The report draws the conclusion itself: “With capability no longer a clear differentiator, competitive pressure is shifting toward cost, reliability, and real-world usefulness.”
Price has moved too. The 2025 AI Index found that querying a model at GPT-3.5 level fell from $20 to $0.07 per million tokens between November 2022 and October 2024, a drop of more than 280-fold. Epoch AI estimates that the price of reaching a given level of performance falls between 9 and 900 times per year, depending on the task.
So the model your competitor uses is almost certainly a model you can use too, through the same APIs and cloud platforms. Something everyone can rent only gets you to the starting line.
Most companies are stuck between pilot and daily use
Getting started is no longer the problem. Drawing on McKinsey survey data, the AI Index reports that 88% of respondents said their organization used AI in at least one business function in 2025, up from 78% a year earlier. Depth is another matter. The same chapter finds AI agent deployment “in the single digits across nearly all business functions.”
The most quoted number of the past year came from MIT’s Project NANDA: 95% of organizations getting zero return on generative AI. It needs careful handling. The report draws on interviews with 52 organizations, 153 survey responses and a review of over 300 public AI initiatives. The authors call some of their figures “directionally accurate,” measured return six months after a pilot, and counted a project as successful only if it had a “marked and sustained” impact. Unsuccessful does not mean zero return, a point critics raised soon after publication.
The headline number is weak. The diagnosis is more useful. The authors found that the divide “does not seem to be driven by model quality or regulation, but seems to be determined by approach.” Most systems, they write, “do not retain feedback, adapt to context, or improve over time.” Pilots fail on brittle workflows and poor fit with daily operations. Those are not model problems. They are design and implementation problems.
The value comes from changing the work
Larger samples point the same way.
BCG’s AI at Work 2026 survey of 11,749 people across 14 markets offers a useful comparison. Among regular AI users with an unclear AI strategy and limited tools, 55% reported measurable business impact. With strong tool access but still no clear strategy, the figure was 60%. With a clear strategy but limited tools, it was 80%. Respondents at companies pursuing workflow redesign were 24 percentage points more likely to see measurable business improvement.
McKinsey’s 2026 State of AI survey shows the same pattern. Among the roughly 6 percent of respondents it calls AI high performers, nearly three-quarters report fundamentally redesigning workflows because of their AI use. Among everyone else, it is one-quarter.
The same BCG survey shows where value leaks. Among frontline employees, 74% are now regular AI users. Yet 66% get limited or no guidance on what to do with the time they save. BCG puts it plainly: “The time individuals save doesn’t automatically translate into value.”
BCG also has a rule of thumb for where the effort belongs. In its 10-20-70 model, 10% goes into algorithms, 20% into technology and data, and 70% into “people, organization, and processes.” Its 2026 Applied AI Index finds that almost 50% of companies now generate value from AI. Among the 7.5% it calls “future-built,” 61% have moved from isolated pilots to a multiyear, enterprise-wide program.
These are self-reported surveys, so read the exact numbers as directional. The direction is consistent across all of them.
The advantage sits in the layer after the model
If the model is a shared input, the advantage lies in what a company builds around it. We work through that layer in four steps: Context, Workflow, Action and Learning. Inside those steps sit six parts:
Workflow redesign. Choose one process and change how it runs end to end, rather than adding a chat window to the old version.
Company context. Connect the system to the documents, records and rules that make an answer correct for this business, not just plausible.
Integration. Put the system inside the tools the team already uses, such as the inbox, the CRM or the document system, so it is part of the work rather than another tab.
Approval points. Decide where a person reviews, edits or signs off, and where the system stops. Write it down.
Adoption. Train the team on the new workflow, not just the tool. Tell people what to do with the time it gives back.
Measurement. Agree on the metric before launch, whether cycle time, error rate or cost per case, and track it afterward.
None of this comes with a model subscription. It depends on knowledge only the company has: how its work actually runs, where its risks sit, what its customers expect. That is also why it is hard to copy. A competitor can rent the same model tomorrow. It cannot rent your redesigned workflow, your connected data or a team that has used the system every day for a year.
The layer also compounds. Each workflow in production teaches the team what to measure and where people need to stay in the loop. That knowledge carries into the next one.
The model still matters at the edges
This argument has limits, and they are worth stating.
At the frontier, models still move fast, and the differences are real. The AI Index reports that frontier models gained 30 percentage points in a single year on Humanity’s Last Exam, a benchmark built to be hard for AI. AI agents improved sharply but “still fail roughly one in three attempts on structured benchmarks.” The gap between the best closed and best open-weight models has reopened, to 3.3% in March 2026. Physical tasks lag far behind: robots succeed in only 12% of real household tasks.
So for work at the edge of what models can do, model choice is a real decision. That includes long multi-step agent tasks, hard reasoning problems and specific modalities such as video. The wrong choice can stall a project. For labs, and for companies whose product is the model, it is the central decision.
For most companies, though, model choice is an engineering decision. It gets revisited every few months as prices fall and new releases arrive. The sensible response is to build so the model can be swapped without rebuilding the workflow. The workflow, the approvals, the data connections and the team’s habits should outlast any single model.
Start with one workflow
The companies that pull ahead over the next few years will not be the ones with exclusive access to a better model. There will be very few of those. They will be the ones that took one important workflow, rebuilt it around AI, decided who approves what, trained the people who run it and measured the result. Then they did it again.
That work is slower than signing up for a model. It is also the part that stays yours.
If you are deciding where to begin, our guide on how to choose your first AI workflow is a practical place to start.
If you already have a workflow in mind, book a fit check. We’ll tell you whether it’s worth a diagnostic.
Frequently asked questions
Do the leading AI models still differ much?
Less than they did. Stanford HAI’s 2026 AI Index reports that the top four models on the Arena leaderboard are now fewer than 25 points apart, against roughly 97 a year earlier. The price of GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024.
Where does the advantage come from, if not the model?
From what a company builds around it. We work through four steps, Context, Workflow, Action and Learning, with six parts inside them: workflow redesign, company context, integration, approval points, adoption and measurement. None of this comes with a model subscription, which is why a competitor cannot rent it.
When does the choice of model still matter?
At the edges. For long multi-step agent tasks, hard reasoning problems and specific modalities such as video, the wrong model can stall a project. For most companies, model choice is an engineering decision that gets revisited every few months. Build the workflow so the model can be swapped.
Why do so many AI pilots never reach daily use?
MIT’s Project NANDA found that most systems do not retain feedback, adapt to context or improve over time. Pilots fail on brittle workflows and poor fit with daily operations. Those are design and implementation problems. The report’s headline numbers are weak, but its diagnosis is useful.
How should a company start?
With one important workflow. Rebuild it around AI, decide who approves what, train the people who run it and measure the result. Then do it again with the next one. That work is slower than signing up for a model, and it is the part that stays yours.
Related reading
Sources
Stanford HAI, AI Index Report 2026, Chapter 2: Technical Performance, April 2026.
Stanford HAI, AI Index Report 2026, Chapter 4: Economy, April 2026. Adoption figures drawn from McKinsey’s 2025 State of AI survey.
Stanford HAI, AI Index Report 2025, Chapter 1: Research and Development, April 7, 2025.
Epoch AI, LLM inference prices have fallen rapidly but unequally across tasks, March 12, 2025.
MIT NANDA, The GenAI Divide: State of AI in Business 2025, July 2025. Preliminary findings.
Futuriom, Why We Don’t Believe MIT NANDA’s Weird AI Study, August 26, 2025.
BCG, AI at Work 2026: Strategy Matters More Than Tools, June 2026.
BCG, AI Is Reshaping Jobs Faster Than Companies Are Reshaping Work, press release, June 3, 2026.
BCG, The Formula for Agentic AI Value (Applied AI Index 2026), September 30, 2026.
McKinsey & Company, The state of AI in 2026: On the road to ROI, August 25, 2026.