The Enterprise AI Advantage Is Not the Model. It's the Glue.

In May 2026, Goldman Sachs published projections showing that token use by AI agents is expected to multiply 24 times by 2030. In that same month, Uber's COO told Business Insider it was getting harder to justify the money spent on AI. The gap between them explains everything about where the agentic enterprise is actually headed. One company sees extraordinary scale. The other is struggling to show it pays off. Both are right.

Everyone Has the Magic Box

The LLM is no longer a differentiator. Benchmarks from ArtificialAnalysis tracked across OpenAI, Anthropic, Google, Meta and Chinese labs show aggregate scores converging sharply through 2025 and into 2026. Sam Altman put it plainly: intelligence will be a utility, like electricity or water, bought on a meter. When every enterprise has access to the same underlying capability, the question stops being which model you use and starts being what you build around it.

This is the moment most organisations are getting wrong.

Pouring Tokens Over Your Org Is Not a Strategy

There is a term for what many enterprises did in 2024 and 2025: tokenmaxxing. Take a general-purpose model, expose it to employees via a chat interface, and hope productivity follows. It rarely did. Only 3.3% of Microsoft 365 users pay for Copilot, according to Windows Central. A Polymarket post went viral in May 2026 when an AI consultant revealed a client had accidentally spent $500 million in a single month after failing to set employee usage limits.

Token prices have collapsed. The fastest models dropped roughly 900 times in cost per year between 2021 and 2025, according to Epoch AI. The problem is not the cost of intelligence. The problem is that unstructured access to intelligence produces unstructured results.

Work Is Just an Algorithm

Every job in an enterprise is a series of steps to accomplish a goal. A workflow. Every enterprise is a collection of those workflows, layered, interlocked, running in parallel across departments. Most of them today are held together by humans doing manual, fragmented, slow work: re-keying data between systems, forwarding emails, chasing approvals, repeating context from one meeting to the next.

That is not a technology problem. It is a design problem.

When you map a complex process like patient onboarding in healthcare, what currently takes two weeks (involving a therapist, a digital health platform, an external supplier, a clinician and the patient, each waiting on the others) can run in two hours with agents handling parallel workflows, automatic notifications, order placement and access provisioning. The human role does not disappear. It shifts left, from executing to verifying.

The question is not whether AI can do this. It already can. The question is whether your organisation is designed to take advantage of it.

The Model Is the Horse. The Harness Is the Enterprise AI Advantage.

An AI agent is an LLM that can act in the real world on your behalf. It observes, reasons, plans, acts, remembers and repeats. It needs a brain (the model), tools (CLI, APIs, browsers, files), memory and a loop that keeps running when no one is talking to it. OpenClaw, the open-source personal assistant that went viral in early 2026 and was built entirely by an agent, is an elegant proof of concept: this is not complicated in principle. What makes it hard at enterprise scale is everything around it.

The harness (the orchestration, memory management, retry logic, routing and governance layer) is where most enterprise AI projects succeed or fail. And critically, using an LLM should not bind you to a single vendor. The model is a commodity. The harness is yours.

This has a strategic implication that many boards are not yet taking seriously: when your operating model becomes software, are you going to rent it from someone else? Kirkland & Ellis, the world's highest-grossing law firm, announced plans to spend $500 million building its own AI technology, as reported by the Financial Times in 2026. They are encoding the collective intelligence of their lawyers into a platform they own. That is not a technology investment. It is a moat.

Context Is the Enterprise AI Moat

The most overlooked component of enterprise AI is ontology. The relationship between your data and language. CRM calls them "customers." Core banking calls them "account holders." Marketing calls them "clients." Nobody is quite sure if they are talking about the same person. When you ask an agent to act on customer data, it will guess. Guessing at scale is expensive.

Context has to be explicitly designed. It means mapping your business entities, agreeing what things are called, defining the rules that govern decisions, and connecting the right data sources to the right agents at the right time. Done well, this is your competitive advantage. A generic LLM trained on the internet cannot replicate your context. It compounds with every interaction: better context produces better outputs, better outputs produce better decisions, better decisions generate better data.

The enterprise that designs its context layer properly will outperform one with a better model. Every time.

The Agentic Enterprise Is Not a Chatbot

There is a maturity curve here. At one end: chatbots, prompt-and-response, individual productivity gains from a grassroots install of Claude or ChatGPT. At the other: a genuinely agentic enterprise, where agents coordinate across functions, share memory, route work, verify their own outputs and operate within defined governance constraints. Most organisations are still at one end. A few are starting to approach the other.

The architecture of the agentic enterprise has five layers. Vision and operating model. Context: trusted data and your system of record. A harness: orchestration, memory and routing. Generative UI that surfaces the right interface for each task. And security and governance built in as runtime constraints, not policy PDFs.

Paul Graham wrote in June 2026 that incumbents failing to generate returns on LLM token costs is exactly what you would expect with any new technology. They cannot use it well, and they are replaced by upstarts who can.

The question is which side of that transition you want to be on.

Humans Are the Long Pole in the Tent

None of this displaces people. Research published by the Financial Times in 2026, drawing on Kharazian et al., found that companies with high AI adoption increased headcount by 10.2% on average across all jobs. Entry-level roles showed a 12% average increase. Apollo's chief economist Torsten Slok found zero evidence of AI-related job losses as of May 2026. Call centre employment in the Philippines continued to grow.

What changes is what people do. The Jevons paradox, described by William Stanley Jevons in 1865 in relation to coal and steam efficiency, shows that when a resource becomes cheaper through technology, demand expands dramatically rather than contracting. Coding productivity has exploded: new websites, new iOS apps and GitHub code pushes all up 20 to 40% year on year as of 2026, according to Financial Times data. Stripe is now merging 1,300 pull requests per week written entirely by agents, with no human code. But someone still needs to understand what the system does and maintain it.

The human role shifts to determining what problems are worth solving, inventing new products and services, and optimising workflows from above. That is a more valuable role. But it requires being inside the design of your AI systems, not downstream of them.

What This Means for You

The enterprises pulling ahead right now are not the ones with the biggest AI budget or the most advanced model access. They are the ones investing in the glue: the context layer, the harness, the workflow redesign, the governance constraints that turn a general-purpose model into a reliable worker doing a real job, following your process, connected to your systems.

At Elsewhen, this is the work we do with enterprise clients across financial services, healthcare, travel and the public sector. Our AI Productivity Platform is not a SaaS product. It is a set of components: Business Context, Headless Agents, Orchestrator, Skills Factory and Generative UI, designed to accelerate delivery while keeping you as the owner of the IP. Four weeks from working session to working prototype. No lock-in.

The model is a commodity. The architecture is the advantage. Build it before someone else builds it for you.