Enterprise AI Isn't Sticky: What Gulf Buyers Must Know

In May 2026, Anthropic quietly overtook OpenAI among paying business users. By July, the gap had widened: Anthropic held nearly 44% share against OpenAI's nearly 40%, according to Ramp's analysis of spending data across more than 70,000 American businesses. ^[1] Then, mid-quarter, OpenAI began closing the gap again. One quarter. Multiple lead changes. Thirty AI years of news cycles compressed into a single earnings period.

For Gulf operators embedding AI into ERP and workflow layers, that volatility is not a trivia point — it is a structural warning.

Businesses Are Switching AI Vendors Like Streaming Services — Here's the Data

The Ramp data ^[1] is worth pausing on. These are not consumers experimenting with free tiers. These are businesses spending real budget through corporate cards and bill-pay products. The fact that vendor preference among this group shifted measurably in a single quarter tells you something clear: enterprise AI has almost no switching costs today.

That is unusual. In most enterprise software categories — ERP, CRM, HRIS — switching costs are the entire business model. Implementation inertia, data migration risk, and retraining friction lock customers in for years. AI models have none of that. When Anthropic ships Claude 4 and GPT-5 lands a week later, a business can literally update one API key and the model changes. The prompt library stays. The spend management dashboard stays. The vendor does not.

Ramp's economist noted that OpenAI was growing faster among the segment in Q3 to date, but acknowledged there was still a month left in the quarter — enough time for the lead to flip again. ^[1] That is the rhythm now: sprint, counter-sprint, repeat.

Why Low Switching Costs Sound Good but Create Real Operational Risk

On the surface, zero switching costs sound like buyer leverage. Competition keeps quality up and prices honest. That is true at the model layer. The problem sits one level deeper.

If your AI deployment is architected for easy model swapping, low switching costs are an asset. If your AI deployment is hardcoded to a specific model — prompt formats, output parsers, ERP field mappings, approval triggers — then every model upgrade is a potential breaking change wrapped in a feature announcement.

Consider the operational texture of a mid-sized Gulf trading company: procurement approvals routed through a WhatsApp-to-ERP bridge (a gap we've written about here), freight cut-offs timed to Ramadan schedules, supplier master data spread across seventeen spreadsheets and two legacy systems. Now embed an AI layer that summarises outstanding POs and flags anomalies. If that AI layer is directly coupled to a specific model's output schema, a model version change can silently corrupt the field mappings. The dashboard still reports success. The operations quietly fail. We've covered exactly this failure mode in our piece on automation silent failures.

Low switching costs at the model layer create high rebuilding costs at the operations layer — unless the architecture accounts for it.

The Layer That Actually Determines Your AI ROI (It's Not the Model)

Enterprise AI deployments sit across several architectural tiers. The model — GPT, Claude, Gemini, Llama — is one tier. Below it and above it are the layers that actually determine whether AI produces business value: data pipelines that feed clean, contextual inputs; API contracts that translate model outputs into structured records; workflow logic that routes those records into ERP transactions, approval chains, and reporting.

This integration layer is the connective tissue. ^[3] It is also the layer that is hardest to replace — and therefore the right place to concentrate architectural investment. A well-designed integration layer is model-agnostic by construction: it defines the contract (inputs, outputs, data types) and lets the model behind it be swapped without touching downstream systems.

The Stripe/OpenRouter move — which we analysed separately — is a commercial signal pointing in the same direction: the industry is commoditising the model layer deliberately, because the durable value lies in routing, orchestration, and integration, not in which transformer weights sit underneath.

This is also why a dashboard is not a decision. Dashboards report on the model layer. Decision layers act on the integration layer. Confusing the two is how companies buy impressive demos and see no operational lift.

What Gulf ERP Buyers Should Demand from Any AI Integration

Before signing any AI contract — whether it is a Copilot seat, a Claude API integration, or a packaged "AI module" from an ERP vendor — Gulf buyers should run four tests:

1. Model substitution test. Ask the vendor: if we replace the underlying model tomorrow, what breaks? Document the answer in writing. If "everything" is the honest reply, the architecture is wrong.
2. Data schema ownership test. Confirm that your input data schemas, output field mappings, and prompt templates are owned by you — not stored in the vendor's proprietary format. Proprietary formats are switching costs in disguise.
3. Integration middleware independence test. Verify that the middleware connecting the AI layer to your ERP (Odoo, SAP, Dynamics — see our Odoo vs SAP comparison for context) is not the AI vendor's own connector. Vendor-bundled connectors create dependency. Standard API layers do not.
4. Monitoring and failure audit test. Confirm there is a mechanism to detect when the AI layer silently produces wrong outputs — not just when it goes down. Silent failures are the most common and least caught. Our ERP AI readiness audit covers how to check for this before you go live.

These are not exotic requests. Any vendor building for enterprise longevity should answer them without hesitation. Hesitation is information.

For context on what healthy AI integration architecture looks like inside connected enterprise systems, the principle is consistent: integration sits at the centre, and the AI model is one pluggable component within it — not the foundation. ^[3]

Tarsyn's View: Build the Foundation, Not the Bet

The Ramp data ^[1] will keep updating. Next quarter, OpenAI may be ahead. The quarter after, a new entrant may hold the lead. That churn is structural, not temporary — because the model layer is genuinely commoditising, and competition is working exactly as intended.

The wrong response is to track that data and try to pick the winning model. Multiply chaos by intelligence and you get articulate chaos. Betting on a specific model is the wrong unit of analysis. The durable asset is the data and workflow foundation underneath it.

What that looks like in practice: clean, documented data pipelines before any AI layer goes live. Vendor-agnostic API contracts. Workflow logic that is version-controlled and testable independently of the model. Monitoring that catches output drift, not just downtime. This is precisely the checklist we run through in our five-step audit before any client commits spend.

We run that audit with clients before recommending any AI purchase — and sometimes the honest answer is: not yet. Fix the spreadsheets first. The case for doing that has not changed just because the model options multiplied. If you want to know where your operation stands before the next model release reshuffles the market again, the Tarsyn Audit is the right starting point.

The model that wins Q3 2026 will not be the model that wins Q2 2027. The integration layer you build this year can last a decade — if you build it right.