Almost every AI conversation we have with a Kuwait or GCC business eventually narrows to one question: ChatGPT vs custom AI assistant. Both are legitimate answers, they cost very different amounts, and the deciding factor is not how advanced the model is — it is whether the answers your team needs live inside your own systems or out on the open internet.
What a ChatGPT subscription already does well
For drafting, translating, summarising and thinking out loud, a team subscription to ChatGPT, Claude or Microsoft Copilot is the best value in business software right now. A marketing lead in Kuwait City writes bilingual campaign copy. A finance manager asks why a variance moved. An operations supervisor turns messy notes into a clean handover. None of that needs a project — buy seats, run a short internal session so people actually use them, and move on. The usual blocker is habit, not capability, which is why AI training for employees tends to return more than any tool purchase in the first quarter.
The subscription stops being enough at a specific moment: when the answer has to come from your data. A generic assistant does not know your 2026 price list, what is on the shelf in Shuwaikh this morning, which clauses your legal team refuses to sign, or what you quoted that customer in March. Staff then paste sensitive extracts into a chat window to compensate — which is the real risk, not the model itself.
What a custom AI assistant actually is
It is almost never a model trained from scratch. In practice it is a commercial model wrapped in three things you own. First, retrieval over your documents and databases, the pattern known as retrieval-augmented generation, so every answer is grounded in your records and can cite them. Second, permissions, so a branch manager sees branch numbers and a salesperson sees their own pipeline. Third, connections into the systems where work actually happens — ERP, CRM, WhatsApp, email, the shared drive. Microsoft's RAG overview describes the same architecture in vendor-neutral terms.
Language is where GCC projects diverge from the templates. Generic tools handle Modern Standard Arabic reasonably, but your real content is mixed Arabic and English with Gulf dialect in the free-text fields — customer notes, delivery remarks, complaints. Getting that right is its own body of work, covered in Arabic AI translation. If you already run a product or portal, the honest answer is usually not a separate assistant at all: you add AI to your existing app where users already log in. And if the data that matters sits in Odoo, SAP or Microsoft Dynamics, the integration is the project — see AI ERP integration.
Five questions that settle it
- Where does the answer live? Public knowledge or general writing — subscription. Your records — build.
- Does it need to be right, or just useful? A draft email can be 80% right. A price quote, a wage calculation or a stock figure cannot.
- Who asks? Ten internal staff can learn to prompt well. Two thousand customers on WhatsApp cannot, so the assistant has to be constrained, grounded and logged.
- Does anything have to happen afterwards? If the output should create a ticket, book a slot or update a record, you need an integration, not a chat window.
- What cannot leave your control? Patient data, salaries, tender pricing and signed contracts usually force a controlled deployment with retention rules you set.
What each option costs
Seats are predictable: roughly KWD 8–12 per user per month for a business tier, so a 30-person team lands around KWD 3,000–4,500 a year with no build risk. A custom assistant is a project: a narrow, single-source internal assistant is typically a few thousand KWD and four to eight weeks, while a customer-facing bilingual assistant wired into ERP and WhatsApp is materially more, plus monthly hosting and model usage that usually sits in the tens of dinars per month until volume grows. Current bands for the local market are in AI automation pricing in Kuwait. The useful comparison is not subscription versus build — it is build cost versus the hours your team currently spends looking things up.
The sequence that works
Buy seats first. Two or three weeks of real use tells you exactly which questions people keep asking that the subscription cannot answer, and that list becomes your build scope — far more reliable than a workshop wishlist. Then build one assistant for the single highest-volume question, with a measurable baseline: time to answer, deflected calls, or errors caught.
That grounding requirement is the whole job. On WCS, a factory timesheets, wages and budgets system, no general-purpose tool could have answered a single useful question, because every answer depended on the company's own hours, rates and budget lines. Platforms with many sellers face the same constraint from the other direction, where catalogue and supplier data drive the answers — the pattern is set out in AI automation for marketplaces.
If you want this mapped against your actual systems, data and Arabic content before committing budget, that assessment is where our AI work starts.