AI Development with Claude
AI agents, automation and Arabic-language products, built on Claude. It is the model I reach for most — for long documents, for Arabic, and for anything where being wrong quietly is worse than being slow — and this page also covers where it is the wrong choice.
Model choice is a build decision, not a brand preference.
Picking a model is an engineering decision, not a subscription
Most "which AI should we use" conversations are really about a specific workload: this volume of documents, this language mix, this tolerance for a confident wrong answer. Those constraints pick the model. The brand argument is downstream of them, and usually irrelevant.
Where Claude tends to win is long-context work — a contract, a year of tickets, a whole codebase — and tasks where you would rather it say it is unsure than invent something plausible. That last property is worth more in production than it looks in a demo, because the failures it prevents are the expensive kind: the ones nobody notices for a month.
Where it does not win, I use something else, and the systems I build are written so that swapping the model is a configuration change rather than a rewrite. Anyone who tells you their AI product is permanently married to one provider has built it wrong.
What actually gets built with it
The workloads where it has earned its place, rather than the full capability list.
Agents that do a job
Read an inbox, classify it, pull the details out, act, and hand over what needs a person. Claude's willingness to stop and flag uncertainty is what makes an agent safe enough to leave running.
Automation with judgement in it
The steps a rules engine cannot express — is this complaint urgent, does this invoice match, is this reply good enough to send. The deterministic parts stay deterministic.
Document processing
Contracts, reports, statements and the long PDF nobody wants to read. Long context is the difference between summarising a document and answering questions about all of it.
Arabic and bilingual work
Where Claude is genuinely strong, and the reason it is usually my default for this region — see the section below.
Building software with Claude Code
Used daily on real work: migrations, test coverage, refactors and the mechanical half of a build. It changes how fast the first version arrives; it does not change who is accountable for the last ten percent.
Evaluation and guardrails
The unglamorous part. What the model is allowed to do unsupervised, what gets logged, what a bad answer costs, and how you would know. Skipping this is how pilots quietly never ship.
Two ways to start
Most enquiries are one of these. The first call is the same for both: thirty minutes on the process, what it costs you now, and whether a model belongs in it at all.
A workload you want automated
Find the workload, build the evaluation, ship the smallest useful version — the process below, end to end. For the inbox, the documents or the handoffs that eat someone's week.
An AI build that never quite shipped
A review of a pilot that stalled or an assistant nobody trusts: usually the scope was too wide, the evaluation never existed, or the model is doing a job code should do. Most of these are fixable for less than they cost so far.
Arabic is where the choice actually matters
Most model comparisons are run in English, which makes them close to useless for deciding what to deploy in Kuwait or the UAE. The real test is Gulf dialect rather than textbook Modern Standard Arabic, messages that switch between Arabic and English mid-sentence, and names and addresses that transliterate three different ways.
Claude handles that mix better than most, which is why it is usually my starting point here — but "better than most" is not "solved", and the only honest way to choose is to evaluate against your own data. A hundred real messages from your own inbox will tell you more than any benchmark.
The build matters as much as the model. Arabic-first means the prompts, the evaluation set, the interface and the fallbacks are built that way from the first version, not translated afterwards.
Where I would not use it
Stated plainly, because the alternative is discovering it after you have paid for the build.
The systems I build keep the model behind an interface, so none of these are permanent decisions.
How a Claude build runs
Find the workload
One process, traced end to end, with a number attached to what it currently costs. Not "add AI to support" — the six ticket types that make up most of the volume.
Build the evaluation first
Real examples from your data with known-good answers. Without it, "it seems better" is the only measure available, and it is not a measure.
Smallest useful version
Narrow scope, real data, running alongside the existing process rather than replacing it on day one.
Decide what it may do alone
Drafting is not sending. Suggesting is not deciding. The boundary is set deliberately and widened only once the evaluation says it should be.
Hand it over
Documented, owned by someone named, with the evaluation left in place so quality can be checked later rather than assumed.
What a project hands over
Scope varies — a review of a stalled build ends in findings and a prioritised fix list — but a full build includes:
Code, prompts, evaluation data and provider accounts are set up as yours from the first day — and the model stays behind an interface, so switching it later is a configuration change, not a renegotiation.
Discuss your AI workload
Discuss your AI workload
A new build or a pilot that never shipped — the first call scopes either. Bring the process and roughly what it costs you now; that is all the preparation needed.
Related
What I've written about Claude
Model comparisons, pricing analysis and what the coding agents are genuinely like to work with.
Common questions
The questions that come up in the first call, answered the way they would be answered there.
Choosing the model
Should we use Claude, ChatGPT or Gemini?
It depends on the workload, and the honest answer usually involves more than one. Long documents, careful reasoning and Arabic push toward Claude. High-volume cheap classification often does not need a frontier model at all. I build so the choice can change without a rewrite, and I have written the comparisons up in detail rather than asking you to take my word for it.
Is Claude good at Arabic?
Better than most, in my experience, particularly with Gulf dialect and mixed Arabic-English input — which is what real messages in this region look like. But it is worth evaluating against your own data rather than any benchmark, and that evaluation is part of how I start.
When is Claude the wrong choice?
High-volume classification a small cheap model handles, anything that must be deterministic, realtime voice, and data that cannot leave a jurisdiction with no compliant deployment available — the list above states it plainly. The wrong-er choice is a model doing a job code should do: a broken process automated is a broken process running faster.
Will we be locked in to one provider?
No — that is a build decision, and I make it the boring way: the model sits behind an interface, and the evaluation set makes any alternative measurable. When a better or cheaper model ships, you swap it in and run the evaluation. Anyone whose AI product is permanently married to one provider has built it wrong.
Pricing and implementation
What does a Claude project cost to build?
It depends on the workload — an agent working an inbox, a document pipeline and a review of a stalled pilot are different sizes of work, and the integration is normally the larger half. Scope is agreed on a short call and you get a clear quote before anything is built; what never happens is an estimate that quietly grows.
What does this cost to run?
Model cost is usually the small part; the integration and the evaluation are the real budget. I cost the running side before building, because an automation that costs more than the process it replaced is not a saving.
How long does a build take?
It depends on the workload and the integrations, so a number quoted before scoping would be a guess. What the process guarantees instead is that the smallest useful version runs early, alongside the existing process — value shows up in the first weeks rather than at a distant launch, and each widening of scope is earned by the evaluation.
What is included in a project?
The workload definition, the evaluation set, the working integrated system, the boundary in writing, guardrails, documentation and a handover to a named owner — the full list is under "What a project hands over" above. A review of an existing build is the smaller shape: findings and a prioritised fix list.
Build and safety
Can you connect it to our existing systems?
Yes, and that is usually where the value is. A model with no access to your CRM, inbox or documents is a chatbot; connected properly it can actually do the work. The integration work is normally the larger half of the project.
What stops it from making things up?
Nothing stops it entirely, and anyone claiming otherwise is selling something. What a production build does is contain it: a narrow job, grounding in your actual data, an evaluation that measures wrong answers, a boundary on what the model may do alone, and fallbacks for the inputs it should not attempt. Claude's willingness to flag uncertainty instead of inventing is why it is my usual starting point for exactly this reason.
Is our data used to train the model?
Anthropic's commercial terms do not use business API data for training by default. The part that deserves more attention is your own build: what gets sent to the model, what gets logged, and where it is stored are design decisions, and they get made deliberately and written down rather than left to defaults.
Do you use AI to write the code you deliver?
Yes, Claude Code daily, for migrations, tests, refactors and the mechanical half of a build. It is reviewed, and it is my responsibility when it ships — AI-assisted does not mean unreviewed. It means more of the budget goes to the parts that need judgement.
After launch and ownership
Who owns what gets built?
You do, from the first day. The code, the prompts, the evaluation set and the provider accounts it runs under are set up as yours — usage, data and billing stay under your control, and nothing is held hostage to the relationship.
Do you provide support after launch?
The handover is designed so you are not dependent on it: documentation, a named owner on your side, and the evaluation left in place so quality can be checked rather than assumed. Ongoing support is scoped per engagement — some teams want it handled, some want to run it themselves.
What happens when a better model ships?
You swap it in and run the evaluation. Because the model sits behind an interface and the evaluation is part of the system, a model upgrade is a measured configuration change rather than a rebuild — which is most of the reason it is built that way.
Do you work with businesses outside Kuwait and the UAE?
Yes — I am based in Kuwait and most of this work runs remotely, in shared documents and calls. Gulf hours overlap comfortably with European and Asian workdays, and the Arabic-first work travels particularly well across the region.
Got a workload you think Claude could take?
Got a workload you think Claude could take?
Describe the process and roughly what it costs you now. Thirty minutes is usually enough to say whether it is worth building, and which model it should be built on.