What AI systems cost to run after launch

A running AI system in the UAE costs 500 to 5,000 AED a month in model and API spend, another 200 to 2,000 in hosting, plus a maintenance retainer of 10 to 20% of the build cost per year. That retainer is the line item most people skip, and it is the one that stops a working system from quietly decaying

The build price is a headline. The running cost is what decides whether the system is still worth it in year two. Below we split the monthly bill into the parts that scale with use and the parts that scale with neglect, so you know which number on the invoice you should actually watch

What does an AI system cost per month after launch?

For a typical SME build in the UAE, the reliable answer is 500 to 5,000 AED in model and API spend, plus hosting and any Business Solution Provider fees, plus a maintenance retainer if you want monitoring and iteration. Those are the bands we quote and the bands we see in post-launch bills from other Dubai teams, and they are consistent with the run-cost table in our AI automation cost guide

The spread inside that range is almost never the model. It is volume and context. A WhatsApp bot answering a few hundred conversations a day on a mid-tier model sits at the bottom. A document pipeline that pushes a 40-page SPA through a large model with full context every time sits at the top. The fix is routing: cheap model for the routine path, expensive model only for the case that needs it. We usually cut a monthly model bill in half that way

What are the recurring cost bands, line by line?

This is the monthly breakdown we hand clients so they can see where the money goes after launch. Treat the numbers as typical bands for UAE builds, not a rate card

The two rows that surprise people are at opposite ends. Model spend can shrink with routing, so it is not a fixed burden. Monitoring is often zero until the first silent failure, after which it is the cheapest insurance you can buy. A system that fails without anyone noticing costs you customers before it costs you budget

What breaks in an AI system over time?

The maintenance item most teams forget is that nothing in a running AI system stays still. Four things rot on their own schedule

Prompt drift. Your workflow data changes. In real estate, the formats of Property Finder and Bayut lead exports shift, developers change off-plan feeds, and the vocabulary in enquiries moves. Prompts tuned against last year's data answer this year's inputs slightly worse every month, until the accuracy you signed off on is no longer the accuracy you have. The fix is a monthly re-evaluation against fresh inputs, not a rewrite of the prompt the first time it fails

Template re-approval. On WhatsApp, every automated message that leaves the 24-hour customer service window must be a pre-approved template straddling the Business Messaging Policy. Marketing templates get re-reviewed and rejected when the copy drifts, when variables are misused, or when approval standards change. Someone has to notice a rejection, fix the template, and resubmit it, or your outbound flow goes dark. Template economics live in our WhatsApp Business API costs piece.

API and integration changes. Your CRM vendor ships a new API version and deprecates the old one. A BSP changes a webhook payload. Meta adjusts template rules. Each is a small change on their side and an unplanned fix on yours. If nobody is responsible for noticing, the integration breaks and you find out from a customer days later

Model deprecation. The model you tested against gets retired or repriced. The provider ships a successor that behaves differently on your edge cases. Your evaluation set, the collection of inputs you check against, is the only thing that tells you whether the new model still meets the standard. That is why an up-to-date eval set is part of the retainer, not a nice-to-have

Do I need a retainer, or can I run it myself?

If you own the code, prompts and infrastructure, and you have an engineer on staff who can read logs and redeploy, then monitoring plus occasional fixes can live in-house. That is a real option for a company with a tech team. Most SMEs in the UAE do not have that engineer, and for them a small retainer is the cheaper answer, because the alternative is a system that degrades until an incident forces an emergency call, billed at emergency rates

A maintenance retainer of 10 to 20% of build cost per year buys a defined job: monitoring for silent failures, prompt and rule updates on a schedule, template maintenance, upstream integration fixes, and a monthly re-evaluation of accuracy against fresh inputs. It is an agreement about response, not a blank cheque. Our guide to choosing an AI automation agency lists the exact support questions to ask any vendor, us included, before you sign

The decision is about who owns the risk of the system degrading. In-house works when the cost of that risk is low and your team can absorb it. A retainer works when the system sits in a revenue path, where a degraded week is money already lost

What happens if the AI vendor disappears?

With a custom build you own the source code, prompts, workflow definitions and data pipelines, so if the vendor disappears, infrastructure and API keys are in your account and a new team can pick the system up from the code. With a closed platform, your workflows die with the subscription, because the automation logic lives in someone else's builder and you never had it exported

This is the strongest argument for checking the ownership clause in the contract before you build. The agency selection checklist covers it directly: the contract should assign you the code and the prompts, and you should hold the Meta Business account and the BSP contract yourself. Retention of the run phase is a transfer-of-risk problem, and it is much cheaper to solve in the build contract than in the middle of a vendor shutdown

How do I keep monthly spend from creeping up?

All three. First, route to the cheapest model that still passes the eval on each path. Second, set a budget alert on the API provider so an unexpected usage spike pages you instead of billing you silently. Third, review the cost bands above against actual numbers every month, because model pricing and your own volume both move, and the only way to know you are still in a sane band is to look at the bill

On our builds the most common leak is not usage growth. It is the same prompt resent over and over with a large context window because nobody re-checked that a smaller model still passes on that path. The Inventory CRM system, which moved about 8,000 units out of Excel and developer PDFs with status accuracy climbing from roughly 60% to about 98%, shows the payoff of tuning the routing and the approval flow before worrying about the model tier. The running cost there stayed low because the expensive calls were reserved for the changes a human had to approve

How much time does monthly maintenance actually take?

This is the artifact nobody puts on the first invoice, so we tracked what the work really involves on a typical pipeline

At the low end that is a few hours a month; at the high end, with several integrations and a document pipeline, it is a quarter of a person's time. Which is why a retainer is usually priced as a fraction of a named engineer's salary with a response SLA, rather than per incident. The fixed monthly cost buys you the guarantee that the work is being done on a schedule, not a person sitting idle waiting for the system to break

What should the monthly review actually check?

Once a month, run the eval set against the current models, check the cost bands above against the real bill, confirm template and integration status, and log what changed. That is the entire job. It does not need a dashboard with sixteen charts; it needs one repeatable checklist that a human works through and signs

The UAE Visa Platform, where passport scans go through PaddleOCR and Gemini before an application is filed, is our clearest case of why the review matters. Processing dropped from several days to between a few hours and two days partly because the extraction quality was re-checked as document formats from intermediaries changed. The model did the extraction; the review schedule kept it correct. Skip the review, and that accuracy drifts with the first change in how the documents arrive

The takeaway

Plan for 500 to 5,000 AED a month in model and API spend after launch, 200 to 2,000 in hosting, plus a retainer of 10 to 20% of the build cost per year to keep it correct. The retainer buys prompt and model maintenance, template re-approval, integration fixes, and a monthly accuracy re-evaluation. If you own the code and have an engineer, run it in-house; otherwise a small retainer beats an emergency call. Whatever you choose, keep an eval set current, because it is the only thing that tells you the system is still the system you paid for

If you want a range for your own system, book the free audit and we will map your build, your usage and your monthly bands, and tell you which line to watch