Every serious buyer eventually asks the same question, usually in the second meeting: "Fine, but what does it cost me every month?" It's the right question, and the AI industry is strangely bad at answering it. You get hand-waving about tokens, or a pricing page that assumes you already know what a GPU-hour is. So this post is just the money. What a Hermes agent costs to keep running, month after month, on each of the two deployment paths, with rough dirham figures and a worked example against the cost it replaces.
One thing before the numbers, and I'll repeat it because it matters: every figure below is an illustrative market figure, not an INS quote. Cloud compute prices move, model prices move, exchange rates move, and your usage pattern is the biggest variable of all. Treat these as the shape of the cost, then get real numbers for your situation before you decide anything.
The two cost shapes
Hermes deploys two ways, and they don't just differ in price, they differ in the shape of the bill.
The cloud path runs through the Nous Portal. You pick a model and a server size, and the agent is always on, in the sense that it responds the moment work arrives. The trick is that it scales to zero when idle: no work, essentially no compute, essentially no charge. Your bill follows your activity.
The on-device path is a private install on hardware you own, running local models through Ollama. Here the bill inverts. You pay for the machine once, then electricity and not much else. No per-token charges, no monthly compute invoice. High floor, then nearly flat.
Neither shape is better. They suit different businesses, which is the rest of this post.
Cloud costs: what scales to zero actually means
Think about the actual rhythm of an agent working for a UAE business. A burst of activity at 8am preparing briefs. Scattered tasks through the day. Quiet evenings. A nearly silent Friday. If you paid for a server running flat-out around the clock, you'd be paying for mostly silence.
Scale-to-zero means the quiet hours cost roughly nothing. What you pay for is the work: the server time while tasks run, plus the model usage. Both are choices you control. A smaller server with an efficient model handles most business admin comfortably; you'd size up for heavy research or document-crunching workloads.
Illustrative ranges at current market rates, not quotes: a lightly used agent, a few hours of real work a day on a modest server, tends to land somewhere around AED 150 to 400 a month. A heavily used one, working most of the business day on a larger server with a stronger model, might run AED 500 to 1,500. I've seen quiet months come in under the price of a decent business lunch, and busy months spike, which is exactly the point, the spike happened because the agent was doing something.
The comparison that makes owners sit up: an always-on server of similar size, rented conventionally, could cost several times the light-usage figure while doing the same work. You'd be paying the difference for idle time.
On-device costs: pay once, then almost nothing
The private path suits a different buyer, usually one whose real constraint is confidentiality rather than cost. But the economics deserve a fair look too.
Upfront, you need capable hardware. A machine that runs local models with 64k+ context well, think a high-spec Mac or a workstation with a serious GPU, runs somewhere in the AED 8,000 to 20,000 range at current retail, illustrative again, and you may already own something suitable. After that, the running cost is electricity, call it tens of dirhams a month even with DEWA's summer rates, and eventual hardware refresh.
Spread AED 12,000 of hardware over three years and you're at about AED 330 a month, flat, regardless of how hard the agent works. So the crossover logic is simple: light or spiky usage favours cloud, because scale-to-zero keeps the bill tiny. Constant heavy usage, or a hard requirement that data never leaves your premises, favours on-device, because the marginal cost of more work is zero. We've compared the privacy side of this choice separately in Hermes cloud vs on-device if that's your deciding factor.
The worked example
Numbers mean nothing without a baseline, so here's the comparison that actually matters: what do the replaced hours cost?
Take a typical Dubai SME owner using the agent the way our clients do, morning briefs, follow-up chasing, drafting, monitoring. Say that absorbs 60 hours a month of work that would otherwise be done by an admin assistant or, worse, by the owner. A capable administrative hire in Dubai costs roughly AED 6,000 to 9,000 a month fully loaded, illustrative market figure, so those 60 hours, about a third of a full-time load, represent something like AED 2,000 to 3,000 of labour.
Against that, a moderately used cloud agent at, say, AED 400 to 700 a month. The agent isn't a person and doesn't replace one for anything requiring warmth or judgement, we're clear about that with every client. But for the hours it does absorb, you're paying somewhere between a fifth and a third of the labour cost, and the gap widens the more you delegate, because the agent's cost grows gently while an employee's cost steps up in whole salaries.
Run the same math against your own numbers before trusting mine. If the agent wouldn't absorb at least 30 or 40 real hours a month in your business, the case is thinner and you should know that going in.
The costs people forget to count
Honesty section. Three line items that don't appear on the compute bill.
Setup is one. Getting from "software exists" to "agent usefully working" involves deployment, configuration, integration, and security hardening, that's our installation service, a one-time project cost, scoped per client. Second, your attention during the first weeks: reviewing drafts and correcting course is unpaid but real, budget a few hours a week early on. Third, model choice discipline. The single most common way cloud bills creep is pointing an expensive model at work a cheaper one handles fine. Part of our configuration is matching models to tasks so the bill stays boring.
For how these figures sit within UAE automation pricing generally, our AI automation pricing guide covers the wider market.
Frequently Asked Questions
What does a Hermes agent cost per month in the UAE?
For cloud deployment, illustrative market figures run roughly AED 150 to 400 a month for light use and AED 500 to 1,500 for heavy use, driven by your server size, model choice, and how much the agent actually works. On-device, you're looking at hardware amortisation, roughly AED 300 to 400 a month over three years, plus electricity. These are market shapes, not INS quotes; we scope real numbers per client.
Does scale-to-zero mean the agent is switched off when idle?
No, and this confuses people. The agent remains available and responds the moment something arrives; what scales to zero is the compute it consumes while nothing is happening. You get always-on behaviour with pay-for-work billing, which is precisely why the model suits businesses with spiky workloads.
Which is cheaper overall, cloud or on-device?
For most SMEs with normal, uneven usage, cloud is cheaper because idle time costs almost nothing. On-device wins when usage is constant and heavy, or when confidentiality rules out cloud regardless of price. If you're at the crossover, roughly where cloud bills would persistently exceed the amortised hardware cost, we'll show you both projections during scoping.
Are there hidden per-message or per-user fees?
Not in the deployment itself. Cloud costs are server time and model usage; on-device costs are hardware and power. For teams, our setup includes access controls with unified billing, so multiple people share one agent and one bill rather than stacking per-seat licences, which is a quiet advantage over most SaaS pricing.
If you want the real version of this post, the one with your usage, your workloads, and actual current prices instead of illustrative ranges, that's a scoping conversation, and it's exactly how our Hermes Agent installation service starts: a deployment assessment that includes the cost projection before you commit to anything. Write to team@ins.ae or call +971 58 995 4553, bring your monthly admin-hours guess, and we'll do the math with you.
