For the optimal experience, view this page on a larger tablet, laptop or desktop device.
A transparent account of how Loupe prices AI: no token markup, a flat charge for the harness that does the work, and a brokering layer engineered to drive your token cost toward its floor.
Every other "AI" vendor in this space buys tokens wholesale and resells them to you at a markup. That model quietly punishes you for using the product, and it puts the vendor's incentive in direct conflict with yours: they profit when your consumption rises.
Loupe takes the opposite position. We pass model tokens through at cost — zero markup. We charge only for the harness: the orchestration, governance, memory, and tooling that turn a raw language model into a jeweler that can actually quote, cost, transfer stock, and close the books correctly.
The fee charged by the underlying AI model. A raw commodity, bought at the same published rate available to anyone. We add nothing to it.
Billed at $0.0025 per tool run. A workflow that touches one tool costs one unit; a workflow that chains seventy-five touches seventy-five. You pay for work performed, nothing more.
A language model on its own cannot run a jewelry business. It cannot see your inventory, respect an approval gate, remember a session, or be trusted with a price. Seven systems around it — the harness — are what make it safe, accurate, and useful. This is what the $0.0025 buys.
Eighteen representative workflows, priced end-to-end. Token is the pass-through model cost; harness is $0.0025 per tool run; total is what one full run costs. The complexity tag reflects how much work each workflow chains — that is what drives cost.
The price list is an honest indication, not a fixed tariff. Three variables move the real figure, and you should understand each before you model your spend.
The same workflow varies slightly each run — model responses differ in length and a tool may fire more than once. On a typical simple workflow of around $0.08, we observe a standard deviation of roughly ±$0.02. Budget the mean; expect the spread.
The ±$0.02 figure describes the simple band. It does not describe a 75-run build like new-warehouse onboarding, which is an order of magnitude larger and far less frequent. Read each band on its own terms.
Token cost depends on which model the workflow requires. As frontier prices fall and our broker routes more aggressively, the token column trends down over time — the harness column stays flat.
All figures are in USD. Token costs are set by model providers in USD; invoiced amounts are converted to your billing currency at the prevailing rate on the date of quote.
You already employ skilled staff who do this work by hand — and you keep costs lean by throwing capable people at problems. The fair question isn't "AI vs. nothing." It's "what does one workflow cost as a Loupe run, versus the same task done by a person?" Here is that comparison, honestly framed.
The point is not to remove your people — it is to redeploy them: payroll should back rainmakers, not paperwork. Every workflow Loupe absorbs is an hour your team spends selling, sourcing, and serving clients instead of fighting a spreadsheet. The cost comparison is decisive; the strategic one is more so.
The comparison above uses round figures. This does the arithmetic properly. It models the fully-loaded annual cost of a mid-level data-entry clerk — salary, statutory months, employer contributions, management overhead, attrition, and the cost of human error — against the same volume of records run through Loupe. Adjust any assumption to match your operation; every number is a default you can move.
| Location | Human (fully loaded / yr) | Loupe / yr | Annual saving |
|---|
We did not pull "a few dollars an hour" out of the air. Every figure in the calculator above is built from a two-stage, country-specific waterfall — first the true cost of employing the person in their own jurisdiction, then the operational cost of running human data entry on top. Everything resolves to USD, then converts cosmetically to your chosen display currency.
True annual cost = employment + management + attrition + error. Per day / hour / minute are simply that total divided by the working-time basis.
Base ₫130.0M → +1 month → ₫140.8M cash → +21.5% contributions (₫30.3M) → ₫171.1M ÷ 26,000 = $6,581 employment cost. Then +15% management ($987) + 5.8% attrition ($384) + error cost ($4,091, from 460 errored records/yr at ~$8.89 each) = $12,043 true annual cost — i.e. $52.36/day, $6.55/hour, $0.109/minute.
Per-location levers (country-correct): monthly salary, FX, employer-contribution %, contribution cap, extra months, benefits. Global defaults (shared): 230 working days, 8 hrs/day, 200 records/day, 1% error rate, 15 min rework, $8 consequence, 15% overhead, 35% turnover, 2-month replacement. Salary data is reconciled from Glassdoor / Indeed / ERI and local boards; error economics rest on the 1% outsourcing-QA benchmark and the 1-10-100 rule.
Loupe routes every workflow to the lowest-cost model that still clears the bar for that task — frontier intelligence only where the work demands it, a smaller model everywhere it doesn't. You get the right answer at the floor price, without ever managing a model yourself. Weigh that against the two alternatives:
| Your option | What it really costs | Verdict |
|---|---|---|
| Subscribe to a single AI provider (e.g. OpenAI) and try to wire it up with your ERP system | You pay top-tier token rates on every task — including the trivial ones a cheap model handles fine — and you still have no jewelry harness around it. You're buying the most expensive intelligence and using it as a calculator. | Overpays and under-performs, no liability cover on AI |
| Build your own AI-native ERP | Years of engineering, a standing team to maintain seven harness systems, evals, and model routing — rebuilt every time the frontier moves. A multi-million-dollar permanent cost centre to replicate what you can rent today. | Huge sunk costs at risk & likely under-performant, no liability cover on AI |
| Use Loupe as your Operating System and broker layer for intelligence | No markup on tokens, a flat harness fee, and routing that drives your per-task cost toward its minimum automatically — maintained by us, improving without your effort. | Lowest cost & highest performance, contractual liability cover on AI where it counts |
Gold and diamonds are commodities; everyone pays roughly the same for the goods. Overhead is where the margin is won — and AI workflow cost is now part of that overhead. Loupe is engineered so that line item only ever gets smaller relative to the work it does.
We run only one or two implementations at a time, done properly by our senior leadership and accelerated by AI — we re-platform your ERP in a fraction of the time a migration used to take. If the numbers in this briefing make sense for your operation, the next step is a deposit to hold your place. Implementation bandwidth is our hardest constraint, the queue is growing, and first movers go first.