We build the tool that does it. One fixed price, agreed before we start, from 1,500 euro. Running on your data in 5 working days. You pay the second half when it passes your tests.
Send us the problem or email jayron@geurs.com Finance team? Start with the day-5 hold list
02 · Checks
01 Live demo
A governed operations agent for delivery orders between apps and kitchens. It finds stuck orders and menus out of sync, fixes what a written rulebook allows, and asks a person for the rest.
Not a restaurant? See a day-5 hold list for a purchase ledger
Synthetic data, labelled as such.
02 Speed and review
Our engineering team took the delivery agent from an empty folder to a public agent you can click in one night. We build with AI coding agents. Nothing ships without review by our engineers, and they read every line that touches your data before it reaches you.
03 7 days · 249 planted incidents
Every action it takes is logged with the evidence and an undo. It grades itself against ground truth, per detector, per day. The numbers below come from 7 replayed days it was never tuned on, in the Gulf world: 95% handled without a person, 5% handed to a person with the evidence attached. The weakest detector is named on the accuracy page.
Numbers from the delivery demo on synthetic data. Your Proof is graded by the tests you write with us.
03 · Day 5Duplicate invoice check
The duplicate check asks one question of every invoice before it is paid: have we seen this one before? 20 sample invoices, 4 held: 3 confirmed duplicates, and 1 false alarm released by a person.
| Row | Supplier | Invoice number | Status | Evidence | Amount (GBP) |
|---|---|---|---|---|---|
| 10 | Waste supplier K | WC-1187 | held: same amount as last month, released by a person | Same supplier and same amount as the earlier invoice WC-1150 (not on this page), inside the date window in your rules. The service month on the invoice is different. | £395.00 |
| 12 | Packaging supplier C | PK-7731 | held: likely duplicate, hold approved by a person | Same supplier and same amount as PK-7714 on row 3, two days apart. Only the invoice number is new. | £2,368.80 |
| 15 | Chilled foods supplier H | CH-55098 | held: likely duplicate, hold approved by a person | Credit note CN-2210 already cancelled this invoice, and its replacement CH-55131 is row 8. Paying it pays for the same goods twice. | £1,184.40 |
| 19 | Dry goods supplier E | DG 10482 | held: likely duplicate, hold approved by a person | Same invoice number and amount as DG-10482 on row 5, sent again with the date changed from Monday to Friday. A space in place of the dash does not hide it. | £3,912.15 |
Nothing on the hold list is paid until a person decides.
A hold list with no false alarms would mean the check is too loose. Row 10 is the false alarm here: the person presses Release, because it is the monthly fee for a different month.
The Proof reads and reports, it writes nothing. Writes come in the Pilot, through your own APIs, each one logged with an undo.
The Proof ends with the amounts found, in your currency: the duplicates already paid and the invoices on the hold list. The day-5 result states them in pounds for a UK ledger, or euro or dollars. So the Pilot decision has a number.
The full hold list, all 20 rows
Every row is synthetic sample data.
04 · Work
Our operations agent for delivery orders (Canon Ops): detectors, a controller with authority tiers, a ledger, self-measured accuracy and an architecture document.
See it runA weekly hours check for a staffing back office: planned, client-confirmed and payroll hours side by side, only the lines that differ on a short list, and each draft invoice checked against the agreed rate, on synthetic data.
A merchant screening service for MultiSafepay's risk team: from a merchant's website address it returns what the business sells, how it takes payment, a category code, reputation checks and the red flags it found, stored with the raw answer for staff to decide.
This site, with its own experiment logging: which version you saw, which link you came from, no cookies.
Our research and outreach engine: targets in, personalised notes out, one approval file, logged sends and follow-ups.
Our venue map and lead score for a city: 4,520 food and drink venues, scored per buyer.
A venue database for a sales and onboarding team, extended and hardened: venues collected from several sources with duplicates merged, the source kept for every field, fill rules checked blind against held-back cases, and a nightly self-test that wrote nothing when it missed too many known cases.
Lead ranking for a sales team inside its own CRM: each matched enquiry got a score from 1 to 10 with a short reason, thin leads stayed blank, and the scoring was checked on held-out records and against shuffled answers.
A visit planner for a field sales team: a starting point and a time window became a short list of venues known to be open, with one map route, and each visit outcome showed on that venue's row for the whole team.
One login for an IT team's internal tools: personal accounts, tools shown by role, each tool checking access itself, and a private page showing who used which tool and how old each data feed was.
Call notes for sales reps: the key moments of each call written onto the matching CRM deal with a timestamp and a quote, and plain code dropped any quote it could not find in the transcript.
A channel-shift analysis for a sales lead: merchants ranked by their slide toward the delivery marketplaces after removing season, market drift and data-feed changes, with gainers printed beside losers so the list was never read as proof.
Direct-debit collection for a finance team: recorded consent, a sign-up payment that became the mandate record and an announce step before every debit, run end to end against the payment provider's test environment.
An audit of the AI already inside the software a sales and support team paid for: built-in features nobody had switched on, tested on sandbox copies before they touched live data, with a one-page guide per team.
A download log for a support team's flyer generator: every flyer saved with its customer, offer and discount code in one running history per customer, so any offer could be traced later.
05 · Write-ups
Give an AI agent authority in your operations without losing control
Reconcile bank and Stripe payouts with your ledger automatically
Get the data out of PDF invoices and into your accounting system
06 · PricesWhat it costs
from 1,500 euro
5 working days
One workflow on your real data, fixed scope, 3 to 5 pass/fail tests written on day 1. Credited in full against a Pilot booked within 30 days.
Not included: more than one workflow, API access to the systems we read from their exports, such as Sage Intacct, NetSuite or SAP.
from 6,000 euro
30 days
The workflow live in your systems, measured against a baseline you agree first.
Not included: new workflows.
from 20,000 euro
3 to 6 weeks
Several workflows or a governed agent with a rulebook, a ledger and its own accuracy measurement.
Not included: ongoing operation (see Care).
from 500 euro a month per live system
Monthly
Monitoring, fixes, small changes, a monthly accuracy report. Hosting in the EU or UK included when you have no cloud team.
Not included: new workflows.
3,000 to 5,000 euro a month
Monthly
Your build requests in a queue you order, delivered one after another. Pause any time.
Not included: several requests built side by side (see Build).
Prices shown in euro, excluding VAT. Your quote comes in euro, pounds or dollars, fixed for the project. Fixed means fixed: the number you get on day 1 is the number on the invoice.
07 · HowDay 0 to Day 5
Half up front. The other half when it passes the tests you wrote with us. If it fails and we can't fix it within 5 working days, the first half comes back.
09 · Data and terms
Either way, nobody on our side reads your raw rows.
10
stackforagents, based in Dubai: a company specialising in AI agents and back-office automation. Our engineering team builds, and the delivery team works with you from day 1 to day 5. Nothing ships without review by our engineers, and they read every line that touches your data before it reaches you.
You own the delivered system. We keep the reusable parts we bring in, and we tell you which those are on day 1. We sign a mutual NDA before you describe your systems. The runbook, the tests and the code let your team, or anyone else, run it without us.
Care covers it: monitoring, fixes, small changes, a monthly accuracy report. Without Care, you get the code and the runbook.
During the Proof the raw rows stay in your systems. We work through read access you grant and can remove, and on day 1 you choose how we see your data. Scrubbed export: our engineers write the masking script, and your IT or your MSP runs it, or we run it with you on a call. It masks names, bank details and personal fields before anything leaves your systems, and the Proof runs on that. Inside your environment: the tool runs in your cloud account or on your server, and we see logs, test results and masked samples, never the raw rows. With Care hosting, when you have no cloud team, the tool and its data sit on a server we run for you in the EU or UK region. More on Data and terms.
Only when a step needs one, and then only the fields that step needs. Finance work runs on written rules: supplier, invoice number, date, amount, purchase order, line, quantity and price. No model sees your data unless a step reads PDFs. Most decisions run on written rules; in the delivery demo's Friday evening, 95 decisions came from rules and 50 from a model. The model could only propose: the agent runs what the written rulebook allows and asks a person for the rest. You choose the model provider and region, or a model that runs inside your environment. With a scrubbed export, a model only sees what the masking script left in: no names, no bank details, no personal fields. A data processing agreement is part of the written agreement on day 1; you get its full text, with a mutual NDA, before you describe your systems. More on Data and terms.
Read access first: a database, an export or an API you grant and can remove. The tool reads it inside your environment, or works on the scrubbed export, as you chose on day 1. The Proof takes 5 working days and runs in shadow mode: it reads and reports, it writes nothing. Writes come in the Pilot, through your own APIs, each one logged with an undo.
CSV and Excel exports from any system, PDFs, Google Sheets, and databases: Postgres, MySQL and SQL Server. Through their APIs we read Xero, QuickBooks Online, Sage Business Cloud, Exact Online, Stripe, Mollie and Adyen. We read Sage 50, Sage 200 and Dynamics 365 Business Central from their exports. Where a system's API needs a vendor partnership, as with Sage Intacct, NetSuite and SAP, we start from its exports.
It runs in your cloud account or on your server; you hold the keys. With Care we watch it, fix it and make small changes. No cloud team? Care includes hosting on a server we run for you in the EU or UK region; you still own the code and can take it in‑house any time. Without Care you get the code, the tests and the runbook, and your team runs it.
Our team works from Dubai. Replies within 4 hours, 08:00 to 22:00 Dubai time, seven days a week. A message sent outside those hours is answered the next morning.
New fixed price in writing before any extra work starts.
Our engineers use AI coding agents to write and test code and to run research. Nothing ships without review by our engineers, and they read every line that touches your data before it reaches you. Every system measures its own accuracy and shows it. The demo grades itself against faults we planted in synthetic data. Your Proof is graded by the 3 to 5 tests you write with us on day 1, on your data, and you keep the test results.
Send us the job your team hates.
or email jayron@geurs.com
Replies within 4 hours, 08:00 to 22:00 Dubai time, seven days a week.