Agent Runs
Rent an AI agent swarm. Get the finished work.
Fixed-price agent runs on your business data — supplier files, invoices, KYC files, ERP migrations, or 20,000 contracts at once. EU-hosted. Evidence for every change.
Swarm Console
Replay of a demo run on synthetic data.
Supplier file (2,400 suppliers): 24 agents processed 2,400 records in 2h 00m of run time, found 263 issues, and supervisors verified all 160 batches. This is a replay of a demo run on synthetic data.
- being processed
- verified
- Invalid VAT number—
- Duplicate supplier—
- No Peppol ID—
- IBAN country mismatch—
- IBAN shared by several suppliers—
Agents in this run
Lead Supervisor / Editor Worker
- 24
- agents
- —
- queued
- —
- running
- 160
- verified
- 2,400
- records processed
- 263
- issues found
- 12.4M
- tokens
- 2h 00m
- run time
What a run returns
A one-page report, a corrected file, and the evidence behind every finding. This sample is for a supplier-file audit.
Sample report on synthetic data.
E-invoicing readiness: 62/100
Invalid VAT numbers71 of 2,400
Format errors and numbers not found in VIES.
VIES response and the stored value, per record.
Duplicate clusters38 clusters (84 records)
Same VAT or same normalised name.
The records side by side, with the matching fields.
IBAN red flags9 accounts
Shared accounts and country mismatches.
Which records share the account, and when it was changed.
Peppol-ready41%
Suppliers with a Peppol participant found.
The identifier looked up and the lookup result.
Your counterparties by e-invoicing deadline
Belgium
18%
In force since 1 Jan 2026
France
22%
Receiving since 1 Sep 2026
Germany
9%
Issuing from 2027/2028
Luxembourg
51%
2028/2029 — bill pending
Four runs to start with
Each starts with a Diagnostic: we measure the problem on your data, then quote the fix.
Supplier and customer file audit
- The problem
- E-invoicing needs clean master data: valid VAT numbers, no duplicates, known Peppol IDs.
- What the run does
- Checks every VAT number in VIES, finds duplicates, flags bank-account risks and looks up Peppol IDs.
- You get
- A readiness score, a corrected file and a fix list with evidence.
Timeline: 1–2 weeksDiagnostic from €2,500
See this run →Invoice intake into your ERP
- The problem
- Supplier invoices arrive by email and are keyed by hand into Odoo, Sage BOB50 or Business Central.
- What the run does
- Reads each invoice, matches it to the purchase order, and drafts the entry for a person to approve.
- You get
- Draft entries in your ERP and a list of exceptions to review.
Timeline: 2–4 weeksDiagnostic from €2,500
See this run →KYC file remediation
- The problem
- Periodic reviews pile up: expired IDs, missing beneficial owners, overdue files.
- What the run does
- Reads each client file, lists what is missing or out of date, and prepares the requests.
- You get
- A prioritised remediation list per client. Your compliance officer decides.
Timeline: 2–3 weeksDiagnostic from €2,500
See this run →ERP migration data preparation
- The problem
- Old data breaks new ERPs: duplicates, orphan records, broken characters, invalid dates.
- What the run does
- Maps the old tables to the new model, cleans and deduplicates, and tests the import.
- You get
- Import-ready files and a mapping document, for you or your integrator.
Timeline: 2–6 weeksDiagnostic from €2,500
See this run →
How runs are priced
Diagnostic Run
from €2,500
An evidence-based report on one data domain.
Build Run
€8,000–€24,000
The fix itself. Up to 70% fundable through the SME Package AI.
Recurring Run
from €400/month
The same checks, every month, on new data.
Flagship Run
from €20,000
Tens of thousands of documents or records in one campaign.
Check 500 suppliers free — nothing leaves your browser except VAT numbers.
The free Supplier File Health Check scores your file for e-invoicing readiness: VAT numbers in VIES, duplicates, bank-account red flags and deadlines by country.
Run the free Health CheckSwarm Scale Calculator
Choose a job and a volume to see how a run would be sized.
Each square is 13 agent tasks; the swarm works on 4 at a time.
- 3,000
- Agent tasks
- 4
- Peak parallel agents
- 16 h
- Estimated run time
- $91
- Estimated compute (list prices)
- 167 h
- Equivalent human work
Price band
€4,000 – €6,500
Hours of work, side by side
Estimates. Compute is typically under 15% of the price — you pay for design, QA and accountability. Assumptions are listed in the code behind this calculator.
Swarm Lab
Swarm Lab: work that sounds impossible
Every contract you have ever signed, read and risk-registered. Your whole supplier base screened in 48 hours. Flagship runs start at €20,000.
Open the Swarm LabHow we keep your data safe
Three ways to give a run access. Pick the least access that does the job.
The run sees
A file you export, with names and personal fields masked if you choose.
We never see
Your systems, your passwords, anything you did not export.
- EU-hosted models: Claude through Amazon Bedrock in EU regions.
- No stored passwords.
- A data processing agreement (GDPR Art. 28) from day one.
- A run log with every delivery.
- A person signs off every risk decision.
- Code and rights are handed over to you.
We build on top of your systems; we don’t take over your infrastructure.
Masked export: The run sees — A file you export, with names and personal fields masked if you choose. We never see — Your systems, your passwords, anything you did not export.
Runs inside your cloud: The run sees — Nothing leaves your cloud account. The run executes there and writes results there. We never see — A copy of your data on our side.
Short-lived read-only link: The run sees — Read-only access to the named tables, for the length of the run. We never see — Write access, admin rights, or access after the run ends.
The e-invoicing clock
Structured e-invoicing is arriving country by country. Your supplier and customer files need to be ready before your counterparties’ deadlines.
461
days
Until Luxembourg businesses in scope must be able to receive e-invoices, as drafted in bill 8815
Bill pending — dates may change
BelgiumIn force
Since 1 January 2026, VAT-registered businesses exchange structured e-invoices via Peppol. Fines: €1,500, then €3,000, then €5,000. Source
FranceIn force
Since 1 September 2026 all companies must be able to receive e-invoices; large and mid-size companies issue them. SMEs issue from 1 September 2027. Source
GermanyIn force
Receiving since 1 January 2025. Issuing from 1 January 2027 above €800,000 turnover, and for all businesses from 1 January 2028. Source
LuxembourgBill 8815 pending
As drafted: receiving from 1 January 2028; issuing from 1 July 2028 for larger businesses and 1 January 2029 for the rest. Source
Today more than 1,400 private-sector bodies in Luxembourg receive e-invoices via Peppol (European Commission). Source
Questions about Agent Runs
What is an AI agent swarm?
A group of AI agents working on one job at the same time. A lead agent splits the work into tasks, worker agents do them in parallel, supervisor agents check each result, and an editor agent assembles the deliverable. In our runs, at most 50 agents work at once.
Is my data safe?
You choose the access: a masked export, a run inside your own cloud, or a short-lived read-only link. Models run in EU regions (Claude through Amazon Bedrock), we sign a GDPR data processing agreement before the run, and we never store your passwords.
Do you need our passwords?
No. We work from an export you prepare, or inside your own cloud account, or with a temporary read-only link that expires at the end of the run.
What does it cost?
A Diagnostic Run starts at €2,500. A Build Run, which fixes what the Diagnostic found, costs €8,000 to €24,000. Recurring runs start at €400 a month and flagship campaigns at €20,000. Every run has a fixed price agreed before it starts.
Can the SME Package AI fund it?
Build Runs are sized to fit the SME Packages – AI scheme, which can cover up to 70% of eligible costs, with the grant capped at €17,500. Eligibility is decided by the Ministry of the Economy; we check it with you before the run.
How accurate is it?
We measure it on your data. Every run includes a sample checked by a person, and the report states the error rate found in that sample. We do not quote accuracy figures from other people’s projects as if they were ours.
What if the swarm is wrong?
Nothing is changed in your systems without approval. Supervisor agents re-check each result, a person reviews every risk decision, and the evidence log shows why each finding was made, so a wrong one is easy to spot and reverse.
How is this different from buying software?
Software gives you a tool to run yourself. An Agent Run gives you the finished result, for example a cleaned supplier file or a contract register, plus the code and rights to repeat it.
How long does a run take?
Most Diagnostic Runs finish within one to two weeks, including the scoping call and your review. The agent work itself usually takes hours, not days.
Scope your first run
Twenty minutes: tell us which data and which problem, and we tell you what a Diagnostic Run would measure and cost.