The help centre describes natural-language questions about Spendesk data and states that Spendesk MCP is not an AI model. Beta write tools can also change data; agree permissions and permitted operations.
The provider’s own sources say different things here. mcp.spendesk.com says “Works with Claude Code, ChatGPT, and Dust” (no Langdock) – inconsistent with help centre. MCP site: 68 tools (36 read, 32 write), 75+ skills
Write tools “in Beta”; “some actions cannot be undone”; cannot approve/reject requests or invoices, pay or schedule payments, create or change cards, manage users, read budgets; data cached briefly; no test mode; Action log “does not list the fields or values that were changed”; assistant may misread ambiguous requests, dates or currencies; prompt-injection risk from document text
The provider’s own sources say different things here. MCP site lists “physical card settings” among write tools while help centre says the assistant cannot “create or change cards” – inconsistency. Platform figure (verbatim, provider-stated evidence): “Reduce manual data entry by 80%”
Paid Spendesk account; Account Owner or Admin creates the connection (Settings → Integrations → MCP); only Account Owners and Controllers can use it; user OAuth sign-in; supported assistant account (ChatGPT needs Business/Enterprise/Edu and workspace admin); write permissions enabled per connection and ticked per user
MCP: included free for customers in 2026; Spendesk platform: fixed monthly subscription plus variable per-transaction fees, unlimited users, no active-user fees
AWS region not named. Data retrieved via MCP is processed by the customer's chosen AI provider under the customer's own agreement – EU location of that processing not addressed
Does it train AI on your data?
Only third-party model training is addressed – users should “review that provider's settings for privacy, data retention, workspace sharing and model training”
The provider’s own sources say different things here. The security page assigns PCI certification to partners; the platform page lists it for Spendesk. Confirm entity and scope. Payment authorisation is distinct from certification.
List “applicable as from 31.08.2026”; OpenAI not listed – for MCP the customer's chosen assistant provider processes data under the customer's own agreement
Export
Custom exports “in CSV or XLS”; accounting exports via MCP write tools (preview or commit up to 31 days); DATEV Unternehmen Online direct export
Support “happy to help in French, English, Spanish or German”; help centre in Deutsch; German site and DATEV content; Spendesk GmbH handles support in Germany
MCP site itself in French and English only; assistant language depends on chosen AI provider (not stated)
Approval before posting
No approvals or payments by AI – “The assistant cannot approve or reject requests or invoices, pay an invoice or schedule a payment”; write actions: “Claude asks for your approval before each tool call, unless you set a tool to Always allow”
Confirmation depends on the assistant; Dust may run tools without asking. MCP site: confirm-before-acting is “an instruction, not a lock”. Write tools can mark payables ready for export and commit accounting exports
Audit trail
“Every tool call, including reads, is recorded in the Action log” with date, tool, status, user and entity, plus correlation ID; log “does not list the fields or values that were changed”
Log retention not stated. Platform: “Das System protokolliert jede Freigabe, Ablehnung und Eskalation”
Questions for the provider
Ask the provider for applicable certificates and attestations, including their scope and validity.
Ask the provider which retention and deletion terms apply, including when the contract ends.
Cost per accepted outcome
Editable hypothetical example · not provider pricing or measured performance. Use this monthly worksheet to test your own quote and quality assumptions.
/ accepted outcome
accepted outcomes / month
5,000 tasks × 80% = 4,000 accepted outcomes per month. Higher entered costs raise unit cost; more accepted outcomes spread the same cost over more results.
Formula: entered monthly costs ÷ (monthly tasks × acceptance % ÷ 100). The numerator includes only the costs you enter. Add applicable licence, usage, integration, human review and rework costs yourself; this is not automatically a fully loaded estimate or a vendor billing model. Capacity released is not realized financial value.
Inputs stay in this browser tab’s session, including reload and language changes. They are not added to the page URL. Continuing prepares an unsent enquiry draft.
A written comparison of up to three AI offerings · €195 net
You get the facts that matter for your decision, each with its source, the questions to ask each vendor and a cost estimate based on your own numbers. First you agree your three most important criteria with Sten Franke.
We deliver your brief by email within a maximum of five working days after payment is confirmed and we have agreed the question and minimum information needed to begin. ETHORITY confirms the start date and delivery date and delivers earlier when the reviewed brief is ready. Missing optional information is recorded and assessed; it does not require you to research the solution yourself. Review before delivery and email support are included.
ETHORITY®