open back-office-agents.md
Back-office agents · Finance and administration
The agent reconciles, posts and proposes the payment. The signature is still yours
We design and deploy agents that take over the mechanical work in administration: downloading files from portals, reading statements, matching against the ERP and leaving exceptions classified. Connected through secure APIs, they also post transactions, issue invoices and generate payment orders based on rules, with a person who approves.
cat caso-medido.md
What we measured in a real case
A fleet management company. Reconciliations of cards, banks, suppliers and a clearing account for expenses paid on behalf of third parties that was kept in a 48-tab spreadsheet. Run in shadow mode against fourteen months of manual reconciliation.
We also measured the uncomfortable part: about a third of the work is still human review. The agent takes the mechanical work and leaves what requires judgment in order.
agent describe --steps
How an agent works
An agent is a process with defined steps, memory of where it is, tools to talk to other systems and a point where it stops to wait for a person.
- Step 1
Ingestion
Goes and gets the data: warehouse, PDF, portal.
- Step 2
Common language
Translates every source into a single transaction format.
- Step 3
Matching
Cascading rules. Exact amount, to the cent.
- Step 4
Exceptions
Classifies what didn't match and suggests the action.
- Step 5
Report
Excel working paper and a console view.
- Step 6
Approval
Pauses until someone with the right role decides.
Rules first, model behind
Matching is deterministic and auditable. The language model reads PDFs, compares names and writes explanations. It never decides whether something reconciles, and if it fails the run continues.
Three ways to connect
Reading from the data warehouse, APIs where they exist and browser robots when the portal offers nothing else. Repeatable runs: reprocessing a month doesn't duplicate.
Security by design
No passwords in code, encrypted credentials, least privilege, personal data that isn't persisted and models running inside your own cloud.
agent levels --list
From processing to operating: four levels
The architecture is the same across all four. What changes is the last step, which stops being a report and becomes an operation. Each level is enabled after a period in shadow mode.
- A1Reference case
Reads and matches
Read-only. Reconciles, classifies exceptions and delivers the working paper.
Read-only
- A2
+ Proposes
Drafts the journal entry, the invoice or the payment order. A person confirms it in the system.
Draft
- A3
+ Posts with approval
Executes the operation through the API after someone with the right role approved it in the inbox.
API + approval
- A4
+ Executes by rules
Acts without waiting when explicit conditions are met: invoice matching purchase order and receipt, approved supplier, amount under the limit. Everything else goes back to A3.
API + rules
Controls for A3 and A4
- Its own identity per agent and least privilege. The one that invoices doesn't pay.
- Operations that don't duplicate. Every request carries a unique key.
- Limits per operation, per day and per counterparty, enforced on the receiving system's side.
- Rules outside the model. A written, versioned policy approved by a named person.
- Dual sign-off above a threshold, a full trail and a stop button.
cat engagement.md
How we work
Diagnostic
2 to 3 weeks- What we do
- We survey the processes, measure volumes and pick the most boring and most measurable one.
- What you keep
- A process map, data sources and a business case with numbers.
Shadow pilot
6 to 8 weeks- What we do
- We build the first agent and run it in parallel with the manual process.
- What you keep
- A read-only agent in production, measured against the reference.
Operation
Ongoing- What we do
- We add processes and move up levels, one at a time, with their controls.
- What you keep
- Your own platform, in your cloud, documented and operated by your team.
| Phase | Duration | What we do | What you keep |
|---|---|---|---|
| Diagnostic | 2 to 3 weeks | We survey the processes, measure volumes and pick the most boring and most measurable one. | A process map, data sources and a business case with numbers. |
| Shadow pilot | 6 to 8 weeks | We build the first agent and run it in parallel with the manual process. | A read-only agent in production, measured against the reference. |
| Operation | Ongoing | We add processes and move up levels, one at a time, with their controls. | Your own platform, in your cloud, documented and operated by your team. |
open el-agente-no-firma.md
The paper
“The agent doesn't sign” tells the full case in plain language: the architecture, how the agents connect to banks, the ERP and portals, the good numbers and the bad ones, and what it takes for them to also operate.
- The problem wasn't intelligence. A match that worked had been blind for three months because nobody downloaded the file.
- The manual spreadsheet didn't tie to the ledger. Nobody had noticed because there was nothing to compare it against.
- A robot that doesn't retry. If the password fails it stops, because an automatic retry ends in a locked account.
run contact.sh
Which process eats your month-end close?
Tell me which one it is and which systems it runs on today. In a first 30-minute conversation we'll see whether it's a good candidate and what we'd need to measure.
Schedule an Intro Call