01
Per-field accuracy
Against a hand-validated batch, field by field: tax ID, date, total, line items.
Why the other one is not enough
Against overall accuracy, which hides that the expensive field is the one failing most.
AI for back office and operations
Automating an administrative process is easy until the odd document arrives. We design for the exception from the start: the system recognises what it cannot process, escalates it with context, and that rate gets measured.
2–6 weeks per automated process · 10 years connecting enterprise systems · 3 actionable findings or the diagnostic is not billed
- Document extraction
- Billing
- Reconciliation
- Internal requests
- Exceptions
- ERP
- n8n
Built on
An engineering team that has already been on the other side.
Ninety seconds: who we are, how we work and what you get at the end.
An agent that takes the event, decides and leaves a record
An email, an order or an alert comes in. The agent queries your systems, decides by your rules, asks for approval when needed and executes. Pick a case and watch it run.
Example: a supplier invoice arrives. The system checks it against the order and flags it when the numbers differ.
01Event arrives
Does this invoice match the purchase order?
context
supplier = southern_logistics
02Plans
Read · invoice 9902
Cross-checked · PO-4471
Detected · discrepancy
Queried · policy 88213
Read · clause 4.3
Decided · open a claim
Queried · CRM
Queried · price list
Drafted · proposal
03Queries the systems
Supplier_invoices3,900 PDF
ERP · purchase_orders4,110
Framework_contracts.pdf22 docs
Payments_email2,100
Policy_wording_2026.pdf84 pp
Motor_exclusions.pdf12 pp
Claims_email1,240
ERP · policies38,400
Price_list_2026.xlsx118 rows
CRM · opportunities2,310
Proposals_202556 PDF
Sales_email4,800
3,900 PDF → 3,812 reconciled
04Human approval
Escalated · human
ApproveReject
05Executes and logs
- Read · invoice 9902
- Cross-checked · PO-4471
- Detected · discrepancy
- Escalated · human
- Logged · case opened
- Queried · policy 88213
- Read · clause 4.3
- Decided · open a claim
- Asked for approval · human
- Executed · created the case in the ERP
- Queried · CRM
- Queried · price list
- Drafted · proposal
- Asked for approval · manager
- Executed · sent and logged
log · 5 steps · who and when
Swipe to follow the flow →
The five processes that hold up
01
Extraction of inbound documents
Supplier invoices, purchase orders, delivery notes, receipts. They arrive as PDFs, photos or emails and somebody retypes them.
02
Billing and reconciliation
Issue, match against payment received, flag the difference. Mechanical, repeated and verifiable.
03
Internal requests
Approvals, letters, access, small purchases. High volume, documented answers, handled by email today.
04
Updates between systems
What gets approved in one place has to be recorded in another, and today a person moves it by copying.
05
Recurring reports
The same report every week from three systems. That is collection work, not analysis.
01
The deciding part
The exception is not an edge case: it is the design
An automated administrative flow works on 85% of documents from week one. The project is decided by the remaining 15%: the invoice with a different format, the supplier who changed their layout, the crooked scan, the case the rule did not anticipate. A system that processes those like the normal ones introduces silent errors into accounting systems, and the cost of finding them later far exceeds the saving of having automated them.
So the design starts there: the system has to recognise when it is not sure — low confidence on a field, a total that does not match the line items, an unknown supplier — and escalate with the document and the reason, instead of filling the gap with the most likely value.
And extraction gets measured field by field against a hand-validated batch. The overall average is useless: a system can be 96% accurate overall and be failing on the tax ID or the total, which are exactly the two fields where an error costs money. It is reported per field, and per field you decide what gets automated and what gets reviewed.
How it is measured
The numbers behind an administrative flow
02
Exception detection rate
Of the documents the system should not have processed, how many it actually escalated.
Why the other one is not enough
Against "percentage automated", which goes up precisely when the system stops escalating.
03
Cost of residual error
What happens to what got processed wrong and went undetected, and what it costs to find later.
Why the other one is not enough
It is the number that decides whether the process gets fully automated or keeps review.
04
Cycle time
From document arriving to record landing in the destination system.
Why the other one is not enough
It is the real saving, and it is what gets compared against the cost of running the flow.
When not to automate
If the process changes every month or the judgement depends on who is looking, automating it produces a flow that has to be rebuilt constantly. Stabilise the process first, automate second, and that order is not negotiable.
And if volume is low, the saving does not pay for the build or the maintenance. We say so on the first call with the numbers on the table, even when it means a smaller project or none.
Related services
AI automation
If someone does it by copying and pasting, it can be automated
Applied where there is volume and stable rules, not where there is expectation.
Integrations and APIs
Connecting the systems that do not talk today
MCP servers against the systems already in operation.
AI agents
Agents that do the work, not just talk about it
LangGraph orchestration, durable state and human approval on steps with consequences.
Software development
Custom software development
The system around the model, not just the model.
Key takeaways
Four things before the call
- 01
Administrative automation does not fail on the normal case: it fails on the odd document, and that is where the project either works or creates new work.
- 02
The right metric is not "what share got automated" but what share of exceptions was detected and escalated. A flow that silently mis-processes 3% is worse than one that escalates 15%.
- 03
Document extraction gets measured field by field against a hand-validated batch, not overall. The average hides that the field that matters is the one failing most.
- 04
It pays where there is volume and stable rules. Where judgement changes case by case, automation creates more review than it removes.
This is where the data comes in
An AI system is worth what its sources are worth. These are the standard connectors; anything with an API or a database connects the same way, and what has no API is handled by file.
SAP
Enterprise ERP
Oracle
ERP and database
NetSuite
Cloud ERP
Salesforce
CRM and service
HubSpot
CRM and marketing
PostgreSQL
Database and pgvector
Microsoft SQL
Database
Snowflake
Data warehouse
BigQuery
Google data warehouse
Databricks
Data platform
Redshift
AWS data warehouse
Synapse
Azure data warehouse
Supabase
Managed Postgres
Workday
Payroll and HR
QuickBooks
Accounting
Sage
Accounting and ERP
Xero
Cloud accounting
Shopify
Catalogue and orders
WooCommerce
Catalogue and orders
Magento
Catalogue and orders
Stripe
Payments and subscriptions
Google Drive
Documents and folders
CSV y Excel
Flat files
Nothing in this category
A RAG system in production: three years, two million users
We built and operated the assistant for a loyalty platform serving more than two million active users across ten countries.
We built the full pipeline: document ingestion and normalization, chunking, embedding generation, vector store on Azure AI Search and Pinecone, and retrieval with grounded generation on LangChain.
We held it above 99% availability for three years.
RAG systems →Frequently asked questions
01 What happens with documents the system does not understand?
It escalates them, and that is the design rather than a failure. The system recognises signals that it is not sure — low confidence on a field, a total that does not match the line items, a supplier it has never seen — and sends the document to a person with the reason. The metric we report is what share of exceptions it detected, because a flow that silently mis-processes 3% costs more than one that escalates 15%.
02 Does it work with scans or phone photos?
It does, and the cost has to be stated. Image quality caps everything downstream: a crooked photo with a shadow drops extraction accuracy below acceptable on the numeric fields. We measure batch quality before promising anything and report what share falls below the threshold. Sometimes the right conclusion is to change how documents arrive before automating how they are processed.
03 Does it integrate with our ERP?
Yes, over API, an MCP server or whatever import mechanisms it exposes, against SAP, Dynamics or the system you run. When the ERP does not allow automated writes, the flow goes as far as leaving the record ready to approve, which already removes most of the manual work and keeps control where the team wants it.
04 Is this RPA or AI?
It is the combination, and the difference matters. RPA works when the process is deterministic: if the document always arrives the same way, you do not need a model. AI comes in when the document varies — every supplier invoices differently — and you have to interpret rather than read fixed positions. A good flow uses rules where rules suffice, because they are cheaper and more predictable, and a model only where it is needed.
05 How many people does it take to maintain?
Fewer than run it today, but not zero, and that is worth saying before signing. An administrative flow needs somebody to review the exceptions and flag when a business rule changes. We hand over the flow documented and with dashboards so the internal team can do that follow-up, and maintenance with us is optional and quoted separately.
06 How long does it take and how is it priced?
Two to six weeks per flow depending on complexity and integrations, with something running from week one. It is quoted with fixed scope, price and date in a proposal 48 hours after the first call. We recommend starting with one process, measuring it, and choosing the next with that number in hand rather than committing five at once.
Fifteen minutes. A concrete answer.
It gets fixed, it gets built, or it is not worth it. And if the diagnostic does not reach three actionable findings, it is not charged.