Business process automation with AI agents
We automate business processes in the back office with AI agents, one process at a time: invoice intake, bank reconciliation, the monthly finance pack. Plain rules check every figure, and a person sees only the cases the rules flag.
Our own finance run
Company finance down to a few approvals
Approvals a month, for company finance
A daily check, AI only on new work
Mail check with no AI; the agent runs only on new work
Meeting notes filed minutes after the call
From call to filed meeting note
Rules first, AI where it earns its place
Most steps end up as plain code.
What goes in each lane
- CodeSame answer every time
- IBAN and VAT ID checksums
- Totals, dates and tolerances
- Duplicates, by content hash
- ModelChecked by rules
- Which attachment is the invoice
- Fields from an unfamiliar layout
- The intent behind a free-text email
- PersonIrreversible decisions
- Payments and refunds
- Mismatches the rules flag
- Anything the first two lanes cannot settle
Project stages: map, automate, monitor
The number of weeks comes out of the workshop, once the process is known.
- Map
- Steps, inputs and exceptions, plus the one metric that decides success.
- Automate
- Built on the client's data and tools, tested on past cases first.
- Monitor
- Scheduled runs. Recurring exceptions become new rules.
Company finance, prepared by an agent
Our own finance runs this way. A person approves the payments and presses Send. LifeOS case study
How the finance run is checked
- Every document
- Plain code parses every PDF and checks the IBAN checksum, that the amount in words matches the figures, that the payer belongs to the right company and that the accountant's IBAN has not changed.
- Every statement
- Parsed totals are checked against the bank's own total line, and payments are reconciled against the statements.
- The model's one job
- Judging which mail attachments are company receipts. Everything that must be exact stays in code, covered by tests.
- What a person keeps
- Money and mail, always. The payment file is imported and approved in the bank app, and the pack goes out when a person presses Send. LifeOS case study
Only the exceptions reach a person
Each arrives with the reason and a proposed fix.
19waiting
- Raised
- 36
- Decided
- 17
Back-office processes an agent can run
| Process | What lands on the desk | A person steps in when |
|---|---|---|
| Invoice intake | Checked ERP drafts | Totals differ from the order |
| Bank reconciliation | Matched transactions | A payment matches nothing |
| Monthly finance pack | Payments and a mail draft | Always |
| Report assembly | The weekly report | A figure leaves its usual range |
What an exception carries
- Reason
- The rule that failed, in plain words: the total differs from order 4471 by three units.
- Source
- A link to the exact page and line in the original document.
- Proposal
- The next step the agent proposes, which the approver accepts or edits.
- Channel
- Email, Slack or Teams. The decision flows back into the pipeline and into the log.
What an agent may do alone, and what always waits for a person, follows the guardrails behind every agent. How agent access is limited and prompt injection is handled: cybersecurity and NIS2.
Process automation: common questions
AI agent vs RPA
RPA replays clicks on a screen and breaks when the layout changes. An agent works through APIs and files, reads documents whose layout varies and routes the unclear cases to a person. For a stable screen and a fixed sequence, RPA or a plain script costs less, and the workshop says so.
Agent or plain workflow automation
A plain workflow fits when every step is a fixed rule between systems with ready connectors. An agent is worth it when a step needs reading: free-text emails, documents from many suppliers, or judgement calls that can be checked afterwards. Many pipelines are mostly plain workflow with a single model step.
Email, CRM, ERP, spreadsheets and WhatsApp
The pipeline connects through each system's API or export: mailboxes over IMAP or Microsoft Graph, CRMs and ERPs through their REST APIs, spreadsheets as files or through Google Sheets, WhatsApp through the Business API. Where a system has no API, the fallback is a scheduled export or an emailed report.
n8n or Make instead of a custom build
n8n and Make fit processes where every step is a connector plus a fixed rule. A custom pipeline fits when steps need validation code, tests for the model, or approvals those tools cannot express. The AI workshop compares both for the specific process, and sometimes the answer is n8n.
Measuring the payoff
The baseline is measured before the build: hours per week, error rate or time to first reply, whichever the process owner already tracks. The same metric is measured after launch from the pipeline's own log, so both numbers come from one method.
Clean data or new software first
Usually neither. Mapping shows which inputs are too messy for rules; those go to the model lane or get a small clean-up step. New software is worth it only when the current system has no way in: no API, no export, no inbox.
AI workshop: one process on the table
The process that costs the most hours is the usual start.
- Studio
- Cluj-Napoca, Romania, EU