Services
Solutions
Use cases
Industries
Tools
Pricing
Book a call
Automation

AI process automation

Automate existing workflows with — from receipt capture to quoting to email routing.

Many companies have digitised their processes — but not automated them. Data is moved between systems by hand, emails sorted by hand, documents checked by hand. That isn't digitalisation, that's just PDF instead of paper. combined with engines can change that.

Existing systems become automated workflows via interfaces: receipts are captured by and posted, quotes generated from requests, customer requests classified and routed, onboarding steps triggered. The data stays inside your own infrastructure.

Which tools we use depends on the case — as a visual engine for most SMB scenarios, for complex business processes with humans in the loop, classic APIs and for integration. The right tool per , not one tool for everything.

Process5 steps at a glance

  1. Process discovery

    Frequency, regularity, manual time, data quality, interfaces — list with cost/benefit estimate.

  2. design

    Tool choice (, , ), trigger strategy (, , manual), data flows, usage points.

  3. Integration

    , , database, file watcher, . Secrets encrypted, if needed.

  4. Test & pilot

    Happy path, edge cases, idempotency, retry, dead-letter queue, parallel pilot phase.

  5. Go-live & monitoring

    Dashboard, alerting, audit logs, health checks, cost tracking, continuous iteration.

From manual flow to automated workflow

Triggers

From a prioritised process comes a technical concept. Which tool, which trigger, which steps, what happens on errors — everything is thought through in advance, not decided while building.

Tool choice depending on the case:

  • n8n — visual workflow engine, ideal for 80% of SMB processes: API integrations, data transformation, AI calls, simple business rules. Self-hosted and open source
  • Camunda — when the process gets more complex: multi-stage approvals, escalations, service levels with time limits, human in the loop. Industry standard with BPMN 2.0
  • Direct API integration — when no workflow engine is needed, just a thin Lambda function or a cron-triggered job

Trigger strategy per process:

  • Cron — time-driven (every morning at 8:00, every 15 minutes, Mondays at 9:00). For reports, backups, token refreshes
  • Webhook — event-driven, the source system actively calls the workflow. For orders, form submissions, payments
  • Manual — a person triggers the workflow with a button. For rare or sensitive cases that require a human decision
  • Polling — when the source system has no webhooks, the workflow asks regularly. Weakest option, but sometimes the only one

Data flows and AI usage inside the workflow:

  • Where is data read, where written, where transformed?
  • At which points does a language model decide — and where a deterministic rule node?
  • Which fields have to be extracted (e.g. from PDF receipts via OCR)?
  • Which validation steps prevent garbage data from flowing on?
Read more

Automation is only as good as the integration. A workflow that isn't connected to real systems is a nice demo — and nothing more.

Typical integration paths:

  • REST / API interfaces — the standard. CRM, ERP, accounting, mail server, cloud storage usually have APIs
  • Webhooks — source system actively notifies the workflow. Faster, leaner, less error-prone than polling
  • Database connection — direct read/write to PostgreSQL, MySQL, MS SQL. For bulk data when APIs are too slow
  • Email / IMAP — incoming emails as a trigger (e.g. applications, orders, complaints)
  • File watcher — new files in cloud storage or network drives trigger workflows (e.g. receipts that land in the inbox folder)
  • RPA — last resort when a system has no API. A click bot in the UI. Works, but brittle

AI in the integration:

  • OCR for receipt capture — a scanned receipt becomes structured fields (date, amount, VAT, supplier)
  • Classification of incoming emails — complaint, request, order, spam
  • Data extraction from unstructured sources — contract details from PDFs, address data from email signatures
  • Enrichment — incoming data is combined with master data from the CRM before it moves on

The integration is consistently secured with secrets management: API keys are stored encrypted, rotated, never visible in the workflow code. In sensitive industries the entire stack runs self-hosted.

Read more
Alerting

Go-live starts the production phase — and at the same time the most important one: will the workflow still work in four weeks when no one is watching? That's what observability is for.

What ships with it:

  • Monitoring dashboard — how many cases run through per day? How many fail? What's the response time?
  • Error alerting — Telegram, Slack or email for critical errors, with a link to the affected case
  • Audit logs — who triggered what when, which data was moved? Mandatory for compliance, helps with debugging
  • Health checks — workflows ping themselves regularly, reachability of source systems is monitored
  • Cost tracking — token consumption on AI calls, API calls to third parties, compute time

Iteration and extension:

  • New edge cases emerge from pilot logs — they flow back into the workflow
  • New use cases in the same domain (e.g. a second receipt type, other suppliers) are added incrementally
  • If an AI model is replaced by a newer one, it happens under control — A/B comparison, then switchover
  • Step-by-step scaling — what works at one site is rolled out to further sites

Over time, the first automated process becomes a collection of reusable building blocks — auth flows, data enrichment steps, error handling. The second and third automation then go significantly faster than the first.

Read more
IdempotencyDead-letter queue

Sounds interesting?

Let's talk it through in a free intro call and see how this would work for you.

Book a callAll services
Book a callBAFA funding