Get in touch

AI workflow automation

Most of the hours we find in AI Shadowing go into moving information between systems: reading an email or a PDF, checking it against the ERP, writing the result somewhere else. We automate those steps with a mix of plain code and AI, and keep a person in the loop wherever a mistake would be expensive.

Discuss a workflow
Run #214: receivables, dailyshadow mode
  1. Fetch open items code 0.8 s
  2. Classify items AI 6.2 s
  3. Check data quality AI 3.1 s
  4. Draft 14 emails AI 9.4 s
  5. Approval person Anna Kiss
  6. Send via Outlook simulated not sent

Every step logged: inputs, outputs, tool calls, tokens

Inputs
Emails, PDFs, scanned invoices, spreadsheets, form submissions
Targets
ERP, accounting software, Jira, CRM, shared drives, Outlook
Human in the loop
Approval before anything is sent or posted
Audit
Every step logged with inputs, outputs and model calls
Timeline
Pilot typically in 4–8 weeks

Good candidates

  • Receivables and billing checks. Daily summaries, reminders to the right person, mismatches flagged before month-end.
  • Document intake. Invoices, delivery notes or contracts read, checked against purchase orders or master data, and filed.
  • Recurring reports. The weekly deck someone assembles by hand from three systems every Monday morning.
  • Ticket and email triage. Incoming requests classified, enriched with data from your systems and routed to the right person.

Poor candidates: one-off tasks, processes that change every week, or work where nobody can say what a correct result looks like.

How a workflow is built

Deterministic steps do what code does well. The model only handles the parts that need reading or judgement, and its output is validated before anything moves on.

  1. 1 Trigger A schedule, a new email or file, or a change in the ERP
  2. 2 Code steps Queries, joins, calculations, business rules: no model needed
  3. 3 AI steps Reading documents, drafting messages, classifying; output checked against a schema
  4. 4 Approval A person approves anything that leaves the system
  5. 5 Action Email via Microsoft 365, a Jira comment, an ERP or accounting update

Every run is traced: inputs, outputs, model calls, tokens and the files it opened.

What you get

  • Workflows as code

    Each workflow defined in version-controlled files your team can read, not hidden in a SaaS editor.

  • Integrations

    Connectors to your systems: Microsoft 365, Jira, databases, file shares, and the APIs of your ERP and accounting software.

  • Validation

    Schemas and checks on every model output; anything uncertain goes to a person.

  • Shadow mode

    A trial period in which the workflow runs on live data but only proposes actions, so you can compare it with your team’s own work.

  • Monitoring

    Run history, failures and costs on a dashboard, with alerts when something breaks.

  • Handover

    Documentation and runbooks for your IT team, or ongoing operation by us.

Shadow mode first

Nothing is sent until you trust it.

New workflows start in shadow mode: they run on real data and record the exact email, Jira comment or update they would make, without making it. Once the proposals match what your team would have done, you switch them on.

Questions

Is this RPA?

Partly. Classic RPA clicks through screens and breaks when a screen changes. We integrate through APIs and databases where possible, and use AI only for the steps that involve reading or writing language.

What if the AI makes a mistake?

Every model output is validated, uncertain cases go to a person, and nothing leaves the system without approval until the workflow has proven itself in shadow mode.

Can it work with our old accounting software?

Often, yes. We have connected to a 25-year-old desktop accounting system through its database and REST interface, read-only and with care for the live system.

n8n, Make or custom code?

For simple flows a low-code tool is fine, and we will say so. When you need validation, audit trails, approvals and tests, we write the workflows as code.