Service · AI and ML engineering

Workflow automation

Take the repetitive back-office work off your team, one flow at a time, with a person where it matters.

Relevant to Manufacturing, Service businesses, Education, Growing businesses

The problem

Invoices are keyed in twice. Enquiries wait in an inbox until someone forwards them. Approvals travel on WhatsApp. Each step is small, and together they are three people's day.

Most of this is rules, not intelligence: if this, then that. Where the input is messy, an email, a scanned form, a photo, a model reads it and the rules take over.

What it costs today

Typical: two to five people's time per department spent moving information between systems, errors from retyping, and approvals that take days because nobody can see the queue.

This is for you if
  • A flow you can draw on a whiteboard, that runs tens of times a day or more.
  • Inputs arrive as email, WhatsApp, PDF, spreadsheet or a form, and end up typed into a system.
  • Someone owns the process and wants it fixed, not only measured.
What we build
  • Automated flows for intake, extraction, validation, approval and entry, with a person's approval where the step matters.
  • Models where the input is messy: reading documents, classifying emails, matching records. Plain rules everywhere else.
  • A queue your team can see: what is waiting, what was done, what was rejected and why.
  • Connectors to your ERP, accounting, CRM and messaging, reading and writing only where you permit it.
How we build itIn this order
  1. 01

    Map the flow with the people doing it

    Every handoff, every exception, every WhatsApp approval. The exceptions are the work, and they only show up when you sit with the team.

  2. 02

    Automate the boring 80 percent first

    Fixed-format inputs and rule-based steps go live in weeks. Judgement steps stay with people until the data says they can move.

  3. 03

    Add models at the messy edges

    Document extraction, email classification, record matching. Each one has its own test set and its own accuracy number, published.

  4. 04

    Keep a person in the loop

    Approvals, low-confidence cases and anything that touches money go to a queue. The queue shrinks as confidence is earned, not before.

  5. 05

    Measure hours and errors

    Hours per week, errors per hundred items, turnaround time. Before and after, same period, in the pilot report.

Data you already haveThis is usually enough to start
  • A sample of the inputs: emails, PDFs, forms, messages
  • The systems where the output lands
  • The SOP, or the person who knows it
  • Volumes per day and the current turnaround
In practice

Illustrative example

Customer orders into the ERP without retyping, for a freight forwarder

Booking requests arrive as emails, spreadsheets and PDFs in a dozen customer formats. Three people retype them into the ERP, then forward confirmations by hand. Mistakes show up as wrong consignee details on the documents.

Read the case study
Typical engagement
Duration2-week assessment, then 6 to 10 week build on one flow
TeamSoftware engineer, ML engineer where documents or text are involved, your process owner
You receiveOne flow live end to end, measured hours and errors, queue in use by the team
PriceFixed price, quoted after the readiness assessment
What you own at the end
  • Running flows on your accounts, with their connectors, in your repository.
  • The queue and the dashboard.
  • Documentation your IT team can follow to change a rule, a field or a recipient.

Everything runs in your environment and is handed over with documentation. Nothing stops working if we part ways.

Questions we are usually asked

We tried RPA and it kept breaking.

Screen automation breaks when a screen changes. We work through APIs, database views and exports wherever they exist, and use screen automation only as a last resort, with monitoring.

Will this remove jobs?

It removes retyping. Every client so far has moved people onto work that was waiting for them. We do not promise headcount reductions and we do not build business cases on them.

Do you use no-code tools?

Where they fit and your team can maintain them, yes. Where the flow is critical, we write it as code in your repository with tests. We tell you which, and why.

Can it run on our own server?

Yes. Most flows run on a small server or your cloud account, and nothing has to leave your infrastructure.

When does this become an agent?

When the steps vary case by case and a judgement is needed between them. Then we build an agent with a harness. Most back-office flows never need to.

Often combined with

Further reading

Not sure if your data is ready?

Take the free 20-minute readiness check. You get a short scorecard and a plain answer about what, if anything, is worth doing next.

Start the readiness checkTalk to us