ATPL

AI · Build with AI

Take the repetitive steps out of the working day.

For owners and managers whose staff copy data between sheets, portals, WhatsApp and the ERP every day: the steps run on their own, with exceptions sent to a person.

Describe the problemAll AI capabilities

Human glue

When the human is the glue

The repetitive steps between your systems, done by software with a log.

Describe the problem

  • Retype

    Every morning someone downloads yesterday’s orders from three aggregator panels and types them into a sheet.

  • Chase

    Month-end means four days of copying invoices into Tally and chasing the ones that do not match.

  • Blind

    You hired a second person for data entry and the backlog is the same.

How it ships

Map, automate, alert

  1. flow/01

    Understand

    We sit with the person who does the job and record every click and every "except when", including the ones they fix without thinking.

  2. flow/02

    Design

    We decide which steps are rules, which need a model, and where a person must still approve, and agree what a failed run looks like.

  3. flow/03

    Build

    The automation runs alongside the manual process for two weeks and its output is compared line by line before it takes over.

  4. flow/04

    Stay

    We watch run logs and exception rates, and adjust when a portal changes its layout or a supplier changes its invoice format.

Steps that run alone

Automation with an exception queue

  • Automated steps

    Scheduled jobs and triggers that move data between your systems through APIs, exports and, where nothing else exists, controlled browser automation.

  • AI for the messy parts

    Where the input is an email, a PDF or a free-text message, a language model reads it into fields; everything else stays plain code.

  • Exception queue

    Anything the workflow cannot handle lands in a queue with the reason, for a person to resolve in minutes rather than discover in a week.

  • Run log and alerts

    Every run is recorded with what moved, what failed and how long it took; a missed run raises an alert.

Primary deliverable

Ships with the engagement

Live brief

01 · Lead outcome

Process map

Every step, system, file and exception in the workflow, written down from watching it done by hand, with the ones to automate marked.

  • 01Automated steps
  • 02AI for the messy parts
  • 03Exception queue
  • 04Run log and alerts

Map the handoff

A chase becomes a flow

The steps, the exception path and who gets the alert — automation you can see, not a black box.

What a client typed

Every night our team pulls the day’s orders from the two delivery aggregators and our own website, matches them against the POS, and prepares the settlement sheet for accounts, and the numbers never tie out first time because of cancellations and discounts.

Brief ready

What it becomes

Problem
Nightly reconciliation of aggregator and website orders against the POS, by hand.
Scale
Daily; three order sources; cancellations and discounts cause most of the mismatches.
Systems
Two aggregator partner panels, the website order system, the POS, the accounts sheet.
First step
A nightly job that pulls, matches and produces the settlement sheet, with mismatches listed on top for one person to resolve.

Automate or leave

Where it fits, where it doesn’t

Saying when this is the wrong tool is the fastest way to be believed about when it is the right one.

Fits when

  • Work that repeats daily or weekly with the same systems and a known set of exceptions.
  • Steps that move data between systems with APIs, exports or at least a consistent file format.
  • Teams that can name the person who will own the exception queue.

Doesn’t fit when

  • Processes that change every month. Automating a moving target is rework, not savings.
  • Work that is mostly judgement with a little data entry; automating the entry saves little.
  • A process nobody can describe end to end. Map it first; the map alone often removes steps.

Your data

The automation runs on your servers or cloud account with credentials you issue and can revoke. Only the messy steps, such as reading a PDF or an email, touch a language model, and those calls carry one document at a time under terms that exclude training. Run logs stay with you.

How we work

Ethos has been pulling third-party and mobile orders into restaurant billing since 2016; ATPL applies the same approach to office workflows. ATPL will confirm which automation projects may be named.

    Chosen for the job

    Node.js · Python · Laravel

    Decide the shape

    Before you start

    Plain answers on scope, data and what a first proof usually takes.

    Workflow automation is software that carries out the repetitive steps between your systems, such as moving orders from a portal into the ERP, on a schedule or a trigger. RPA is one technique inside it, driving a screen like a person; ATPL prefers APIs and exports and uses screen automation only where nothing else exists.

    The ones that are daily, boring and already well understood: order reconciliation, invoice entry, report preparation, follow-up reminders. ATPL starts with the single step that costs the most hours per week and has the fewest exceptions, proves it, then adds the next.

    Yes, by design. ATPL builds every automation with an exception queue: anything that does not match a rule is set aside with the reason for a person to resolve, and a missed or failed run raises an alert rather than failing silently.

    Usually only for the messy parts. Most steps are plain rules in code, which are cheaper and more predictable; a language model is added only where the input is free text, an email or a scanned document that rules cannot read.

    Continue to