ATPL

AI · Build with AI

Software that plans the task, does it, and reports back.

For teams with repeatable multi-step jobs: chasing documents, reconciling orders, preparing reports. An agent runs the steps, uses your systems, and tells a person what it did.

Describe the problemAll AI capabilities

Still clicking

A job that is still clicking through

Software that plans the task, does it, and reports back.

Describe the problem

  • Clicks

    A job takes forty minutes of clicking between five systems, and it happens two hundred times a month.

  • Exceptions

    The steps are known, but every run has small exceptions that a simple script cannot handle.

  • Trust

    You want automation you can trust, which means you need to see what it decided and why.

How it ships

Watch, then trust

  1. run/01

    Understand

    We watch the job being done by hand and write down every branch, including the ones people say never happen.

  2. run/02

    Design

    The agent’s tools, limits and hand-off points are agreed on paper first; the model is the last thing chosen.

  3. run/03

    Build

    It runs shadow mode against real work for two weeks, with a person doing the job alongside, before it acts alone.

  4. run/04

    Stay

    Runs are reviewed weekly at first, then monthly; model and prompt changes go through the evaluation set.

One job, end to end

A scoped agent for one job

  • Tool connections to your systems

    Read and write access to the ERP, CRM, inbox, spreadsheets or APIs the job needs, each with its own permission.

  • A review surface

    Every run leaves a log a person can read: what it saw, what it did, what it could not finish and why.

  • Guard rails

    Spending caps, rate limits, allowed actions and human approval on anything irreversible.

  • Evaluation set

    Thirty to fifty real past cases the agent is tested against before and after every change.

Primary deliverable

Ships with the engagement

Live brief

01 · Lead outcome

A scoped agent for one job

One workflow, defined end to end, with the tools it may call and the point at which it must stop and ask.

  • 01Tool connections to your systems
  • 02A review surface
  • 03Guard rails
  • 04Evaluation set

One sentence in

A job becomes a run plan

Describe the multi-step job. The brief that comes back is tools, limits and the hand-off — not a chat reply.

What a client typed

Every Monday someone downloads the supplier invoices from the portal, matches them to our purchase orders in the ERP, and emails finance the ones that don’t match.

Brief ready

What it becomes

Problem
Weekly manual invoice-to-PO matching across a supplier portal and the ERP.
Scale
One run a week; roughly 300 invoices a month.
Systems
Supplier portal, ERP purchase orders, finance mailbox.
First step
An agent that matches and drafts the exception email; finance sends it.

Agent or not

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

  • Repeatable jobs with a clear finish line and systems the agent can reach through an API or a browser.
  • Work where a wrong action is reversible or can wait for a person to approve it.
  • Teams that can give thirty real past examples to test against.

Doesn’t fit when

  • One-off tasks. The cost of building the agent is only recovered when the job repeats.
  • Decisions with legal or financial consequences that cannot be reviewed before they take effect.
  • Processes nobody can describe. If the steps are unknown, start with workflow mapping, not an agent.

Your data

The agent runs on your accounts and your data. Prompts, logs and outputs stay in infrastructure you control; model calls go to the provider you choose under its enterprise terms, and nothing is used to train anyone’s model.

How we work

ATPL builds agents on the same engineering discipline as its products: scoped, logged, tested against real cases, supported after launch.

    Chosen for the job

    Python · Node.js · Next.js

    Decide the shape

    Before you start

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

    An AI agent is software that is given a goal and a set of tools, decides the steps, carries them out across your systems, and reports what it did. Unlike a chatbot it acts rather than answers, and unlike a script it handles variation in the input.

    Workflow automation follows fixed rules you write in advance. An agent uses a language model to decide the next step, so it copes with messy inputs like emails, PDFs and portals. Use automation when the rules are stable, an agent when the inputs are not.

    Only within limits you set. ATPL gives each agent the minimum permissions for its job, caps what it can spend or send, logs every action, and routes anything irreversible to a person for approval. It runs in shadow mode first.

    Four to eight weeks for one well-defined job: two to map the process and gather test cases, two to four to build and connect it, two in shadow mode alongside the person who does the job today.

    Continue to