Design preview. Product screens are illustrations, not screenshots, and the built-in flows shown are proposed, not yet confirmed.

Repeatable AI starts with the process.

Map the work once. Decide what the AI does alone and what a person checks. Your AI follows it the same way every time.

Why AI in a chat window is not repeatable.

Ask an AI the same thing twice and you get two different results. Ask two colleagues to do it and you get two more.

The method lives in whoever is typing, so nothing can be reused and nobody can see how the answer was reached.

One request, four routes, four results: you on Monday, you on Tuesday, a colleague, and another colleague.

AI gives consistent results when the process is mapped first: the steps, the order, which steps the AI does alone, which a person reviews, and which need approval. Most teams know their work well. Turning it into something an AI can follow is the hard part.

Abstraction is where you do that.

Three ideas

Map the process first

Describe how the work gets done. Abstraction drafts it as a flow of steps you can see, check and improve.

A person where it matters

Mark each step as done by AI, by AI with a person checking, or by a person. Nothing runs until someone approves it.

Your own AI

Use the AI you already pay for, or run local models on your own hardware. You are never locked into one provider.

Your process, drawn as steps you can see.

Describe the work, let Abstraction draft the flow, and run it the same way every time.

Describe

Say how the work gets done, in plain words.

Most teams know their work well. Turning it into something an AI can follow is the hard part. Start by describing how a piece of work gets done.

New flowIllustration

How does this work get done?

Every month we take the sales export, compare it with budget, write the commentary and send it to finance to sign off.

Draft the flow

Draft

Describe the process. Abstraction draws it.

  • Abstraction drafts it as a flow of steps on a canvas.
  • Correct the draft: add, remove or reorder steps.
  • Mark each step as done by a person, by AI, by AI with a person checking, or automatically.
Monthly report pack · draftIllustration
  1. Analyse the exportAI
  2. Compare with budgetAI
  3. Write the commentaryAI, person checks
  4. Sign offPerson
  5. Add a step

Approve

Raise a ticket. Approve it.

  • Start a flow by raising a ticket in plain words.
  • A small model on your own machine reads every request first and holds back anything vague or sensitive.
  • Nothing runs until a person approves it.
Ticket · October report packIllustration

Run the monthly report pack on the October sales export.

Read on your machine firstClear

ApprovalWaiting for a person

ApproveHold back

Run

Let the AI follow the flow.

  • The AI follows the flow, and every step is shown on its own.
Running · October report packIllustration

ApprovalApproved by a person

  1. Analyse the export
  2. Compare with budget
  3. Write the commentary
  4. Sign off

Review

See exactly where it went wrong, and fix only that.

  • The result comes back to you for review.
  • If it is not right, send back that one step with a note, instead of starting over.
  • Every run keeps a record of what was asked, who approved it and what the AI did.
Review · October report packIllustration
  1. Analyse the export
  2. Compare with budget
  3. Write the commentarySent back: “Use the revised budget figures.”
  4. Sign off

Accept resultSend back this step

Share

One way of working for the whole team.

  • Publish a flow so everyone on the team can run it.
  • Anyone can write their own flows; published flows are reused across the team.
  • One board shows all the work, from people and AI alike, with limits on how much is open at once so nothing piles up.
  • Live work shows which AI is working on what, right now.
Team boardIllustration

To approve 1 of 3

Supplier contract check

Running 2 of 3

October report pack

Tender response, section 4

For review 1 of 3

Leave policy check

Published to your team: Monthly report pack

Fix the step, not the whole conversation.

Every step in a flow stands on its own. When a result is not right, you see which step went wrong and send back that one step with a note. No more starting a chat again to fix one thing.

One step goes back with a note. The rest stay as they are.

Chat

Change a flow, or work directly, by talking to your AI.

  • Adjust a flow in conversation instead of on the canvas.
  • Talk to the AI you signed in with, cloud or local.
  • See how much of your AI subscription you have used.

Start with a flow that already works.

Abstraction comes with flows for common work. Follow the guided path, then change it to fit your team.

No prompts required. You follow a guided path instead of working out what to ask. When you want to change something, the chat is there.

Build an app and get it production-ready

Building an app with AI is easy to start and hard to finish safely. This flow takes you from an idea to an app that is ready for production, without you having to work out the prompts.

  1. RequirementsYou approveDescribe what the app must do. Abstraction turns it into a clear specification for you to approve.
  2. BuildAIThe AI builds the app in steps you can follow.
  3. TestAIThe app is tested against the specification.
  4. Security reviewAI, then you checkThe app is checked for security problems, and a person reviews the findings.
  5. RecommendationsBack to youYou get recommendations for taking the app to production, based on what was built.

Abstraction stops at the recommendations. Putting the app into production stays with you and your IT team.

More built-in flows Proposed

  • Turn a process into a flow

    Describe how you do something. Abstraction drafts the flow, you correct it, and you publish it to your team.

  • Review a document against a checklist

    Checks a contract, policy or supplier document against your team's own rules, flags the gaps, and a person signs off.

  • Respond to a tender or RFP

    Extracts the requirements, drafts each section from your past material, then runs a compliance check and review.

  • Monthly report pack

    Takes a spreadsheet or export, analyses it, drafts the commentary, gets a second opinion from a different AI, and goes for review.

Make it yours. Every built-in flow can be copied and changed. Add a review step, remove one you do not need, and publish your version to your team.

Add what your team needs.

Plug-ins are sold separately. Once bought, they are added to your Abstraction platform and use the same tickets, board and flows.

Meetings

Meeting notes that stay on your own computer.

  • No bot joins the call.
  • Recording, transcription and the summary all run on your own PC.
  • Speakers are named in the transcript.
  • Import a recording made elsewhere, and export notes to Word, PDF or Markdown.
  • Follow-up emails are prepared as drafts in your mailbox. Abstraction never sends them for you.
  • Works with English and Afrikaans.

Data harnesses

Data agents that build inside your data platforms.

Describe the change you need in plain words. A data agent works inside the platform to build it: changing an app, updating a semantic layer, landing new data in Bronze and carrying it through the medallion layers, or creating a pipeline. A person stays in control of every change.

The same agents can answer questions about your data and bring back a report.

Available for: Qlik and Microsoft Fabric.

What a data agent does

You describe a change. Each step is shown on its own, so you can see what was built and how.

  1. Connect
  2. Read the model
  3. Plan the change
  4. You approveThe exact change
  5. Try it firstNothing changes yet
  6. Make the change
  7. Read it back
  8. Your review

Built for control

  • Reading runs on its own. Changing anything does not.
  • Every change waits for a person to approve the exact change.
  • Each change is tried first without changing anything, then read back afterwards to confirm it landed as approved.
  • Steps that cannot be undone, such as publishing or deleting, sit behind their own confirmation.

Talk to us about plug-ins

Cloud or local, the choice stays with you.

Use the AI your business already pays for, add local models beside it, and keep a person in charge of what runs.

Any provider

Sign in with your Claude or ChatGPT subscription, with an API key, or both. Your AI costs stay on your own account.

Local models

Open models can run on your own hardware, beside a cloud AI or instead of it.

A check before anything runs

A small model on your own machine reads every request first and holds back anything that is unclear or sensitive. Then a person approves it. Nothing runs without that approval.

Your data stays where you put it

Abstraction runs on your own computers and uses your own AI accounts. With the Meetings plug-in, recording, transcription and summaries stay on your own PC. Data harnesses sign in to your data platforms with your own account.

What is coming next.

These are being built now. They are not available yet.

Platform

  • Process intelligenceComingLearns how work really runs and suggests which flows to build or improve next.
  • Improvement loopComingPropose a change to a flow, test it, score it and keep what works.
  • Memory across the platformComingOne place that remembers context across tickets, chats and meetings.
  • Tickets from emailComingRaise a ticket by sending an email.
  • Mac app and simple installComingAbstraction on a Mac, and a one-click install on an ordinary laptop.

Plug-ins and data harnesses

  • Harnesses from other providersComingThird parties build their own harnesses and offer them for Abstraction.
  • More data agent workflowsComingMore kinds of build, behind the same approval and read-back, and writing documentation for a data model.

Meetings

  • Ask across meetingsComingAsk a question and get the answer from all your past meetings.
  • Meeting briefsComingA short brief and agenda before each meeting.
  • Checked summariesComingMeeting summaries checked against the transcript, with a source for each point.

Why “Abstraction”?

A process written down for an AI is abstract. You read it step by step and picture it in your head, and so does the AI. When the result is wrong, it is hard to tell where the two of you missed each other.

Abstraction makes the process visual, so people can see it, check it and improve it.

Models are very good. What makes them unreliable is a poor foundation. Flows show how the work gets done, tickets say what is needed, and a person stays in control. That is the foundation.

The team

Abstraction is being built in South Africa by its three co-founders.

Be one of the first teams on Abstraction.

We are working with a small group of early teams. Tell us a little about yours.

We use these details only to contact you about early access, and delete them on request. Privacy notice to follow.