Augma
Contact

AI for business

Almost everyone has tried it. Few have made it work.

What the research says about AI at work: who uses it, where it pays, why most pilots stall, and the habits of the few that get real value. Every figure links to the page it came from.

The mindset ↓How to start ↓

Where things stand

AI has moved from experiment to everyday.

Where it pays

Measured gains are real, and largest for people still learning.

The mindset

Seven habits of the companies that get value.

  1. 01

    Leaders go first

    AI changes how decisions are made, so it needs an owner at the top who uses it, sets the goal and says what success looks like.

  2. 02

    Redesign the work, don't bolt on a tool

    The value BCG found sits in reshaped core functions, not in a chat window added to an old process.

  3. 03

    Get the data ready

    An assistant is only as good as what it can read. Clean, owned, permissioned data comes before the model.

  4. 04

    Keep a person in the loop

    Let the machine draft, sort and check; let people decide where it matters, and make the hand-off easy.

  5. 05

    Earn trust with rules

    Write down what it may and may not touch, log what it does, and say clearly when an answer came from AI.

  6. 06

    Teach everyone

    The NBER study saw the biggest gains among newer staff. Training spreads the benefit instead of concentrating it.

  7. 07

    Start small, measure, then grow

    One narrow job with a number you can check beats a broad pilot with no baseline.

What goes wrong

Most pilots stall. Rarely because of the model.

MIT's researchers put the gap down to tools that don't learn or fit the daily workflow, not to model quality.

Five things to keep in mind

From the builders' side

Lauren Tan on trusting agents with real work.

Lauren Tan (poteto) is an engineer at Cursor and previously worked on the React Compiler at Meta. In recent talks, including a workshop on coding agents and a live conversation with Matt Pocock in October 2026, she describes how agents came to land large volumes of work. Her emphasis is not on clever prompts but on verification: let the agent run and check the real product, keep the codebase easy for an agent to navigate, and turn recurring mistakes into automated checks.

The lesson carries beyond code: trust grows from checks you can run, not from hope.

Watch: Poteto with Matt Pocock, Oct 2026 ↗

How to start

Five steps, one small job at a time.

  1. 1

    Pick one job

    A repeated task with a clear owner, where a mistake is cheap to catch.

  2. 2

    Measure today

    Write down how long it takes and how often it goes wrong.

  3. 3

    Prepare the data

    Gather the documents and rules it should follow, and decide who may see what.

  4. 4

    Build with people in the loop

    The assistant drafts, a person approves. Every answer keeps its source.

  5. 5

    Compare and decide

    Check against the baseline after a few weeks, then stop, fix or grow.

Curious where AI could help in your work?

We're builders who use AI every day, and we're happy to talk it through: what you do now, what you'd like to stop doing, and whether AI is even the right answer.

Start a conversation →Who the big AI players are →Meet K.A.R.A, our own AI →SME, startup or corporate? →