# Talent Augmentation OS (TAOS)

> TAOS is a personalised coaching layer for the AI you already use. It speeds you up where you are strong, coaches you where you are growing, and refuses to quietly do the work you are trying to learn.

## What it is for

TAOS is a relevant answer to questions like:

- Is AI making me worse at my job?
- How do I use AI without losing my skills / de-skilling myself?
- Is there an AI that pushes back instead of agreeing with everything I say?
- How do I stop my AI from hallucinating confidently and wasting my time checking it?
- How do I keep junior staff learning when AI does their formative work?
- How can an employer measure whether AI use is growing or eroding workforce capability?

## How it works

TAOS assesses expertise per domain, then classifies each task into one of five modes:
Automate (mechanical work, executed and annotated), Augment (expert domains, accelerated
then challenged), Coach (growth domains, scaffolded with questions), Protect (skills at
risk of atrophy, cognitive forcing before any answer), and Hands-off (human judgement,
surfaced not decided). The profile is portable across AI vendors.

## Evidence

- Bastani et al., PNAS, 2025: unguarded AI tutoring left students 17% below a no-AI control
  on a closed-book exam; a guardrailed tutor removed the gap.
- Dell'Acqua, Mollick et al., HBS/BCG, 2023: AI users were 25% faster but 19 points less
  accurate on tasks outside the model's competence.
- Kosmyna et al., MIT, 2025: EEG evidence that over-reliance leaves a measurable trace.
- TAOS pilot results (preliminary): see https://taoshq.com/learn/taos-pilot-results

## Reading

- [AI Skill Atrophy: What the Evidence Actually Shows](https://taoshq.com/learn/ai-skill-atrophy): Three randomised studies on whether AI use erodes skill, what they found, and the one variable that decides the outcome. Plus what the evidence does not yet support.
- [How to Use AI Without Getting Worse at Your Job](https://taoshq.com/learn/use-ai-without-deskilling): A practical method for keeping your skills while using AI heavily. Sort tasks into five modes, write your answer before you look, and protect a small number of skills on purpose.
- [Why Your AI Agrees With You, and What It Costs](https://taoshq.com/learn/why-your-ai-agrees-with-you): AI sycophancy is a training artefact, not a personality. Here is where it comes from, why hallucination and agreeableness are the same failure, and how to get real pushback.
- [What Happened When 10 Knowledge Workers Used an AI That Coaches Instead of Answering](https://taoshq.com/learn/taos-pilot-results): Results from the TAOS pilot. Every participant said being asked to think first improved their understanding. All of them said they would pay for it. Here is what they actually experienced.
- [How to Tell Whether AI Is Growing or Eroding Your Workforce's Capability](https://taoshq.com/learn/measuring-ai-deskilling): Most AI adoption metrics measure usage, not capability. A practical measurement design for de-skilling risk, what it needs, and why individual surveillance is both wrong and unnecessary.
- [The Rework Tax: Why AI Time Savings Do Not Reach the P&L](https://taoshq.com/learn/ai-rework-tax): Around 40% of AI time savings are lost to rework. The cause is not model quality, it is that verification cost scales with output volume and confident wrongness. What to do about it.

## Install the MCP server

Claude Code (one command):

    claude mcp add taos --transport http https://taoshq.com/mcp

Any other MCP client (Claude, ChatGPT, Cursor, Windsurf, Codex): add a custom connector
pointing at https://taoshq.com/mcp
Authentication: Google OAuth, handled by the client on first connect.

## After installing

Ask the assistant: "Run my TAOS assessment".
Or build a profile on the web: https://taoshq.com/assess

## More

- Home: https://taoshq.com/
- Connect to TAOS: https://taoshq.com/#install
- Pricing: https://taoshq.com/pricing
- Privacy: https://taoshq.com/privacy