---
title: "Be the AI Person Your Company Can Rely On"
description: "5 source-backed AI claims from Be the AI Person Your Company Can Rely On, with key statistics, context, and the original AI Adopters Club source."
url: "https://kbanc.com/claims-library/be-the-ai-person-your-company-can-rely-on"
source: "https://aiadopters.club/p/be-the-ai-person-your-company-can"
date: "2026-10-02"
topics: ["implementation", "strategy", "measurement"]
generated: "2026-10-04"
---

# Be the AI Person Your Company Can Rely On

By Kamil Banc | October 2, 2026

## Claims

1. **Perceived Speed Diverged** (source summary): METR's 2025 coding trial found perceived speed and measured completion time diverged among experienced developers.
2. **Start Simple, Add Complexity** (direct quote): Anthropic's agent-building guidance recommends starting with the simplest solution and adding complexity when it earns its place.
3. **Asking Isn't Approval** (source summary): Anthropic's permission documentation explicitly distinguishes instructing an agent to ask before sending from approval controls.
4. **Name a Change Reviewer** (source summary): NIST's voluntary guidance recommends naming who reviews changes to the prompt, tool, or rules.
5. **Measure the Whole Job** (Kamil's interpretation): Measure the whole job because drafting savings can be erased by added checking and correction time.

## Evidence

### Quote
> "Telling an agent to ask before sending doesn't establish an approval control." - Kamil Banc

### Key Statistics
- **20 minutes saved drafting versus 25 minutes added checking**: The author's illustration of why measuring the full job, not just the AI-assisted step, can reveal net added work of five minutes.
- **Ten representative historical test cases**: The recommended sample for a first look at performance, explicitly framed as not proof of a company-wide gain.
- **Three years advising companies on AI**: The author's stated practitioner experience underpinning his return to basics: understand the work, involve the right people, and measure.
- **2025 and 2026 METR studies on AI coding speed**: The 2025 trial found perceived speed and measured completion time diverged; the 2026 follow-up encountered selection and time-tracking problems.

## Context
Kamil Banc's methodology is grounded in direct observation rather than tool comparison: he advises booking time with the person who actually does the work, walking through the last completed case with files open, and examining a recent failure to surface hidden rules, workarounds, or defects. Before any AI test, he insists on recording baseline metrics for the whole job, including turnaround time, active preparation time, checking, corrections, and recurring cost, with waiting time kept separate from active work. His practitioner framework divides responsibilities explicitly: existing systems supply approved data, code enforces calculations and validation, AI drafts commentary and flags discrepancies, and people resolve exceptions and authorize outputs. The approach is deliberately incremental, testing on authorized historical cases, piloting with a small group against agreed acceptance checks, logging failures as test cases, and naming who maintains the workflow if it earns its place.

## Source
- Original: [Be the AI Person Your Company Can Rely On](https://aiadopters.club/p/be-the-ai-person-your-company-can)
- Cite: kbanc.com/claims-library/be-the-ai-person-your-company-can-rely-on

## Primary Evidence
- [METR’s 2025 coding trial](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) (metr.org; supports claim 1)
- [Anthropic’s agent-building guidance](https://www.anthropic.com/engineering/building-effective-agents) (anthropic.com; supports claim 2)
- [Anthropic’s permission documentation](https://platform.claude.com/docs/en/managed-agents/permission-policies) (platform.claude.com; supports claim 3)
- [NIST’s voluntary guidance](https://airc.nist.gov/docs/AI_RMF_Playbook.pdf) (airc.nist.gov; supports claim 4)