Derrick Edward & Misha Afon

Derrick Edward (business / co-founder) and Misha Afon (tech / co-founder) of Harnex AI, a Wellington-based generative-AI consultancy, sit down with Andy for the podcast's first episode dedicated entirely to AI in construction. The pair met at AI meetups while both working in NZ enterprise (Misha came from the telco market that "trialled 3G/4G/5G" in NZ, Derrick is the technical engineer of the pair); they describe a "human-first" service model where the firm understands the client's processes before deciding whether generative AI, traditional ML, or no AI is the right fit. Their core product is education and consulting, not selling tools. Derrick's "perfect intern with perfect recall but terrible judgment" frame for current generative AI lands as the episode's keystone metaphor. The conversation traces three layers: (i) a tier-2 construction adoption diagnosis ("I've done it this way for 20 years; why change?"), (ii) the data-security trap (the free chatbot economy, "when it's free, you're paying with your information") plus the emerging local-models-on-device alternative ("not there yet"), and (iii) the agentic-AI accountability question, if an agent gets a cost forecast wrong and it propagates into a BI report and then into a board pack, where does blame sit? Their answer: human-in-the-loop checkpoints at every workflow boundary, and cost-proportional accuracy expectations ($1k/month tools vs $50k/year enterprise agents). The episode closes on Andy's parenting story, using "Zena" (his named ChatGPT bot) on voice mode to explain selfishness to his seven-year-old son and then having it generate a multiple-choice quiz, plus the 21-day Roblox earn challenge teaching consistency and resilience. First episode under sponsorship, Andy's first paid sponsor Daniel Small / Elemental QS announced at the close, the start of the YWA monetisation arc.

Key Topics Discussed

• First all-AI-in-construction episode. earlier episodes touched AI (EP30 Hamish Race / Cody AI; EP37 Kieran Mackenzie / Pressing vision-safety; EP53 Jono Lockwood / Timescapes; EP68 Maria Mingallon / Mott MacDonald enterprise AI); EP66 is the first to make AI the spine.
• Harnex AI as a service-first AI consultancy. predominantly service, productising on the side; tagline-equivalent "humanise it, simplify it, practically help you leverage it." Pre-decision: understand the business first, then choose between generative AI, traditional ML, or no AI.
• NZ adoption pattern, perfect test market, terrible adoption speed. Misha frames NZ as the trial market for telco generations 3G/4G/5G but slow to actually adopt. ~OECD productivity gap quoted at "35% behind average, 37-38% behind US." NZ recorded negative 1.8% productivity last year overall.
• Three dials to anchor adoption. Derrick's framework: AI adoption should be justified against (1) increase revenue, (2) reduce cost / efficiency, (3) reduce risk. "Generative AI can 100% do all three if you adopt the right strategy and educate yourself."
• Top-down adoption is the precondition. "Leaders that adopt AI top-down create an AI-centric culture." Anything mid-tier alone burns out fast. Andy: "It's mandated within my company."
• 75-80% of employees already use generative AI. Derrick quotes "every report", so the corporate question is no longer "should we" but "how do we govern what's already happening." Banning ChatGPT is not a control; education is.
• The data-security calculus. DeepSeek (Chinese servers) vs ChatGPT (US servers); "when it's free, you're paying with your information." Free-tier inputs become training data unless explicitly switched off. Enterprise contracts ($25/user, Andy's deal, to $100-150/user) are the only zero-data-retention path.
• SOC 2 + HIPAA compliance vocabulary. Misha and Derrick's quick primer on the certifications that actually matter when assessing an AI tool's data handling. Construction industry largely unaware of the compliance vocabulary.
• Local models, promising but not there yet. Llama and similar open-source models can run locally on a phone or MacBook GPU and address pure data-privacy use cases. Currently 100GB+ on disk, fans run hot, accuracy not at ChatGPT-Enterprise level. "It's the future, but the future hasn't arrived."
• Andy's multi-model workflow. starts with ChatGPT Enterprise, cross-checks against Claude, Perplexity, NotebookLM, last resort DeepSeek. "Every model has biases I can't see, the answer is to triangulate, not to trust one."
• Generative AI vs agentic AI, accountability question. Andy's hypothetical: an agent generates a cost forecast that's wrong and the report propagates into BI dashboards. Whose fault? Misha and Derrick: human-in-the-loop checkpoints at every workflow boundary, and the deliverer of the final output owns it.
• Cost-proportional accuracy. $10/month chatbots will hallucinate; $1,000-$50,000/year agentic systems infer that cost in their accuracy and orchestration. Understand the $/accuracy curve before deploying.
• "AI is a perfect intern with perfect recall but terrible judgment". Derrick's keystone framing. AI gets domain breadth, humans keep judgement and ownership. Generate 10 options, use domain knowledge to pick.
• Verification rule. Derrick + Misha personal rule: any AI output must be verified before action; once actioned, you own the decision. Aligns with standard professional-services PI/liability framing.
• Andy's parenting + AI. 7-yo son and 3-yo daughter. Voice-mode ChatGPT (Andy's bot named "Zena") used to teach selfishness through scenarios + multiple-choice quiz. "Zena is parenting my kid. And I didn't think I could have done that myself."
• The 21-day Roblox earn challenge. Andy's son had to complete 7 daily tasks (math, English, drum practice, etc.) for 21 consecutive days to earn Roblox; missed-day reset to day 1. Lesson: consistency + resilience. Son cried, started over, completed it.
• Skills children will need. Andy's two-step view: (1) fundamentals first (math, grammar) without technology, (2) prompting-with-purpose layered on top. "Without fundamentals you've got nothing."
• Voice mode as the future modality. Misha's framing: text prompting will give way to natural voice conversation. "Communicating yourself effectively might be all you need."
• Sam Altman's "please and thank you" anecdote. multi-million-dollar compute cost of polite prompts. Backed by the emotional prompting research paper, politeness + caps + "this is imperative for my career" objectively improves output. Andy hadn't been doing this; will start.
• Sponsorship arc launches. Andy's first paid sponsor Daniel Small / Elemental QS announced at episode close. Marker for YWA monetisation timeline.

Notable Quotes

"AI is a perfect intern with perfect recall but terrible judgment. So you're not going to give an AI system judge power and autonomy over everything. You own that decision. Let it be a tool.", Derrick

"When it's free, you're paying with your information. That's how the machine language is being trained. Remove that information when you start paying, or even in free models, switch off data training.", Derrick

"75 to 80% of your employees are using generative AI. Just teach them what to put in and what not to put in.", Derrick

"The biggest misconception is that AI is a genie that can solve all your problems. It is not. But if you use it well, it can alleviate a lot of it.", Derrick

"We don't use it. We abuse it.", Andy, on AI inside Strategic Planning Co

"It's not a matter of if a breach will happen. It's a matter of when.", Derrick on cyber-security data risk for cloud-hosted AI tools

"The companies doing this are going to be getting all of the market share. The Xerox that never took on internet, the Blackberry that never took on the mobile phone, that's the position you're in if you don't adopt.", Derrick

"Without your fundamentals, you're stuck. Once you have those fundamentals, then it's about how to nurture that curious mind in a pathway that uses this technology for good.", Andy on parenting

"Zena is parenting my kid. I didn't think I could have done that myself.", Andy on ChatGPT voice mode explaining selfishness

Guest Background

Derrick Edward, co-founder of Harnex AI (Wellington, NZ). The technical engineer of the pair ("the nerd"). Came up through enterprise AI experimentation; describes constant attendance at NZ AI meetups during and after work. Self-described nerd; Misha describes him as "the tech guy."

Misha Afon, co-founder of Harnex AI. Background in telco where he learned the NZ market behaviour ("perfect demographics to trial new tech, slow to adopt"). Self-described as the business-side counterpart to Derrick, though the conversation makes clear both are technically fluent. New father at the time of recording, son age unspecified, but referenced as relevant context for the parenting discussion.

Note on naming. The official spelling per the spreadsheet (07_PODCAST/Podcast Guests.xlsx Guestlist) and the email domain harnex.ai is Misha Afon and Derrick Edward. The audio transcription renders the company as "Hanx AI" / "Hanxi" and the surname as "Aon", these are transcription artefacts; the emails misha@harnex.ai and derrick@harnex.ai resolve to Harnex AI as canonical. Both spellings appear once in this wiki: the official spelling on this page; the as-spoken artefact in the raw transcript. See Open Question below.