HatchWorks

HatchWorks We build AI-native solutions — and use AI to build software better, faster, smarter.

09/03/2026

You asked the system a question. You got a good answer.

A week later, you asked the same question and got a slightly different one.

That is the moment trust dies, and it is why in-house legal teams have been slow to adopt AI tools that look impressive in a demo. A company's risk baselines are not probabilistic. Payment terms are net 60 unless someone senior says otherwise. Disputes follow a defined procedure. A lawyer needs that applied identically every time, because the lawyer carries the liability and the model does not.

On the latest Talking AI clip, RiskVantage AI co-founder and CTO Emad Khazraee explains the architecture built around that constraint. A deterministic ontology holds the legal reasoning and produces the same answer every time. Language models sit underneath it, handling document understanding and drafting, supervised by the ontology rather than trusted to decide.

Consistency, not capability, was the product requirement all along.

👇 Link in the comments

For years, Emad Khazraee told his co-founder that building a legal AI company was a bad idea.He is now the co-founder an...
09/01/2026

For years, Emad Khazraee told his co-founder that building a legal AI company was a bad idea.

He is now the co-founder and CTO of one.

On the latest Talking AI episode, he walks Matt Paige through the objections he kept raising. A nice interface over a frontier model has no moat, because Anthropic or OpenAI will ship a better one overnight. An assistant bolted into Word never sees the transaction, so it never gets smarter. And frontier token prices are subsidized heavily enough that the business built on them is not really a business yet.

What changed his mind was not a market. It was an architecture. A deterministic ontology owns the legal reasoning, and small domain-specific language models handle the language.

The best reason to trust a founder is watching them try to talk you out of it first.

👇 Link in the comments

A typical MCP setup, five servers and 58 tools, spends around 55,000 tokens before anyone types a word.A Claude Skill co...
08/28/2026

A typical MCP setup, five servers and 58 tools, spends around 55,000 tokens before anyone types a word.

A Claude Skill costs 50 to 100 tokens until the moment it is actually needed.

Both are ways to give Claude more capability. The difference between those two numbers is the entire argument for Skills.

A Skill loads in three layers. The name and description stay resident, and that is the 50 to 100. The full instructions load only when the work in front of Claude matches, usually 1,000 to 5,000 tokens. Bundled references, scripts, and assets load only if the task reaches for them. That is why hundreds of Skills can sit on one Claude instance without slowing the model down or crowding out the conversation.

Which puts unusual weight on one thing most teams rush. A Skill fires on its name and description alone. If the description does not say what the Skill does, when to use it, and the phrases your people actually type, it never activates. Helps with stuff is not a description.

The carousel below has the three layers, the description mistake, and where to start.

👇 Link in the comments

If you were building your company from scratch today, what would you do fundamentally differently because of AI?That was...
08/27/2026

If you were building your company from scratch today, what would you do fundamentally differently because of AI?

That was one of four questions we put in front of a group of executives in Atlanta last night, and it was the one that took the longest to answer.

We could have run another presentation about where AI is headed. Instead we put the hard questions on the table and let the room debate them.

The other three:
1. What separates AI investments that create real business value from those that simply create activity?
2. As AI usage scales, how do you decide which model should do which work, and when does it make sense to customize or own more of the model layer?
3. As AI moves from recommending actions to taking them, where do humans still need to stay in control?

There were not always easy answers. That was the point.

The conversation is shifting. Less about whether to adopt AI, much more about where it creates value, how it changes the way companies operate, and how much autonomy we are willing to hand it.

Thank you to everyone who joined us and brought their perspective to the table. These are exactly the conversations we need more of.

08/26/2026

There's a lot of token maxing going on right now, and it suits the labs just fine.

On the latest Talking AI, Zapier co-founder and CEO Wade Foster makes the case for something less glamorous and more effective: hybrid agent workflows, which he thinks are genuinely underrated.

Some jobs need a model. Unstructured input, writing code, generating email. Plenty of others are better served by deterministic steps that run reliably and cheaply.

The point of blending them is that you aren't just letting a model run amok and spend your money to arrive somewhere vaguely useful. You get something purpose-built for the task.

Spending more tokens isn't the same as getting more done.

👇 Link in the comments

08/25/2026

Your automation works perfectly until the day it doesn't.

Deterministic workflows are wonderful right up to the edge case. Something unexpected arrives, or the input is a wall of unstructured text or an image, and the rails that made it reliable are exactly what make it brittle. So you patch in another rule, then another, and slowly rebuild the mess you automated away.

On the latest Talking AI episode, Zapier co-founder and CEO Wade Foster frames the way out. Agents are almost the mirror image: they handle the messy and unexpected, but with weaker reliability, higher cost, and more latency.

Neither is the answer alone. Match each step to the mode that suits it.

👇 Link in the comments

08/22/2026

The jobs apocalypse is the wrong thing to be losing sleep over.

On the latest Talking AI clip, Tom Scott, CEO of Wrike, doesn't deny AI will cause disruption. Any transformative technology has a jobs impact, and this one is no exception.

But he pairs that with something he holds firmly: he's a humanist. Humans evolve. They create new roles and find space for themselves, the way they always have.

So he spends far more of his time worried about how we build the next generation of full stack professionals than about any apocalypse.

Disruption is real. Fatalism is a choice.

👇 Link in the comments

08/20/2026

Benchmarks are how we measure AI progress.

They're also why so many people feel let down by these models.

Because a benchmark is a proxy for real work, and a model that optimizes against the proxy doesn't automatically get better at your job.

On the latest Talking AI episode, Zapier co-founder and CEO Wade Foster describes the pattern: a headline score makes a model look extraordinary, then you hand it one of your own tasks and it disappoints. Both experiences are real.

Trust benchmarks for direction. Trust your own evals for decisions.

👇 Link in the comments

The AI coding industry is measuring the wrong thing.Lines per hour. Tokens per sprint. Percentage of code "AI-generated....
08/19/2026

The AI coding industry is measuring the wrong thing.

Lines per hour. Tokens per sprint. Percentage of code "AI-generated." Every one of those numbers can go up while the thing you actually shipped gets worse.

Here's our CEO Brandon Powell on it:

"Most of the market is still measuring AI coding by how fast it produces lines. That is the wrong scoreboard. What our clients need is code that passes review, holds up in production, and shows a return they can point to."

That belief is the whole foundation of GenDD (Generative-Driven Development) — and this month it was named Code Generative AI Solution of the Year in the 2026 AI Breakthrough Awards. 🏆

The program is in its ninth year and drew more than 5,000 nominations from companies in over 20 countries.

Proud of this one. And prouder of the teams across four countries who proved the model works before anyone handed us a trophy.

Link to the full announcement in the comments. 👇

Wade Foster runs an automation company that stands to gain from AI hype.So he published the number that undercuts it.On ...
08/18/2026

Wade Foster runs an automation company that stands to gain from AI hype.

So he published the number that undercuts it.

On the latest Talking AI episode, the Zapier co-founder and CEO points at AutomationBench, his company's own benchmark of real cross-app business tasks. The top frontier model on the leaderboard completes them 18.1% of the time.

His conclusion isn't that agents don't work. It's that turning a model loose on an entire workflow is the wrong bet right now. Blend deterministic steps where reliability and cost matter with agents where judgment and messy data live.

The most trustworthy voice in a hype cycle is usually the one with the least incentive to be honest.

👇 Link in the comments

Address

5256 Peachtree Road NE Ste. #140
Atlanta, GA
30341

Opening Hours

Monday 8am - 6pm
Tuesday 8am - 6pm
Wednesday 8am - 6pm
Thursday 8am - 6pm
Friday 8am - 6pm

Website

Alerts

Be the first to know and let us send you an email when HatchWorks posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.

Contact The Business

Send a message to HatchWorks:

Shortcuts

Share

Category