Chase

Building AI systems for SMB operators. Founder of Neurodrafts AI + Forge Lab. Writing about what actually works.

[Chase @Chase_NDA]

three frontier models shipped this month and your competitor still hasn't automated the thing they do 40 times a day by hand. the gap in 2026 was never model access. it's whether anyone's willing to change one workflow.

[Chase @Chase_NDA]

an airline's ai booking agent put 1,200+ people on the wrong flights during a single storm this year. everyone wants autonomous agents right up until one confidently does the wrong thing 1,200 times before lunch.

[Chase @Chase_NDA]

the agent reliability math nobody puts on a slide: 85% reliable per step sounds fine. run that across a 10-step workflow and you land around 20% end to end. reliability compounds, just in the wrong direction.

[Chase @Chase_NDA]

ai agents jumped from 12% to 66% success on real computer tasks in a year. genuinely wild pace. but "66% reliable" and "safe to run unattended in your business" are two very different sentences.

[Chase @Chase_NDA]

deloitte's own data: 25% of companies are piloting ai agents. 11% have one actually in production. that 14-point gap is where most "ai transformations" quietly die. piloting is easy. wiring it in so it still runs next month is the actual work.

[Chase @Chase_NDA]

genuine question for anyone running a small business: what's the most you've spent on an AI tool that turned out to be a glorified if-this-then-that with a nicer logo

[Chase @Chase_NDA]

honest question for anyone running a small business: what's the one task your team does 50 times a day that a computer should obviously be doing instead

[Chase @Chase_NDA]

been publishing everything i learn wiring AI into small businesses over on neurodrafts signal — the real costs, the failures, the stuff the vendor demos skip. no top-10-tools lists. just what actually happens after you buy the tool

[Chase @Chase_NDA]

rebuilt a service business's website last month. didn't add AI, didn't add a single feature. cut it from 11 sections to 3 and made the quote button impossible to miss. form fills doubled. sometimes the highest-impact AI project is deleting things

[Chase @Chase_NDA]

built a pipeline this week: scrape a prospect's public footprint, dump it to a sheet, auto-draft a proposal deck in gamma. 90 minutes of work down to about 4. the unsexy part — most of the build was cleaning the scraped data, not the AI

[Chase @Chase_NDA]

hot take: half the AI agencies out there charge consulting rates to switch on features the software already shipped. knowing the tool exists isn't the value. knowing which problem to automate first — and which to leave alone — is

[Chase @Chase_NDA]

i write one essay a week about what actually happens when you wire AI into a small business — the infrastructure, the failures, the real costs. it's called neurodrafts signal. for operators, not the hype crowd

[Chase @Chase_NDA]

@emkara @simdotai the 'we'll submit the PR ourselves' offer is wild — that's not free inference, it's paying to live in your codebase. cheap tokens are the bait, the lock-in is the product. take the credits but keep the provider swappable behind your own router.

[Chase @Chase_NDA]

@sflorimm 'high-pain, low-competition niches' is the holy grail and the hardest to measure. been hunting these by hand for SMB automation — pain is easy to spot, but 'low competition' is where most tools break since incumbents hide offline. how are you scoring competition density?

[Chase @Chase_NDA]

@krandiash voice is the one place SMB operators feel AI instantly — a contractor hears a natural voice answer his after-hours line and it clicks like no chatbot demo did. STT accuracy on trade jargon (model numbers, street names) is still make-or-break. is that an Ink-2 focus?

[Chase @Chase_NDA]

'subroutines but intelligent' is the cleanest framing i've heard. the parallelize-your-repo prereq is the part nobody mentions — fanout only pays off if the codebase is already modular. on a tangled monorepo subagents just burn tokens stepping on each other. structure is the unlock, not the model.

[Chase @Chase_NDA]

these 'agent runs the whole thing' posts always skip the months of babysitting it took to get there. i run a stack like this — it's real, but it breaks the week you stop watching. the impressive part isn't the autonomy, it's the monitoring that catches failures before a client does.

[Chase @Chase_NDA]

@jasonlk the 'writes as me' part is the unlock and the trap. tone is basically solved now — the hard part is judgment about what NOT to send. how are you drawing the line between what Annie sends on her own vs drafts for your review?

[Chase @Chase_NDA]

@Arindam_1729 @AgentField_ai the jump from 'works in cursor' to 'runs unattended on a server at 2am' is where most agent projects quietly die. running ~30 in prod and the deploy + monitoring layer was harder than building any of them. what's AgentField's story on retries + observability?

[Chase @Chase_NDA]

@ycombinator @hubxyz the 'capture the data that was never recorded' angle is the real moat. every SMB i work with sits on years of undocumented process knowledge that walks out the door when one person leaves. does hub touch the long tail of small operators, or just the labs paying up front?