Greg Isenberg on where local AI pays, and three startups he would build with it

Key points

1. 1

   Ask if local is good enough

   Isenberg sets aside the usual question of whether a small model beats the biggest cloud model. His test is whether it can do this job, and whether running it on the user's own machine makes the product better. He keeps cloud models for deep research and hard reasoning. He points local models at private files, customer data, offline field work, fast responses, voice input and internal tasks that repeat.

   “is this model good enough for the job and does running it locally make the product better”

2. 2

   Local first pass, cloud escalation

   He expects most serious products to mix the two. His example is a tool for a professional services firm. A local model reads the sensitive drafts and strips out the private details. A cloud model sees only that clean version, when the customer wants deeper reasoning. A person approves anything important before it goes out.

   “You basically have local handling the private files as a first pass and then cloud handles the heavy thinking when you need it”

3. 3

   The filter for a local AI business

   Before his three ideas, Isenberg gives his screen. He wants small, cash-flowing businesses that need no venture money and fix a painful workflow. He looks for five traits together: sensitive data, repeated review work, bad incumbent software, costly mistakes, and work that happens close to the device. All three ideas pass it.

   “I look for a customer with sensitive data, repeated review work, bad software usually, uh expensive mistakes”

4. 4

   Home health QA: start as a service

   Idea one is a desktop reviewer for home health agencies. It checks visit notes, care plans and dictated notes against each other before billing or audit. It would flag, say, dizziness noted with no vital signs recorded. He would start it as a service for five small agencies, run the reviews with local AI and check each one by hand. The issues that keep coming back become the product.

   “I would write down the 20 issues that keep showing up and those issues become the checklist and then the checklist eventually becomes the product”

5. 5

   Restoration reports: sell with old jobs

   Idea two is an offline phone app for water, fire and mold restoration crews. The technician walks the site taking photos and voice notes. The app drafts the report and flags gaps before they leave, such as a ceiling photo with no moisture reading. He would pick one niche, study the owners' templates and their early-2000s software, and open with a demo on their own past jobs.

   “send me three old jobs and I'll show you how fast your techs could create reports”

6. 6

   Pre-send review sells itself

   Idea three checks outgoing drafts at professional services firms. He calls it schmuck insurance. It flags a wealth advisor's line that sounds like a promised return, or an accountant's number that does not match the attached file. He would start with email review for independent wealth advisors, interview 10 of them and turn their worries into a checklist. The firm already pays a person to check drafts, so the product speeds up an existing habit.

   “this is so sellable because the buyer understands this behavior, and they already asked someone to check the draft”