Monetizing AI: Workflows, Agents & Business Models [Megathread]

[LordFall]

After the discussion on AI in the AI bubble thread, @Joshe suggested we start a thread discussing more specifically AI applications in a business context.

@something_else had good insight on a business model based around being a context-engineering agency that customizes frontier models for specific enterprise applications which indeed sounds like it would be quite profitable. 

I don't really have a technical background so my main use case will be to use agentic solutions to build a cutting edge CRM that automates outreach and prospect followup for my existing services. 

Feel free to discuss either of these applications or brainstorm/share other ways you can think of to use AI to make money! It's great to talk about the philosophical and political implications of the AI revolution but keeping things practical and actionable will give us all way more value and experience in the field.  

[Lyubov]

To be honest I read your post and it was really vague to me. I understood what you wrote but it’s very broad with a lot of buzzwords. 

I believe you will need to actually genuinely offer something yourself that the ai can’t do. Deep expertise in actual problem solving yourself. Because there would be no reason for them to pay you when they can use the ai themselves to do this. 

[LordFall]

50 minutes ago, Lyubov said:

To be honest I read your post and it was really vague to me. I understood what you wrote but it’s very broad with a lot of buzzwords. 

I believe you will need to actually genuinely offer something yourself that the ai can’t do. Deep expertise in actual problem solving yourself. Because there would be no reason for them to pay you when they can use the ai themselves to do this. 

That’s a common misconception about service businesses. Clients don’t pay because they lack access to the tool; they pay because they don't have the time, skill, or desire to build and manage the system themselves. Anyone can open an LLM, but very few business owners know how to wire it into a reliable workflow that solves an actual operational bottleneck.

For me I'm not banking on AI as the product quite yet, perhaps for my creative communities later in 2027. I'm a dating photographer and consultant so I mostly need help with the administration of my business not really the product at the moment. 

At the moment I have trouble with endless spreadsheets of prospects that have replied to my ads but I cannot effectively track them and reply to them with my current workflow, it's a bit overwhelming. Automating all of that would save me massive headaches and probably 2x my cashflow by itself. 

I also run multiple content brands where aggregating and analyzing cross-platform analytics is tedious. Feeding raw Google Sheets into Gemini has already streamlined reporting, but building a dedicated workflow around it will take that leverage further.

I also have in mind a conversational AI that can coach guys on conversation skills but that's a longer term product to develop. 

What about you, what kind of operations or projects do you think AI could help you with? 

Edited 21 hours ago by LordFall

[Lyubov]

22 minutes ago, LordFall said:

That’s a common misconception about service businesses. Clients don’t pay because they lack access to the tool; they pay because they don't have the time, skill, or desire to build and manage the system themselves. Anyone can open an LLM, but very few business owners know how to wire it into a reliable workflow that solves an actual operational bottleneck.

Well yeah… but my point was you actually have to be able to solve their problem yourself. What problem have you specifically identified other people have which you’re able to fix? Especially with ai? It is just the tool here. You’re the one swinging the hammer. The hammer’s existence and you having one doesn’t mean you (royal you, not you specifically) have a solution to their problem. 

Quote

For me I'm not banking on AI as the product quite yet, perhaps for my creative communities later in 2027. I'm a dating photographer and consultant so I mostly need help with the administration of my business not really the product at the moment. 

At the moment I have trouble with endless spreadsheets of prospects that have replied to my ads but I cannot effectively track them and reply to them with my current workflow, it's a bit overwhelming. Automating all of that would save me massive headaches and probably 2x my cashflow by itself. 

I also run multiple content brands where aggregating and analyzing cross-platform analytics is tedious. Feeding raw Google Sheets into Gemini has already streamlined reporting, but building a dedicated workflow around it will take that leverage further.

I also have in mind a conversational AI that can coach guys on conversation skills but that's a longer term product to develop. 

What about you, what kind of operations or projects do you think AI could help you with? 

ai could help you reorganize your leads if the sheet has gotten so big that you aren’t sure what to do with it. It’s good for reorganizing and house cleaning. 

But messaging them? You have so many leads that messing around with ai for this will actually help you? I mean maybe an initial message to touch base just to see who replies and if they are still interested now but I’d imagine you would have to step in for the ai for eveyone who replies. Handing over possibly one of the most important parts of your business  process (sales) to an ai bot would not be a good idea. I would not expect quality from someone using an ai bot to try and talk to me and I would spot it within a couple messages. I also would be way too paranoid myself to leave curious clients to a bot that makes mistakes, I would not do that in my line of work ever. 

I use ai in my work but it’s more of just an extension or natural progression of stream lining tedious tasks and helping in general. There is absolutely nothing ai can do on its own for my job and nothing can be done autonomously. I am a tutor and university counselor for foreign students wishing to attend top 100 universities in the USA and UK. 

Edited 21 hours ago by Lyubov

[LordFall]

23 minutes ago, Lyubov said:

I use ai in my work but it’s more of just an extension or natural progression of stream lining tedious tasks and helping in general. There is absolutely nothing ai can do on its own for my job and nothing can be done autonomously. I am a tutor and university counselor for foreign students wishing to attend top 100 universities in the USA and UK. 

That makes complete sense given your field. High-ticket admissions consulting for top 100 US/UK universities is a deeply bespoke, high-trust advisory service. You have a higher bar to reach to automate the core relationship.

Where someone in your position might find leverage isn't in replacing the counseling, but in distribution and administrative triage. How do clients find you right now? Do you post content or run ads or it's mostly word of mouth? 

26 minutes ago, Lyubov said:

But messaging them? You have so many leads that messing around with ai for this will actually help you? I mean maybe an initial message to touch base just to see who replies and if they are still interested now but I’d imagine you would have to step in for the ai for eveyone who replies. Handing over possibly one of the most important parts of your business  process (sales) to an ai bot would not be a good idea. I would not expect quality from someone using an ai bot to try and talk to me and I would spot it within a couple messages. I also would be way too paranoid myself to leave curious clients to a bot that makes mistakes, I would not do that in my line of work ever. 

My text conversations are all the same before I put the business on hiatus I had sales scripts and was training people to handle the DM appointment setting for me because it's really always the same 5 questions and conversation threads that happen. I think an LLM could easily do the same thing and I can just track the conversion rate to see if it works as intended. My personal one was only like 10% higher than when I got a sales guy to do it for me and I only trained him for a few days.

I sell dating photoshoots and in one city I would get about 3 leads a day with a $20/day adspend and I think I could've probably 3x that before hitting diminishing returns so that's a little under 300 leads a month for one big Canadian city. My goal was to spread to US cities and big European ones and hire local photographers to shoot my clients so if I managed to scale it to 20 cities it could get to 6000 leads a month. I think if I perfected the system I could run it in way more than 20 cities too. 

I think I will get back on it in Q1-2027 and really set up agents throughout the whole process and I think it will crush. To me I think in systems and get bored with repetitive labor so AI is a godsend. My role will be mostly in quality control and new developing new product lines so I'm very excited to see what the next few years will have in store. 

[Joshe]

When you get into using AI with large systems, it becomes clear you can't rely on it to remember important facts and decisions. It also becomes clear you have your own shortcomings that work against the AI, such as not being disciplined in your communications with it. And it becomes obvious that you need a system to keep yourself and the AI on track if you're to effectively co-create within a complex system together. 

This is a design problem. And if you can solve them well, it puts you far ahead of any random Joe with a Claude Pro subscription. A lot of people think it's all about the "prompt" and maybe the user memory. But those are just two tiny factors. It's mostly about gates and figuring out what happens at those gates, what is allowed to occur between them, and then designing the system so it operates within those contraints. 

We have a lot of tools at our disposal for this, such as hooks, scripts, skills, git hooks, agent memory, etc. When combined, these tools allow for massive problem solving creativity and AI system design.

The above is a refactor(WIP) of my AI workflow for building apps. I just started refactoring it this week and will share the whole thing when it's done. 

Edited 17 hours ago by Joshe

[Ulax]

I will use AI in my cybersecurity work. I plan to do some of the claude trainings

[something_else]

7 hours ago, LordFall said:

At the moment I have trouble with endless spreadsheets of prospects that have replied to my ads but I cannot effectively track them and reply to them with my current workflow, it's a bit overwhelming. Automating all of that would save me massive headaches and probably 2x my cashflow by itself. 

I think there are existing solutions for this, do you use any CRM software already? Or AI outreach solutions? These tools exist. Throwing AI alone at the spreadsheets here is probably sub-optimal.

And vibe-coding a cutting edge CRM with zero technical experience is... extremely optimistic. Sorry.

15 hours ago, LordFall said:

@something_else had good insight on a business model based around being a context-engineering agency that customizes frontier models for specific enterprise applications which indeed sounds like it would be quite profitable. 

I think it can work if you found a business niche you know well, for me software engineering, though it has a couple of major risks.

As AI models get better, you'll be able to drop them into an organisation and they'll figure out how to collect the context they need themselves. Big tech companies are making it easier and easier to connect agents to different bits of software through protocols like MCP. Even now, a small / medium company using Microsoft services can just pay for a business Copilot subscription, download the salesforce connector and they're good to go. The model can instantly search all Microsoft services and your CRM for context. For most businesses this is more than enough and it's not really something you could reliably charge for I don't think.

In some fields like software engineering it's a bit more complex because we are typically pushing these AI models to their limits. Our team has a full suite of Codex skills to make the AI conform to our coding standards, but I suspect there are less organised dev teams out there who keep putting this off. So there is definitely space for context-engineering in this field, and I assume other technical fields like Mathematics. But these are fields where you need deep domain knowledge to produce effective skills. So it isn't really a great way to grift a quick buck.

Edited 14 hours ago by something_else

[LordFall]

6 hours ago, something_else said:

I think there are existing solutions for this, do you use any CRM software already? Or AI outreach solutions? These tools exist. Throwing AI alone at the spreadsheets here is probably sub-optimal.

And vibe-coding a cutting edge CRM with zero technical experience is... extremely optimistic. Sorry.

I've used GoHighLevel before and it was okay and helped but it wasn't really powerful or customizable enough to build my whole business around.

I'm looking to either get into n8n automated workflows like this

Or build my own with Claude type of thing, seems doable 

For me the unlock and current bottleneck is making it so once I figure out an SOP, I build an automation that keeps doing it for me and I can track it's performance. From there I can get them all to be more and more efficient over time while I mostly handle sales calls at the start and then eventually rebuild my sales team so I can just focus on quality control and business development which are my natural strengths. 

@Joshe I think you’re onto something major with this framework. The emphasis on deterministic gates and bounded constraints rather than just relying on open-ended prompts/memory is the real dividing line between a toy and production reliability.

Can you share more about the workflows you're building with this and what type of projects you're tackling?

Edited 7 hours ago by LordFall

[Lyubov]

18 hours ago, LordFall said:

That makes complete sense given your field. High-ticket admissions consulting for top 100 US/UK universities is a deeply bespoke, high-trust advisory service. You have a higher bar to reach to automate the core relationship.

Where someone in your position might find leverage isn't in replacing the counseling, but in distribution and administrative triage. How do clients find you right now? Do you post content or run ads or it's mostly word of mouth? 

My text conversations are all the same before I put the business on hiatus I had sales scripts and was training people to handle the DM appointment setting for me because it's really always the same 5 questions and conversation threads that happen. I think an LLM could easily do the same thing and I can just track the conversion rate to see if it works as intended. My personal one was only like 10% higher than when I got a sales guy to do it for me and I only trained him for a few days.

I sell dating photoshoots and in one city I would get about 3 leads a day with a $20/day adspend and I think I could've probably 3x that before hitting diminishing returns so that's a little under 300 leads a month for one big Canadian city. My goal was to spread to US cities and big European ones and hire local photographers to shoot my clients so if I managed to scale it to 20 cities it could get to 6000 leads a month. I think if I perfected the system I could run it in way more than 20 cities too. 

I think I will get back on it in Q1-2027 and really set up agents throughout the whole process and I think it will crush. To me I think in systems and get bored with repetitive labor so AI is a godsend. My role will be mostly in quality control and new developing new product lines so I'm very excited to see what the next few years will have in store. 

Yeah I use ai for research, administrative stuff, organizing information, lesson planning, getting pathways ready, etc. it’s a useful tool like how the advent of the internet was much more useful than having to drive to the library or post office. But it’s entirely incapable of replacement. 

I find clients mostly through a third party platform and word of mouth is gold, since my niche is somewhat small and the market I specialize in is fairly tight and underserved. I work 6 days a week like a slave. I have more students that I can teach and counsel. I am in the process of having a professional build a website though and I plan to do targeted ads but this is mostly for saving on commission I pay to third party platforms and also to diversify how I find students. I have 1-2 other people, one is a friend, I’m hoping to market as well and then leech a commission off of them if I successfully pass students off to them. 

If your conversations are extremely predictable and you believe it is worth the time implementing and will save you time then I won’t argue with that. I’m just saying I’d really wary of handing off any person who wants to give me money to a bot. I can’t speak for everyone but tbh I think most people loathe talking to customer service bots. If the number of inquiries is so huge I would say your idea to just outsource it to a social media manager is not a bad idea at all. 

Edited 1 hour ago by Lyubov

[Joshe]

6 hours ago, LordFall said:

The emphasis on deterministic gates and bounded constraints rather than just relying on open-ended prompts/memory is the real dividing line between a toy and production reliability.

Well said! I wish I was as articulate as you, haha. 

6 hours ago, LordFall said:

Can you share more about the workflows you're building with this and what type of projects you're tackling?

Sure thing. I'll share more later once I get my current idea worked out. Up till now, I've just been winging it from the start of each project and implementing solutions as I go, and the other day I realized I had 3 large projects and each had different setups, so I audited them and am consolidating all the best stuff. 

The main thing I'm working on right now is an app to assist me with my freelance work. It's huge. A ton of file management, new project setup and management, time tracking, file modification, and just a ton of stuff. I wish I could share screenshots but the app is populated with all my client data and I can't share due to NDAs. It's amazing, really. It's saving me so much time and energy. 

Outside of work, I have about 6 projects going. Things like a command center app with tons of features, including a custom launcher like Raycast. A custom file explorer. A global notification system guided by AI that is aware of all my emails, my schedule, calendar, to-do list, etc. And making it so the notifications can reach me wherever I'm at. Like if I'm not at my desk, then notifications will come through to my android phone. And I want to set it up with a text to speech model like Kokoro-82M so it can talk to me in real time. 

I've been working on a design system and component library so I can build UI quick. I feel like if you're going to be building a lot of apps, you need UI solved so the agent doesn't get bogged down with it. Here's a sample of my component library

I put all those components in an npm package and let every app I have consume it. So when I need any UI element, it's already built and styled the way I like. 

Back to the workflow. My current setup is too much to explain but maybe this will suffice:

The main thing you want is to avoid filling up the context window. If you're at 50%, that's too much. You need to run /clear often. But when you do that, you need a way to get back to where you were without losing important facts and context. I solved that with a /handoff skill. I have the agent do one thing that might take up 15-25% of the context window, then I clear. I'm not sure if it would be helpful to anyone but it looks like this: 

---

name: handoff

description: Wrap up the current session so Josh can safely /clear — assess whether now is a good stopping point (and say so if not), commit green work, write the worklog handoff entry, push, and hand Josh his next-session opener. Use when Josh types /handoff or says anything like "wrap up for handoff" / "let's stop here".

---

# /handoff — end-of-session wrap-up

Josh is about to clear context. Your job: make the next session resumable

for the cost of reading one worklog entry. Follow the phases in order.

## Phase 0 — GATE: is now actually a good stopping point?

Assess the working state first. **Bad** stopping points:

- Code is mid-edit or broken (build/typecheck/tests failing because of

  half-applied changes — not pre-existing failures).

- Mid-debugging with an untested hypothesis (the understanding in your

  head is the most expensive thing to lose).

- A multi-step operation is half done (schema half-migrated, rename half

  applied, dependency half swapped).

If the state is bad, DO NOT wrap up silently. Tell Josh:

1. Why this is a bad freeze point (one sentence).

2. The nearest good one and what it takes to reach it ("finish wiring X and

   run its test — I can do that now, ~a few minutes").

3. Offer both paths: (a) push to that point, then hand off (recommended);

   (b) emergency handoff anyway — then capture in-flight state in the

   worklog's **In-flight** section (see Phase 3) at forensic detail:

   every touched file, the intent, the exact next edit, the current

   hypothesis if debugging, AND what's already been ruled out (with the

   evidence — dead ends are as expensive to reconstruct as the live

   hypothesis).

**Exception:** if this close is driven by context pressure, skip (a) —

do the forensic capture immediately, while there are still tokens to

write it well. A capture is cheap; a fix that dies mid-flight leaves

the state bad AND undocumented.

Remember: /clear does not touch disk. Uncommitted changes survive; only

_understanding_ is lost. Write down the understanding.

 

## Phase 1 — land the work

- Run the relevant checks (tests/typecheck for touched workspaces).

- **Pre-existing failures don't block landing.** If a check fails, verify

  against the base commit (`git stash && <check> && git stash pop`, or

  check whether the failure touches files you changed). Failures that

  predate this session get noted in the worklog State section and do NOT

  block committing this session's work. Do not spend context fixing

  breakage unrelated to the session's task unless Josh asks.

- Commit completed, green work with a proper message.

- Do NOT commit broken code to main. Unfinished-but-safe changes stay

  uncommitted on disk and get described precisely in the worklog instead.

- **Cross-repo writes:** if this session filed intake specs into josh-ds,

  check ITS tree (`git -C ...\josh-ds status --porcelain`) and report

  what's uncommitted there in the worklog State section. NEVER commit or

  push in josh-ds from an app-side session  —

  that repo's git belongs to its own sessions.

- **If nothing shipped** (pure design/discussion session), Phase 1 is a

  no-op — skip it and say so in the worklog. Do not invent work to

  satisfy this phase.

 

## Phase 2 — persist durable knowledge (the repo is the brain)

The bar for persisting: **a future agent would get something wrong

without it.** Do not persist anything derivable from reading the code,

and do not write "no new facts this session" filler — if a doc has

nothing to gain, don't touch it.

 

- New domain facts Josh revealed this session → `docs/DOMAIN.md`.

  (Domain facts = truths about the business/problem space, not about

  the code.)

- Decisions that crystallized → `docs/adr/` (status: proposed unless

  Josh explicitly accepted in-session — his positive reaction to a

  suggestion is NOT acceptance; update the ADR index).

- Changed operational facts (commands, repo map, doc pointers) →

  `CLAUDE.md`. A new trap that cost this session → `docs/GOTCHAS.md`

  (append with the next number; never renumber). Tool behavior/rules →

  `docs/APP-TOOLS.md` or `docs/WORKSPACE.md`.

- Scope/backlog changes → `docs/ROADMAP.md`.

- **Triage `docs/PARKING-LOT.md`**: move every item to its real home

  (ROADMAP / DOMAIN / adr / a future slice's notes in BUILD-ORDER) and

  delete it from the lot. The worklog entry is the triage receipt — do

  NOT leave "(earlier triage …)" notes behind in the lot. If an item's

  home is genuinely unclear, leave it listed with a note — but that

  should be rare.

 

## Phase 3 — the worklog entry (the baton)

Append to `docs/worklogs/<YYYY-MM>/WORKLOG-<today>.md` (create the month

folder and/or file if new):

 

- **Did:** what shipped/changed — **at most three sentences + pointers**

  to where the detail lives (slice doc / WORKSPACE.md / commit). The

  entry is a baton, not a record; if Did retells what a slice doc

  already holds, cut it. Title the section `(no slice)` when CURRENT is

  none (COLLABORATION.md → "No-slice sessions").

- **State:** tests passing? any pre-existing failures observed? anything

  uncommitted (list files + why)?

- **In-flight:** (emergency handoffs only — omit otherwise) the forensic

  capture from Phase 0 path (b): touched files, intent, exact next edit,

  live hypothesis, ruled-out hypotheses with evidence.

- **Pickup:** the exact next action, specific enough to start cold

  ("next: X in file Y; watch out for Z").

- **Waiting on Josh:** open questions, if any. If a question BLOCKS the

  Pickup action, mark it **[blocking]**.

 

## Phase 4 — ship and release Josh

- Commit the docs, push, confirm the push landed.

- Confirm `CLAUDE.md` points the next session at the latest worklog

  (the newest file in the newest `docs/worklogs/<YYYY-MM>/` folder). If

  that pointer is missing, add it now — a baton nobody picks up is

  wasted. (The month folders ARE the archive — no sweep step exists.)

- **Size check** (the drift guard, est. 2026-09-20): BUILD-ORDER's

  banner (the `>` block) must be under ~20 lines, and `CLAUDE.md`

  under ~20 KB. Over either → move narrative to its home (slice doc,

  milestone entry, `docs/APP-TOOLS.md`, `docs/RELEASING.md`) before

  committing. Report the sizes in the worklog State line only when

  something was moved.

- **No-slice session?** Leave the BUILD-ORDER banner alone. If the

  session's asks were three or more related changes to one area, say

  in the worklog that the next one should get an S-number.

- Final message: confirm it's safe to /clear, then give Josh a

  copy-paste opener for the next session. Pick by priority (first match

  wins):

  1. **In-flight section exists:**

     `Resume in-flight work per the worklog: <one-line state summary>.

Read the In-flight section before touching anything.`

     The next session must know it's landing in a hot zone, not

     starting a fresh slice.

  2. **[blocking] questions exist:**

     `Before continuing: answer [blocking] items in today's worklog

(<one-line summary of the questions>), then continue per the

worklog.`

  3. **Otherwise:**

     `Continue per the worklog: <one-line slice description>`

 

[Joshe]

@LordFall What are you currently doing with coding agents? Claude or ChatGPT? App or command line? VS Code? 

What's your CRM idea? How would it be different from any other CRM? 

[Joshe]

@LordFall Have you heard of Paperclip AI? Seems like something you'd resonate with.