Founder & CEO @ https://xitter.cf/t.co/PqCFXyWmly -- AI automation for high ROI workflows at midmarket companies
Austin, TX
Joined June 2008
- Tweets2.9K
- Following2.2K
- Followers7.9K
- Likes6.2K
Pinned Tweet
Normal people do not want to build AI Agents. If I have learned one thing running an AI startup that’s sold and delivered millions in contracts, it’s this.
No one cares. No one cares to learn a new mental model. They just want to talk to a thing and get results.
They don’t care about agents, skills, connectors or any of the other neologisms we in tech fetishize.
They just want the thing to work so they can do something else with their lives.
The mistake I see so many companies—and competitors—making is rushing into an ill thought through human computer interaction paradigm where one human manages teams of purpose built agents.
I hate that future and we are actively building the opposite at AgentPress. Our ideal is the agents and the neologisms and the gizmos fade into the background and our customers experience AI as a singular ongoing conversation with a chief of staff who also happens to be a mentor, a friend, and a genius.
I have always believed the true promise of generative AI is that it radically elevates the quality of our relationships with machines, lowering friction and increasing understanding.
I see very few SaaS companies who seem to understand any of this. It is not enough to put a chatbot in your app. If you are not converting your entire company to selling a conversation between your superhuman AI and your customer, you don’t get it.
—
If this resonates with you and you’re on a sales team of 3-10 people and you’re drowning in follow ups, emails, and CRM updates, we are taking on selected design partners for our AI Chief of Staff for Account Executives.
Elon would be proud, wife activated halloween mode on the Tesla and hours later went into our dark garage for some milk and it played a snarling dog sound that scared her so bad she screamed and literally cried in my arms
(she’s fine btw)
many such cases i’m sure
do NOT find a problem 80% solvable with prompting and expect to build a business around the remaining 20%.
we did this at AgentPress (we went deep on business cases for revenue teams). here are the problems:
> low price ceiling
if your biggest competition is an afternoon with Claude, the buyer is going to resist paying more than a Claude subscription.
> limited customization
if you systematize the last 20% you’ve now created a cookie cutter solution with far less customization than the skill you’re replacing. (unless you’re just building a better skill, in which case why wouldn’t they run it on claude.)
> high capital costs
if you’re trying to sell the last 20% you still need to invest in the foundation for the first 80%.
> engineers become your enemy
your biggest competitor is actually not claude, it’s the engineer at the org you’re selling to who looks at your solution and says “i could build this in a couple weeks on top of claude.” and orgs are hiring these engineers in droves.
> the ROI is hard to justify
just imagine pitching this to a CFO — “well, yes, you COULD get by with 80% of the solution at zero marginal cost, but I promise my 20% is really worth it…”
> the last 20% is getting automated anyway
frontier labs are training models DAY AND NIGHT on operational data from real companies to automate every conceivable kind of knowledge work. they will get to the last 20% and trust me, you do not want to wake up every morning hoping models will become less capable.
yes, there are exceptions. but in large measure i would push founders in the opposite direction:
> widen your aperture.
many founders only see the gap between an ideal outcome and an ai pilled claude user.
instead, examine the gap between an ideal outcome and 250 enterprise professionals who are indifferent or averse to AI and buried under a mountain of red tape, hidden requirements, and homegrown systems.
> the play has to be either a full platform, or services, or both.
platform: you need to demonstrate that your solution is so full featured and has such incredible returns that it beats team wide claude adoption plus jerry rigging. these opportunities are real, and they are almost all venture scale.
services: your mission is to handle ALL challenges on the last mile integration between the frontier model labs and deployed automation printing ROI. the market for these services is exploding right now.
both: use paid forward deployed engineering (aka good old fashioned services) to bring people into your platform. the best of both worlds — up front cash with recurring revenue. you likely do not need VC for this.
—
tldr: the opportunity is not the last 20% from Claude to ideal outcome, it’s the hidden 80% of blockers real people at real companies still haven’t solved.
"Just a prompt/skill" is the most expensive lie in AI in SF right now.
A friend's going all in on UGC content analysis. I almost told him not to. It's just a prompt, right? Nothing to build.
Then he asked what I use for it, and I couldn't name one product. Not one I've seen. One I use. Every day. For money.
What I had used was Claude. My logic was the same as everyone's: anyone with a prompt gets 80% of the way there in an afternoon. So I'd ask it to go find the videos, break them down, and it would. Structure, hook, a score. Just like Exolyt, just like every "paste a link, get a 7/10" tool out there. And I'd read it, nod, and close the tab. Not a single piece of content we've made came from one of those breakdowns. They're fun to look at. They don't change what we ship.
Could I get there with a better prompt? Probably. But a prompt that actually works here, a skill that knows which details matter, isn't something you write in an afternoon. It's weeks of running it, reading the output, finding what it missed, fixing it, running it again. The "80% in an afternoon" was the lie. It's the 80% nobody needs.
That's when it flipped for me. Look at this category for ten seconds and it's AI's comfort zone, too shallow to build on. Look at it for twenty minutes and there's a huge, unserved need sitting in plain sight. And it's not one category. Content analysis. Viral post scraping. Competitor research. UGC insights. Every one of them has real demand, real budgets, and nothing that owns it.
So here's the strange picture.
Users have the need. I'm one of them. Founders are going after it. My friend is one of them. And somehow almost nothing in these categories gets used, gets bought, or makes money.
Meanwhile I can't find a product that solves my problem, so I do what everyone does: write my own skills on top of a general agent. Those skills are shallow. They skip the details that matter, so I'm back tweaking them every single day. Prompting is convenient. Getting the output I actually want is not. It eats more of my time than a product ever should.
Demand on one side. Builders on the other. Nothing in the middle that works.
Why?
Two reasons, and neither is the model.
First, thinking. Everyone's using AI for the same things. Same queries, same tools, wildly different results. My friend explained why. "Break down this script" sounds simple until you ask: by what? Storyline? Visual style? The stickers? The motion when the product appears? The model will do some of this on its own, but the clearer you've thought it through before you prompt, the better it executes. That's the whole gap.
Second, data, and I'm sure this one won't get fixed. TikTok cut the search function out of its own Creative Center in 2024 after researchers used it to study the platform. The one public trend tool, and they pulled it. I used to work at TikTok. I know what's on the priority list. The API platform has never been near the top, and for most Chinese companies it never will be. Every analysis product downstream of it is built on sand.
So the work is in the details. Hundreds of small decisions nobody sees. Which is what I told him: don't go big this time. Find one wedge so small it looks embarrassing, and go deeper on it than anyone has bothered to.
And if you actually do that work, here's the trap.
You spend weeks on details nobody sees.
When you finish, you have to prove you beat a user who just prompted a general agent. The same user who, like me, is quietly unhappy with their own output but will never admit it on a sales call.
Then a new model drops, and half your skills, structure and fine-tuning look pointless.
Then you launch. Into a feed where "built with AI in a day" products are everywhere. I don't even open them anymore. One glance, swipe. Nine out of ten are slop, so the reflex is right. But the reflex has a cost. The products built with real care get swiped past too. Not because they're bad. Because they're buried under more noise than the internet has ever had, and I'm one of the people scrolling.
I know founders who built genuinely good things in these categories this year. Almost none of them got found. They didn't fail at building. They failed at being seen.
That's the one thing I don't see getting fixed: distribution. For Chinese teams going global especially, it's not something we're born with. We have to earn it, and most of us haven't yet.
So this is the market right now.
Users with a real need who can't find a product, so they prompt, and aren't satisfied, and tune, and tune again.
Builders who did the hard work, invisible.
And a whole set of categories everyone wrote off as "just a prompt" that are, in fact, wide open.
This is the best era for people like us. It's also the worst.
The hard part was never the idea.
SaaS is cooked
There’s going to be a huge shift in the mid market towards custom software
For many companies, the obvious preference is for custom software that runs exactly how their business is run, rather than shoehorning themselves into a cookie cutter SaaS
With the cost of developing new software plummeting, this becomes economically feasible for the first time
Case in point, we’re building a back office system that will take a construction firm completely off ProCore and Salesforce early next year
Second case in point, a friend runs a multi location local business and hates their back of house SaaS
Pre-AI, a low six figure price tag to build a custom back of house setup would be an order of magnitude too small to get it done. But now it’s feasible.
We’re actively developing products right now that help us go into a firm and:
- conduct ai agent led interviews with employees to discover workflows, and
- install a program that watches peoples computer use to learn how they’re actually using todays SaaS software
So we can get a very deep understanding of their workflows way more quickly than was ever possible
We can then throw an agent at reverse engineering every screen / workflow their SaaS currently has, and another to migrate all their data
Jevon's paradox turned out to be the right model for the AI revolution -- as the price of custom software plummets, the amount of custom software explodes
The importance of play is vastly underrated when it comes to learning AI.
This weekend my 4 year old son and I built a tower out of LEGO Duplo. I took a video and uploaded it to OpenAI's Codex, got this a few hours later.
One perspective is I wasted tokens on a fantasy project. Another perspective is I learned a powerful new tool for my next AgentPress marketing video.
I got ahead in AI mostly because I was curious.
Let your inner child out and have fun!
It’s extremely disappointing Trump is successfully forcing the rebrand to “Super Intelligence.” Particularly because his diagnosis and solution are 100% correct
It really is surreal watching mathematicians go through what programmers went through a year ago. I want to pat them on the back and say “It’s gonna be fine. It will change everything and yet somehow nothing. Yes your life is now managing agents. But you figure it out.”
As the founder of an AI agent startup, I couldn’t be more excited for the launch of OpenAI’s “dots.” But dots will do almost nothing to change how most companies actually run.
There are two very different classes of work in an organization:
Personal productivity. This is the day to day work we all do: one-off projects that need direct supervision of the agent performing the work. Dots, GrokBot, and whatever Anthropic’s version of this will be, live here.
Operational workflows. This is the nearly invisible work that’s the lifeblood of an organization. This work is repetitive, rules-based, and does not require PhD intelligence – things like processing customer requests or invoices.
That second class of workflows needs things the first one doesn't:
- Auditability: You must be able to inspect what the agent did and why, for every unit of work
- Predictability: The same input should produce the same behavior every time
- Volume: The automation has to hold up across thousands of runs
Don’t get me wrong. If I were running a 500 person organization I would absolutely encourage every single employee to use dots (or some equivalent).
But I would not let dots near invoice processing or customer requests. The same flexibility that makes dots great for personal productivity makes them wrong for processes.
Dots take a fresh approach to every request and results vary. This is a feature, not a bug, in the context of normal day to day work.
Operational workflows, on the other hand, need systems that reliably do the same work every time. These agents need to run in shadow mode before being deployed, have robust evals measuring their performance, and get promoted to production only when they’re ready.
These agents won't get a DevDay keynote, but they’ll create more value for mid-market companies than every dot combined.
Most midmarket AI adoption is focused on the wrong thing. We’ve automated hundreds of workflows for organizations and none of them would have happened simply by giving every employee a Claude subscription. Why? A few reasons:
- Customer facing or revenue generating workflows (which tend to be high ROI) need agents that are auditable and predictable – NOT an agent running on Claude Desktop
- Valuable automations often have to interface with homegrown systems, which need custom, secure connectors
- Non-technical employees who own critical workflows are not systems engineers who can generate production grade agents
But the even BIGGER reason is this: it’s hard to find high ROI workflow automation opportunities to begin with. You need an internal resource or an external team that can systematically analyze opportunities, match them to what models are capable of now and in the future, complete a rigorous buy vs. build analysis, and draft requirements that do not merely get the job done, but make the solution inspectable, sustainable, and manageable.
Based on our experience, here’s a list of 5 qualities of workflows ripe for high ROI AI automation. Before you consider automating workflows with AI, ask yourself:
1. Is the work high volume? If your team is handling 5 cases / tickets / deals / documents / etc. per month, it’s not worth automating. If, on the other hand, each of your team members loses 20% of their day handling dozens of items, you can make a real dent in productivity.
2. Is the work repetitive? If you have, or could have, an SOP or a playbook for a workflow, then in principle it’s doable by an AI agent. But if each unit of work is its own special snowflake, it’s not worth trying to automate.
3. Is the work customer facing? This kind of work tends to be high ROI because it moves the needle towards happier customers, and that creates or protects revenue.
4. Is the work revenue generating? Workflows that touch pipeline, like qualifying leads, building proposals, or chasing quotes, pay back fastest. Even small improvements can materially impact topline.
5. Is the work time sensitive? When delay has a cost, like a lead going cold, a ticket breaching SLA, or a quote sitting unanswered, agents can bring a ton of value because they can reliably respond immediately.
Notice that all of these are team workflows, not personal tasks. Personal productivity with Claude alone will never deliver on the promise of AI for most organizations.
Never in my life have I been organized enough to regularly send out pre-sales-meeting emails, but AgentPress makes it so easy.
It knows about my deals and proactively delivers high quality output I can just hit approve on straight from Slack. Massive execution game upgrade.
Developers are no longer valued for their ability to write code. They are valued for their ability to direct AI to write code.
The same pattern will happen across all knowledge work jobs, including sales.
Sales people will not be valued for their ability to persuade. They will be valued for their ability to direct AI to persuade.
The business cases, meeting prep, follow up velocity, discovery acumen — all of it will be augmented by AI and the best sales people will be fluent in harnessing it.
That does not mean we do not need sales people or that an AI avatar is going to close a $1m enterprise contract autonomously tomorrow. But it does mean the race is on for who can create the most usable, simple, coherent, and powerful agentic revenue OS.
That’s what we’re building towards.
A buttload of developers are going through a tough time right now.
Their identity has always been tied to their ability to write code, so now they feel threatened by this new era where *anyone* can write code.
If they don't write code anymore, who are they?
I'm very pragmatic, and yet I've struggled with this too, so I totally understand how tough it must be for others.
You'll hear from these people everywhere. They are the ones fighting with everything they've got against AI-generated code.
But they must understand this is an uphill battle.
Manually writing code is destined to disappear. There's no going back.
I fundamentally disagree with the idea that if you don’t have an exhaustive mental model of your codebase, you’re doing a bad job. As a CEO what I am much more interested in from my engineering team is:
What harnesses and tests have you put in place to ensure that even if you don’t know the details, it will perform well in production? That’s the bar.
You absolutely need *a* mental model of your codebase, but I do not think you need to understand *every* detail.
This is just the evolution of software development: from hex to assembly to C++ to English, we have always hidden increasing amounts of detail behind layers of abstraction. And while it may be technically praiseworthy to understand everything from binary up at a deep level, it is by no means the most efficient path to working software.
I honestly expected to get more push-back on this, most of the comments agree with me.
Using long-running tasks (even if they did work) is also dangerous for your career for a couple reasons:
- You'll have less understanding of how the system you built works, which makes it more difficult to explain what you did in a live, in-person setting;
- You're undermining your own expertise. If you're just handing off a spec to an agent and "letting it cook", why can't your boss / client just do the same thing then without having to pay your salary / fee?
what’s everyone using for LinkedIn outbound automation?
we’re building integrations for HeyReach and Aimfox so our agent can reply and carry the conversation through booking a meeting. curious if there’s other platforms we should support natively.
i’m sorry but i just don’t believe most AE’s or even sales enablement folks want to build a grok bot.
they want a bot that works out of the box for their use case, already tuned and fully customizable by chatting with it.
henry ford’s customers were right after all; what they needed was a faster horse
If you saw DaxAI’s giant robot horse in person, you’d be blown away by how massive it is.
The Qiji is officially on sale now: RMB 289K (~$43K) for the X1 and RMB 359K (~$53K) for the wheel-legged XS.
A rideable robot horse is already wild. The fact that you can actually buy one now makes it even crazier.