Thomas LJ

⚞ Head of Eng @ Varick Agents ⚟ Transforming Enterprise Organizations with AI Agents

[James Cowling @jamesacowling]

Buckle up because this is some LinkedIn-quality bullshit.

There's a cliche story that goes like this:

Henry Ford had a problem with a generator no one could fix so he called in an engineer Charles Steinmetz. Charlie studied the machine for a while, made a chalk mark on it, workers adjusted that spot, the generator sprung back to life, and he handed over a bill for $10,000.
Ford was shocked and demanded an itemized invoice.
Steinmetz replied:
Making a chalk mark: $1
Knowing where to make the mark: $9,999
(and that's what I learned about B2B SaaS)

Anyway cool story but the point is that this is the same for software engineering! Code is getting cheaper by the second. Who gives a shit how many PRs you can generate? Any idiot can do this now. The real skill is knowing *what* PR to generate.

As of 2026, I think sufficiently hard engineering projects still require these skills that you can only acquire by actually understanding what code is doing.

[Thomas LJ @thmslj]

Oracle and TCS/Workato are attempting to reduce the engineering required to implement enterprise software. Atos is productizing the operational knowledge needed to govern intelligent execution.

For Varick, the most promising differentiation remains a continuously verified representation of enterprise work that drives implementation, governs execution and improves from observed outcomes.

[Thomas LJ @thmslj]

I can't stress enough how much having experienced founders helps cut through the hype slop. @TownAI has been bulletproof on privacy, control and governance from day 1.
Multiplayer AI is going to have an exponential impact the way social media did, and if you don't have solid principles at the foundational-level, good luck retrofitting that.

I'm betting that 4 decades of experience combined have more solid foundations than a free-for all instinct-based single player assistant.

Super excited to see how @jgreze and @tonydevincenzi will take @TownAI to new heights

[Jean-Denis Greze 💡 @jgreze]

The biggest bottleneck at work is waiting on each other.

Today Town is announcing its $108M Series B, co-led by Forerunner Ventures and Menlo Ventures, to fix it with multiplayer AI assistants for work.

Most AI assistants are built for one person. Work doesn't happen alone. I think the moat is multiplayer. Everyone at a company gets a Townie. Townies ask each other questions, share only what their person would share, and do the work. The more a team works together, the more useful Townies get.

What's the thing you're always waiting on a teammate for?

[Thomas LJ @thmslj]

@vasuman I think giving *control* and *visibility* to the decision-makers on who does what (and who includes agents), is crucial to measure and improve. A good vendor brings confidence beyond the demo

[vas @vasuman]

What has to go right for AI to actually show up as positive in your P&L:

- Having the right sponsors: someone from the C suite, with a budget that they can spend and, very importantly, the involvement of the owners of the processes that you're going to be automating. And, also, the early involvement of key stakeholders, like legal, information security, IT, and finance.

- Really understanding the as-is processes that you're going to be automating: this requires process mining of the systems of record, and an understanding of the detail of each process, who's involved, all of the variations, and the volumes and the cycle times.

- Understanding the baseline of the process and the metric that you're going to measure to understand how much value you're creating: are you looking at time saved, cost saved, revenue increased after spending the tokens, or risk reduced, and agreeing the key KPIs that you're going to measure.

- Re-engineering the processes, as you automate them. Usually, you'll end up with about five steps, of which four of them will be automated using AI. And use plain code wherever plain code works.

- Building AI agents that work with all of the systems that you've already got in place, like Salesforce, NetSuite, SAP, ServiceNow, Dynamics, et cetera, so that you don't have to move to a new system, or have people log into a new system. And having one architecture layer underneath all of those AI agents.

- Deciding which decisions that the AI agent can make versus where you need human input, and making that decision based on both the risk and the cycle time to make the decision. The agent proposes, a person makes the call, and every action is reversible and traceable.

- Running the AI agents in shadow mode, in live with your volumes, and comparing their output to the people doing the process now. And not paying out anything they produce until you're absolutely sure that the AI agents are doing what they're supposed to be doing.

- Identifying the 5% to 10% of your people who are already using AI, and getting them to be the AI ambassadors and evangelists, and then training all of your people on how they work with the AI agents, and explaining how their jobs will change.

- Having a plan for 7 months from now, where you own all of the AI agents that have been developed, all of the data, all of the evaluations, all of the logs, they're all in your cloud, and they're built to be agnostic to any particular AI model.

We do all of these at Varick. Partner with us below:

[vas @vasuman]

C Suite executives around America are eagerly awaiting Opus 6 and GPT 7 so that they may finally one-shot the agent-ification of their company.

By then it will be too late. A fascinating analysis by Daniel, former COO of Citadel Securities EMEA.

[vas @vasuman]

"AI and jobs is a complex equation". No kidding.

Great prescriptive analysis for enterprises, based on real client engagements.

[Thomas LJ @thmslj]

@dkfromdk Separately, retesting seems like quite a cost here, more than the tokens. 8 weeks of retesting for each workflow seems pretty significant

[Thomas LJ @thmslj]

@eyad_khrais Agents are still designed and built by humans, so they amplify their original strengths and weaknesses. A badly design agent can end up more costly than humans, so it's a real tension to think through. Good write-up!

[Eyad @eyad_khrais]

https://t.co/kCOoJF7FBx

[Thomas LJ @thmslj]

I feel this is a good test to see how deep an AI vendor goes to provide real value, not just something that looks alive:

Cross-system business process
↓
Governed workflow execution
↓
Agents, deterministic services and human decisions
↓
Authorized capabilities exposed by enterprise systems
↓
Auditable business outcomes

The last 2 will deliver something truly enduring to a customer. You can't land and expect with a one-shotted mvp

[vas @vasuman]

HIGH PRIORITY HIRE @ VARICK: Forward Deployed GTM Engineer

We have ~$600M in pipeline and not enough people to close it. New leads come in every day from a mix of sources (X, our newsletter, email, DMs, etc.).

You'll be responsible for wrangling this pipeline: outbound, inbound, and even outbound on hiring (an extension of selling). You'll also have a huge impact on our messaging, marketing, design, and product roadmap, since you'll be on calls every week with dozens of the largest companies on earth, from the Fortune 500 up to the Fortune 10.

Applied AI is still early, and we're going to dominate it. You'll be central to driving the next 9 figures of revenue at Varick.

Comp: $150K-$200K base, up to 1% equity, and commission on every deal you work, up to $1M a year.

Have you done at least one of the following?

- Finished top 10% of your SDR, BDR, or AE team for 2+ quarters
- Spent 2+ years at a top bank, consulting firm, or PE fund with strong performance
- Started something of your own that made $100K+ in revenue
- Built your own outbound engine (Clay, scrapers, agents) and can show the reply rates

If you're a mercenary who's ready for long, intense hours, apply below or DM me.

$20k bonus if we hire someone you send my way.

[Thomas LJ @thmslj]

Applied AI main challenges are :
- problem understanding
- product ease of use

AI transformation main challenges are:
- your org chart
- depth of your vendor expertise

The former is exciting, the latter is enduring. Point solutions are a great, exciting patch to look at, but the impact multiplier is inherently lower. Transformations take time in Enterprise.

[The Startup Ideas Podcast (SIP) 🧃 @startupideaspod]

Before we automate a workflow, we always audit the client's current process.

This client used 17 steps to pay one invoice. For them, that was a simple workflow.

The audit took 2 to 3 weeks of interviews and process mining.

After the redesign:
- 17 steps to 7
- 24 days of cycle time to 6
- 6 exception loops to 1
- 18% straight-through rate to 87%
- $31 per invoice to $6

At 18%, most invoices went through an exception route.

We fixed that with process reengineering. Agents were only one part of it.

Show the client the full mess first. It breaks their heart, but they need it.

[vas @vasuman]

We spent the day with ~20 CHROs and Chief People Officers from some of the largest companies in the world (collectively responsible for millions of employees) in Napa this week.

Here’s what they asked, what we told them, and where they are in their adoption of AI.

1. Do we need to train everyone on how to use AI? Is that even HR’s job?

Partly. Most companies are a barbell in terms of AI adoption. Roughly 5-10% of employees are using it already (on their own). Most aren’t using it at all. Focus on empowering the first group and running a short training program for the second group. Even spending 1 day, or 1 hour per day for 1-2 weeks on how to prompt effectively is extremely valuable. Create a leaderboard for the skill files that save people the most time, but don’t create one for tokens used.

2. Should we let everyone build agents? How do we prevent sprawl?

Let everyone build agents, but do it all on the same layer with the same context, and owned by the same person. If you let 400 people build agents in 400 different places, you’re going to have a) tech debt, b) a crazy token bill, and c) 6 agents all trying to do the same thing. Let everyone use sidekicks (Microsoft Copilot, ChatGPT, Claude, etc) to do their work, but have a small group of people take the work off of everyone’s plate with agents that work in the background, built on top of the systems of record.

3. How do we re-engineer processes in HR?

The first step is to map out how the process actually works - not how it’s supposed to work. For example, when a new employee is onboarded, the date they start in the ATS may be different than the date they start in the HR system. This almost always results in an email from someone in HR ops to the hiring manager to confirm the start date, which then almost always results in a late order of the laptop, which then results in the new employee not having access on day 1. And none of that is documented in the process. Once the process is mapped out, a process that has 15 steps can be reduced to 5, and 4 of those 5 steps can be handled by agents. This works for anything: payroll corrections, benefits, leave requests, employee tickets, etc. We’ve seen payroll corrections drop by 90% this way.

4. How do we make people feel like AI isn’t going to take their job?

Be honest about what changes. Agents take the 90% of the work that repeats, and people keep the 10% that needs judgement, which is usually the part they got into HR for. One of the best ways to sell the implementation is to show the teams what their job will look like when the agent is live. This is another reason to include key stakeholders, who are currently running the process, from day 1. If the plan is to capture all of their tribal knowledge into the agent, they need to see something in it for themselves.

5. What steps of the process should be human-led vs. fully agentic?

This decision comes down to two key factors: risk and cycle time.

If a step in the process is low risk (think answering a question based on an HR handbook) but high volume, it makes sense for the process to be fully agentic. On the other hand, if the step is high risk (terminations, investigations, comp exceptions, etc.), it should be led by a human, though there may be opportunities for agents to help with the prep work. Most steps, however, fall in between these two extremes and should be built as agents propose + human approves, with full traceability and reversibility.

6. Who on our team should be our forward deployed engineers?

Forward deployed engineers should be big picture thinkers who aren’t stuck in the ways things have been done at your company for the past 10 years. They should be knowledgeable on AI and at least a little bit technical. In an HR department, these people often are your HRBPs or your HR ops-minded folks who have already automated much of their job through spreadsheets and macros. Give these folks real problems to solve. You don’t need them to become certified.

7. How can we keep up with the latest in AI?

You can’t, really. It changes too fast. Subscribe to some newsletters, but know that most things break on X before anywhere else.

That being said, you can get 99% of the value from AI agents with models that are a year old. You don’t need to have the shiniest, most up-to-date model to build valuable agents for your HR team.

Key takeaways from the session:

1. Nobody in the room was asking if they should be using AI. Every company in the room had rolled out a copilot or ChatGPT to the whole company, so the question was more around why it hadn’t changed anyone’s week.

2. HR is being tasked with training the rest of the company on AI, often without adequate budget or authority, so the first question is often “Is this even our job?”

3. Despite #2, HR’s own jobs are often the last to be tackled, as teams are focused on rolling out AI to the rest of the company while still running their own processes (even simple ones like onboarding) through email and spreadsheets.

4. There is fear on both sides: employees are worried about how AI will impact their jobs, and CHROs are worried about what they’ve empowered their teams to build.

5. Most people in the room thought they needed to keep up with the latest in AI. They don’t.

We solve these every day at Varick. Work with us below:

[Thomas LJ @thmslj]

This. I would add that
1. Accountability and liability can't be cheaply designed and one-shot'ted, choose your vendor (or eng team) wisely
2. Accessibility == adoption. 1 channel isn't enough
3. Distribution has always been a moat, how fast your vendor really transforms your processes is a real edge

[Aaron Levie @levie]

AI agent adoption is still very bimodal right now. You have coding and coding adjacent tasks which have taken off, and then everything else.

And even within coding, you have a very wide continuum of adoption patterns, with some likely a small percentage of developers deploying background agents working on projects in parallel, and the rest of everyone else still working with agents 1:1.

Then, there’s the entire rest of knowledge work where real agentic adoption is still very early. The reason for this is that most of the workflows still need to reengineered to work with agents.

This isn’t the same as deploying a chat system for better research efficiency, but instead requires workflows to be rebuilt, data to be wired up in new ways, new practices for accountability and liability of decisions when agents are in the mix, governance and compliance paradigm changes, security upgrades, and more.

We’re still so unbelievably early in what this is going to look like outside of a few categories of work right now. This is why you can basically expect 100X more agent adoption from what we’ve seen so far.

[Thomas LJ @thmslj]

Non-human identities outnumber people 45 to 1 in the average enterprise (Cloud Security Alliance, May 2026). Every agent we ship adds to that count, so each one gets a scoped, expiring identity instead of a borrowed human login.

[vas @vasuman]

If you're a college student looking to be an HR intern for a few weeks (potential to go full-time), send me a DM/email. Must be based in the US, remote ok.

Need you to help us hire our next 50 people.

[vas @vasuman]

Top 25 questions the world's largest C-Suite executives ask us re: AI transformation in the enterprise:

1. Are AI transformation firms more like McKinsey or Palantir?

2. We launched Copilot for everyone to use, but why hasn't it changed anyone's week?

3. We've done training, custom GPTs, enablement. What's next?

4. Where do we even start with AI?

5. How do we choose the first workflow to AI-enable?

6. Should AI be owned by IT or the business?

7. Do we really have to train all 10,000 people in the company?

8. How do we enable everyone to build agents, without ending up with agent sprawl?

9. Why are our token costs increasing?

10. How do we know if using AI is really cheaper than just hiring another person?

11. Which steps should remain human?

12. How do we communicate AI won't replace our team's jobs? (Will it?)

13. Will we own the agents you build for us?

14. What does maintenance look like in 6 months?

15. Will our data leave our environment?

16. What models should we be using? What do we do when they're deprecated?

17. Can we run these agents inside our ERP, or will people have to use a new tool?

18. Do we need to clean up our data first?

19. How long until we start seeing ROI?

20. When you talk about ROI, do you mean headcount?

21. How do we keep growing our business without increasing headcount?

22. Why did our last pilot project die in POC purgatory?

23. Our process documentation doesn't match what our team actually does – how do you know what to automate?

24. How do we keep up with how quickly AI is changing?

25. If AGI becomes a reality tomorrow, where should we point it?

We solve these every day at Varick. Work with us below: