SYSTEMS BY DESIGN
The last 18 months I've spent inside real businesses implementing AI systems. Not demos. Not pilots. Actual production systems that need to work on a Tuesday morning when no one's watching.
And here's what I've learned that most AI content won't tell you:
Agents are impressive. Workflows are reliable. And reliability is what actually scales a business.
Let me break down the org chart.
At the top, you have the C-suite. Executives using AI as a thought partner, a research assistant, a decision-making co-pilot. This is agents at their best — helping smart people think better, faster.
Agents are genuinely brilliant for individual productivity. Writing emails. Building decks. Researching a topic. Thinking through a problem out loud. Think of how people talk about Cursor — it's not replacing engineers, it's making them 10x. The engineer is still essential. The agent just removes the friction around them.
That's the sweet spot for agents right now: making skilled people faster, not replacing the skill itself. YET - we are getting close.
Then you come down to the ground level — the actual work that runs the business. The repeatable tasks. The processes. The stuff that needs to happen 50 times a week without fail.
This is where autonomous agents break down. And where workflows win.
Investment Firm
Real-World use case
We've built both of these at Phlo. Here's what they actually look like.
The Workflow: Investor Onboarding
Before: an associate spending 2–3 hours per investor — chasing emails, collecting documents, manually working through a 100-point ID checklist.
Now:
Investor fills in an online form — consistent, structured, every field captured the same way
KYC provider fires automatically — identity verified against the same standard every time
AI step runs separately — researches the investor, enriches the database, flags if they're qualified and available
Associate opens a record that's already verified, researched, and pre-qualified
Result: minutes instead of hours. 99% completion rate.
That 99% isn't an accident. It comes from structure. Each step does one thing and hands off cleanly.
Give that same job to an agent with MCP and it could change records ask the wrong questions, not properly secure their ID….
The Agentic Approach: Market Research & Deal Intelligence
Instead of a research analyst manually pulling from five data sources — we connect multiple MCPs and APIs and query Opus across all of it. Allowing it to build dashboards find anomalies and talk to data
No fixed steps — it reasons across everything and decides how to approach the problem
Surfaces patterns, anomalies, and trends a human analyst would take days to find
Can write its own Python scripts or build Excel models to process and visualise the data
Then uploads the finished output directly into the document space
This is where it differs from the workflow's. It can actually act — but only within its own environment. It's not touching the live database, it's not modifying investor records. It has a sandboxed space to think, build, and produce. The output lands where you need it. The rest of your systems stay untouched.
That's the meaningful distinction. Not "can it do things" — but "what can it touch." A workflow's AI step is contained by design. An agentic approach has more freedom, but that freedom is bounded by its environment, not by someone watching over it.
One owns the process. One amplifies the analyst. Both have AI inside them.
Harvey AI is worth $11 billion. And they've figured this out.
Harvey is an AI platform built specifically for law firms. In under two years they've hit $190M ARR and are currently raising at an $11B valuation.
I'll be honest — I know nothing about law. But I do know one thing about it, and I learned it from a TV show. Mike Ross in Suits said it best: "Law is a precise endeavour."
In law, being close enough isn't close enough. Accuracy isn't a nice-to-have — it's the entire job.
That is why Harvey has two products. An Assistant — where lawyers use AI as a thinking partner, always with a human reviewing the output. And Workflows — where routine legal processes are automated with far greater reliability and scale. Over 100,000 lawyers across 1,000+ firms are using it today.
The assistant helps lawyers think. The workflow scales their impact.
They didn't try to replace the lawyer with an agent. They broke the job into its component parts and figured out which parts a workflow could own.
Build the dishwasher first.
Map your highest-volume repeatable tasks. Pick one. Build a workflow around it. Measure the success rate. Then do it again.
That's how you actually implement AI in a real business.
See you next week.
— Daniel
P.S. I'm not saying agents aren't good — they're genuinely excellent. As a co-pilot, a thought partner, a way to get through your work faster, they're one of the best tools we've ever had. But if you truly don't want to do the task at all — if you want something running while you sleep, without you — that's a workflow. Know which one you're building.