Startup Talks #1 · Startup Madeira
Judgment is
the moat.
When almost anyone can build almost anything, the scarce skill is deciding what deserves to be built.
Victor Lipov · victorlipov.com01
The question everyone asks
"What will AI
replace?"
Maybe that's the wrong question. The interesting one isn't what gets replaced — it's what becomes scarce when building is nearly free.
Judgment Is the Moat02
The shift
Building just got cheap.
AI collapsed the cost of production. One person can now ship what used to take a whole team — a feature, a model, a company's back office — in a weekend.
Judgment Is the Moat03
The catch
Easier execution doesn't make
a better business.
When everyone can build almost anything, building stops being the advantage — everyone has that now. The bottleneck moves upstream, to the decision about what to build at all.
Judgment Is the Moat04
The thesis
AI doesn't replace
expertise.
It removes the excuse for not having it. The tool amplifies whoever holds it: give it to someone with judgment and it compounds — give it to someone without and it just ships the wrong thing faster.
Judgment Is the Moat05
Who's saying this
Fifteen years on the business side
of the systems a company runs on.
Credit risk, controlling — the decision systems, not the front-end. I was never the person shipping the most code; I was the one deciding what the system should conclude. Now I build and run those systems as a one-person business. That seat — deciding what's correct — is what this whole talk is about.
Judgment Is the Moat06
The work that's left
The judgment behind questions like:
- What is worth building?
- When to delegate to AI — and when a human must stay responsible?
- When to continue, pivot, or stop?
- A promising product, or an impressive demo?
Judgment Is the Moat07
Where the value moved
From production
to judgment.
The work used to be making the thing. Now the thing is cheap — so the value is upstream: knowing which problem is real, what "correct" looks like, and when to stop. Taste, not throughput.
Judgment Is the Moat08
The founder's trap
An impressive demo
is not a business.
AI makes the shiny thing almost free — a demo that wows in a meeting and does nothing on Monday. The work is the unglamorous part: the edge cases, the audit trail, the decision no one sees. Do the work, not the shiny thing.
Judgment Is the Moat09
A small piece of evidence
A €15 answer.
The question was the work.
One small AI job surfaced that almost a thousand leads in a client's system had never been touched. The AI cost fifteen euros. The value was knowing that was the question worth asking — and that part you can't download.
Judgment Is the Moat10
What this looks like in practice
I run production systems
as a one-person business.
Not by using AI as a coding assistant — by directing it. Two real, running systems follow — and one turns out to be the backbone of the other. In each, watch where the machine does the work, and where I deliberately kept a human in the seat.
Judgment Is the Moat11
Case study · 01
A Bid Desk That Shows Its Work
Score, approve, control — auditable public-procurement triage for a manufacturer
Victor Lipov · Case studies12
Collect
Score
Approve
Control
A bid desk that
shows its work.
Public-procurement triage for a manufacturer. A machine narrows the field; a human owns every bid or no-bid, and can defend each one.
Victor Lipov · Case study13
The problem
Someone has to read every tender.
Hundreds of public tenders are published every month. A person has to read each one and decide, by hand: bid or skip? Miss a fit and you lose the contract; chase a bad one and you burn a week on a losing bid.
A Bid Desk That Shows Its Work14
The scale, one running month
One month, live: 109 public notices in, 103 matched to the catalog, 3 worth a human's time. The hard part is not reading them, it is defending the skips.
A Bid Desk That Shows Its Work15
The instinct
"Let the model just decide."
The obvious build is one LLM call per tender: read it, output bid or no-bid, done.
Why that is the wrong build
A model that silently rules out a public contract you never saw is a liability, not a filter. When you skip a €2M bid, "the AI said no" is not an answer you can give.
A Bid Desk That Shows Its Work16
The design
Score, approve, control.
Score
Approve
Control
Three layers. The machine narrows, a human decides, the system records why. It is how I built credit scorecards for fifteen years, pointed at tenders.
A Bid Desk That Shows Its Work17
How a tender moves
The machine narrows and disqualifies. A human owns the bid decision. Every path ends in an audit record.
A Bid Desk That Shows Its Work18
Score — narrow the field
Rules first, model last.
- Procurement-code rules — tender codes matched to the catalog
- Keyword rules — brand, IP-rating, control-gear patterns
- An LLM judge — reads the notice, ranks fit, explains itself
The model is the last word, not the only one.
A Bid Desk That Shows Its Work19
Approve — the human gate
The machine only ever says no.
AI proposes, a human disposes
The system auto-disqualifies the clear noise. Anything that might fit goes to a person. It never auto-approves a bid, the accountable "yes" stays human.
That gate is not a limitation of the system. It is the system.
A Bid Desk That Shows Its Work20
Control — the part almost nobody builds
Every decision, on the record.
Append-only: which rules fired, the thresholds frozen at decision time, the evidence, and who approved. A real decision from the running system, anonymized.
A Bid Desk That Shows Its Work21
What I did not build
Where I stopped.
- No auto-submission — the system never files a bid
- No ERP wiring until the desk has earned trust
- No AI signature — the machine never owns a decision a person answers for
Restraint is the engineering: knowing where the machine stops.
A Bid Desk That Shows Its Work22
Software that makes a call
should be able to defend it.
The scoring is cheap. The trail you can inspect is the product — and it stands on a catalog the next system serves, so the two compound.
Victor Lipov · victorlipov.com23
Case study · 02
One Rule, Every Variant
Thousands of product codes from a few dozen rows, and read back again
Victor Lipov · Case studies24
Axes
Grammar
Code
Decode
One rule,
every variant.
A product configurator that generates thousands of valid article codes from a few dozen database rows, and reads any code back into a configuration.
Victor Lipov · Case study25
The problem
One product line, thousands of variants.
Every size, housing, LED module and sensor combination is a real, orderable product with its own article number. Maintain that list by hand and it's wrong the day you add an option.
One Rule, Every Variant26
The instinct
"One row per finished product."
Enumerate every variant in the database, one record each.
Why that build rots
Add one option and the table doubles. It drifts out of sync, nobody can maintain it, and a pasted code still cannot be turned back into a configuration. The data model is wrong, not the size of the table.
One Rule, Every Variant27
The design
The article number is a positional grammar, not a stored string. Each segment resolves from a catalog row, so the code space is the product of the axes, computed on demand.
One Rule, Every Variant28
And it runs in reverse
Paste any article number and the same grammar decodes it: product line, size, and every component, straight from the code. Generation and lookup are one rule read two ways.
One Rule, Every Variant29
Why it compounds
The catalog and codes are exposed as an API — and that API is the product source of truth behind the tender bid desk. That wasn't luck: I modelled it as an API so the next system could stand on it. One system's output becomes the next one's backbone.
One Rule, Every Variant30
Model the axes,
not the combinations.
Get the data model right and the hard things get cheap: thousands of codes from a handful of rows, decodable, and reusable by the next system you build.
Victor Lipov · victorlipov.com31
If you remember one thing
Deterministic where you can.
AI where you can't.
Deterministic scripts where possible; AI only where the work genuinely can't be reduced to rules. Scripts are predictable, testable, and cheap to run — reach for the model where judgment is irreducible, and keep it on a leash when you do.
Judgment Is the Moat32
The one seat that stays human
AI proposes.
A human signs.
You saw it in both: the machine narrows, scores, drafts — and stops at the decision someone is accountable for. In anything that matters — money, health, a real call — that seat does not transfer to a model. Build so the human owns it, and the gate becomes the most valuable thing you have.
Judgment Is the Moat33
victorlipov.comAI raises what
everyone can build.
Your judgment decides what's worth building. That's the moat — and it's the one thing the tools can't hand you.
Victor Lipov · I build systems, then I run them.
Victor Lipov · Startup Talks #134