OWNING vs RENTING AI: A REALITY CHECK FOR GCs.

The two categories of AI you're buying

Treat AI tools as two different categories that happen to share marketing language:

Category A — Generic intelligence-as-a-service. ChatGPT Enterprise, Claude for Work, Microsoft Copilot. These are general-purpose models you query for arbitrary writing, analysis, summary. The model itself is the product. You don't need it to know your projects deeply.

Category B — Workflow automation that uses AI. RFI triage, daily log synthesis, change-order risk scoring, blueprint audit. The AI is a component; the actual product is "this process happens automatically." You need it to know your projects.

Per-seat SaaS makes sense for Category A. It usually doesn't for Category B, and that's where most GC AI spend is going.

Why per-seat works for Category A

You don't customize ChatGPT to your company. You query it. Each user gets value individually; the value scales with users; the cost scales with users. That's the right shape of pricing for that shape of value.

Microsoft, OpenAI, Anthropic spend hundreds of millions training models. They have to amortize that across millions of seats. The unit economics force per-seat pricing. It's reasonable.

Why per-seat is the wrong shape for Category B

Workflow automation has a different value curve. Once the system exists, the marginal cost of one more user querying it is roughly zero. The expensive part was the build — ingesting your projects, fitting the rules to your shop, training the classifier. After that, ten users or a hundred users costs the same to run.

Per-seat pricing for workflow automation is the SaaS vendor's preference, not the customer's. It maps the customer's cost to a curve that the customer doesn't actually incur. That's where the margin lives.

The five-year math, written out

Take a typical 30-user GC adopting an RFI AI tool at $50/seat/month. That's $18,000/year. Five years: $90,000.

Now consider building the same workflow as a custom Marapone-style system: ~$9,500 one-time. Five years: $9,500 (plus optional ~$6,000/year retainer if you want one — call it $40,000 over five years with full support).

That's a $50,000-80,000 difference for the same workflow. For a 30-person GC, that's enough to fund the next module — or just go straight to the bottom line.

And the gap widens with headcount. At 100 users, the SaaS bill is $60,000/year; the custom build is still $9,500.

When renting is genuinely the better call

Three honest cases where renting wins:

1. You're under 5-10 users and the SaaS price is small enough to absorb without thinking.
2. You want to be live this week, not in two weeks.
3. The workflow is genuinely generic — what you'd build for yourself wouldn't differ much from what's already on the market.

If you fit any of those, sign the SaaS contract. Don't overthink it.

When owning starts to pay

Four signals that you've crossed the line:

• Your projected annual SaaS spend on AI tools is past $20-30k.
• You're customizing each tool with your own templates, rules, and scripts anyway.
• Your data needs to stay on your network — drawings, owner emails, fee schedules.
• You want the AI to ingest from sources the SaaS vendor doesn't connect to.

Two of those four and the conversation is worth having. Three or four and it's a no-brainer.

The asset-on-the-balance-sheet question

One framing GCs underestimate: a custom build is a capital asset. It's depreciable. It's something you own. A SaaS subscription is an operating expense that disappears the moment you stop paying.

If your CFO is the kind that thinks about ratios, this matters. If they're not, it still matters — just less visibly.

A simple decision rule

Before signing the next AI subscription, ask: "If I scale this to 50 users for 5 years, what does the bill look like, and what could I have built instead?"

Sometimes the SaaS still wins. Often it doesn't. The math just hasn't been done.