How the AI's expertise is layered
Every trade area has a brain and a bookshelf, at two scopes — DarkBird's baseline and your org's layer. Here is exactly what the AI reads before it asks a question.
When the AI asks a scoping question or recommends a card, it is not improvising. It has just read a stack of expertise assembled for that exact moment — and understanding the stack explains both why the AI is useful on day one (before you've taught it anything) and why it gets meaningfully better when you invest in the Expert Center.
The brain and the books
Every expert domain is two separate documents with strictly separated jobs:
| Expert Profile ("the brain") | Knowledge ("the books") | |
|---|---|---|
| Contains | How a senior practitioner thinks: what they always ask first, what makes them nervous, estimation instincts, the failure modes that blow estimates | What the field knows: technology overviews, real vendor landscapes, deployment scenarios, labor benchmarks and gotchas |
| Product names? | Never — a profile should still make sense in three years even if every product changes | Yes — vendor tiers, product characteristics, and manufacturer quirks belong here |
| Changes when… | Your team's judgment evolves | Products, specs, or field conditions change |
The split is deliberate: when a manufacturer ships a new product line, the books update and the brain is untouched. When you decide estimators should always ask about ceiling type, the brain updates and no facts move. The two quote each other's numbers — a profile's "45–60 minutes per drop" matches its knowledge doc's effort benchmarks — but every fact has exactly one home.
The two scopes: DarkBird's baseline and your layer
Both documents exist at two scopes, and the AI merges them at interview time:
- DarkBird's global layer — 30 expert domains maintained by DarkBird, spanning the trades integrators sell: surveillance, networking, access control, AV, cabling, wireless, fire alarm, DAS, data center, building automation, power protection, residential smart home, healthcare communications, retail rollouts, and more — plus cross-cutting experts (Project Management, Compliance, Service Contracts) that join every interview. This is the industry baseline your org inherits on day one, no setup required.
- Your org's layer — everything you add in Settings → Expert Center: knowledge sources per category, per-category Expert Profile additions, and your org-wide Expert Profile. Your layer adds to and overrides the baseline with your standards — your labor rates, your preferred brands, your house rules.
Inheritance rule: your org uses DarkBird's baseline expertise until you customize it. When you save your own org-wide Expert Profile, yours takes over as the cross-cutting voice. Your per-category content always layers on top of the global baseline for that category — you never start from a blank page.
What one interview turn actually reads
Before the AI responds to a single interview answer, the server assembles, in order:
- The org-wide expert baseline — cross-trade estimator judgment (unit economics, scope control, phasing), yours if customized, DarkBird's if not
- The relevant domain experts — for a camera project: the surveillance brain and books, plus cabling and networking when the scope pulls them in; domains activate based on your catalog and the project's content, so unrelated experts never cost you a token
- The always-on experts — Project Management, Compliance, and Service Contracts join every turn regardless of trade
- Your org knowledge and wiki — spec sheets and notes you've added per category, plus the org wiki that review insights maintain automatically
- Your calibration multipliers — so recommended hours already reflect how your actuals have been running
That stack is why two orgs asking about the same 40-camera project get different questions and different hours: the baseline expertise is shared, but your layer — knowledge, profiles, and calibration history — makes the output yours.
Highest-leverage order for a new org: confirm your labor rates in a Knowledge Category, then write the org-wide Expert Profile (or run the Interview Expert wizard), then let calibration accumulate from real project reviews. Each layer compounds the others.