Three sourcing models. Each is right for some buyer. The under-discussed move is treating the vendor as a multi-year operational partner, not a one-time procurement.
Chapter 2.48 min readHow to evaluate it
The landscape map in 2.3 answers the category-level question. This chapter handles the sourcing model within your chosen row.
Each model trades off differently. None is universally better; the right answer depends on what your company has and doesn't have.Model 1
Build · DIY on RAG infrastructure
Pick a vector database, pick a model provider, wire it with your engineering team, and author content in-house. A fully owned, fully controlled stack, with the full operational burden.
When build is right
An in-house ML team you want to use. A multi-quarter runway. An unusual content shape needing custom architecture. You want full control. Some large manufacturers pick build for the right reasons.
When build is wrong
A 1-50-person team without AI/ML staff. Time-to-value in months, not quarters. Most mid-market built-environment manufacturers.
Model 2
Buy · an enterprise platform
License Glean, Coveo, or Algolia, engage their professional services, and bring your own IT team. The platform handles search-and-retrieval; your team handles content, integrations, and operations.
When buy is right
An enterprise with mature IT and budget for enterprise software. Generic content. A 3-6 month implementation runway. Other use cases (employee search) that improve the unit economics.
When buy is wrong
A mid-market cost mismatch. No IT team for a 6-month implementation. Content that needs vertical-specific authoring the platform doesn't produce.
Model 3
Hire · a managed AI service
Engage a vendor whose product is a managed AI customer support service. The vendor handles content authoring, monitoring, monthly reporting, and tuning. You bring the source material and the brand.
When hire is right
Real product documentation, no AI/ML team. Time-to-value in weeks. Content vertical-specific enough that authored content is a moat. A multi-year operational relationship is acceptable.
When hire is wrong
You want to own the AI implementation. An enterprise with budget for a platform. Your trade isn't one the vendor has authored content for, so they would build from scratch on your account and lose the managed-service value.
The under-discussed move
Evaluating the vendor as a multi-year partner
This is the move most evaluations skip, and it matters most for the hire path, because the vendor relationship is the architecture. It isn't a software contract you can rip out. It's a multi-year operational relationship where the vendor's long-term posture matters as much as the current product.
Six dimensions · click to explore each
Transparency about limits
Does the vendor name what they can't do? Vendors who hedge during evaluation are the same ones who go quiet about limitations after the contract signs. Good vendors tell you up front exactly which query categories they'll fail on and which content shapes won't fit.
That honesty is the signal.
Roadmap visibility
Does the vendor share what's coming? Multi-year relationships compound on the roadmap. A vendor who treats their roadmap as a proprietary secret signals you'll find out about changes when they hit production. A vendor who shares context signals a co-design posture.
Vertical depth
Does the vendor speak your trade? Not generally. Specifically. Can they name the top three failure modes installers in your category hit? Can they discuss which content shapes they handle well in your vertical and which they don't?
Vertical depth either exists or it doesn't. The conversation reveals it within ten minutes.
Trade-specific experience
Does the vendor have demonstrated experience in industries with requirements like yours? You want vendors who've done this work in your trade or one structurally adjacent (security to fire to access control; HVAC to building automation; industrial sensors to instrumentation). Direct reference customers in your trade are best; adjacent-trade experience with transferable content patterns is the next-best signal.
Ask the vendor what they'd need to learn to serve your trade versus what they already know.
Onboarding investment
How seriously does the vendor treat the first 90 days? Onboarding is where the relationship is most exposed. A vendor casual about onboarding planning ("we'll figure it out as we go") is signaling operational discipline you should worry about. A vendor with a detailed week-by-week plan, named owners per workstream, and explicit pilot success criteria is operating at the level the relationship requires.
Multi-year commitment posture
Is the vendor optimizing for a long arc or a quick win? Watch contract terms, renewal posture, and the ease of getting your data out. Vendors who make exit difficult are relying on switching costs. That's not the relationship you want with someone running operations for you.
The question every managed-service vendor should answer
"If we wanted to switch off your service in 18 months, what would that look like?" Vendors who answer directly (export format, transition timeline, what stays with you) operate from a partner posture. Vendors who deflect operate from a lock-in posture.
Decision
How to decide
Pick the model that matches the resources you actually have, not the ones you wish you had. Then, if you hire, evaluate the vendor on the six dimensions above.
The most expensive project failures come from sourcing-model mismatches. Get sourcing right early and the rest becomes tractable.
What good looks like
A CS director who has made the sourcing decision well:
Has placed their company on the landscape map honestly
Has chosen build, buy, or hire for the right reasons, not the easiest
If hiring, has evaluated the vendor across the six dimensions
Can articulate to the board why this sourcing fits this company
Has asked the "what does exit look like" question and used the answer