A company gets excited about AI, stands up an internal team, and spends a year or more rebuilding a version of something a vendor already runs at scale. The team may even ship something that works. The spend still does not pay back.
The question worth asking before funding that team is not whether the team can build the thing. Most competent engineering teams can. The question is whether building is the right call for this company, on this workload, right now.
The three conditions
Building an AI capability in-house is only defensible when all three of the following hold at once.
- Volume. The process runs often enough, at high enough cost, that even a partial win covers the cost of the team that built it.
- Patience. The organization can fund the effort past the first one or two quarters without pulling the plug before it proves out.
- Bench. An engineering team exists that can build and maintain the system without starving the core business of its attention.
Missing any one of the three does not just slow the project down. It produces a specific, predictable failure.
What happens when one condition is missing
Three combinations turn up in practice, each missing exactly one condition.
- Volume and patience, no bench. The team lacks the depth to ship and maintain the system. The project stalls into a permanent, half-built backlog.
- Volume and bench, no patience. The team can build it, but the organization runs out of attention before it proves out. The project is killed in the third or fourth quarter, just before it would have turned the corner.
- Patience and bench, no volume. The team builds something that works. The underlying process was never expensive enough to justify the build, so it never pays back the investment.
Only the center, with all three conditions present at once, makes building the right call. Checked honestly against this list, most companies are missing at least one.
Buy by default, except for what makes you different
For undifferentiated work (support queries, invoice matching, quotation follow-up, HR and IT helpdesk requests), a missing condition means: buy. A vendor's product has already been proven on volumes most companies will never see on their own, and buying it does not cost the company anything it could not otherwise sell.
The carve-out is differentiated work: pricing logic, quality judgment, scheduling rules, anything that is the actual basis of competition. There, buying the workflow off the shelf hands the same capability to every competitor paying the same subscription fee, even at low volume. Owning the data and the evaluation layer for that specific process still matters, independent of whether the three conditions above are met.
Two free actions before you decide
Neither of these costs anything beyond time, and both should happen before any build-or-buy conversation goes further.
- Measure the baseline before touching AI. Know what the target process costs today (time, error rate, headcount, whatever the unit is) before evaluating a build or a buy. Without a baseline, there is no way to judge whether either choice worked, and no way to judge a vendor's claim.
- Ask any vendor whose workload proved the product. What volume was it proven on, and what was the before-and-after. A product proven on a real, high-volume workload is a different thing from a product proven on a demo. No credible answer to that question is a reason to keep looking.
Frequently asked questions
How do I know if my company should build an AI capability in-house?
Check the workload against three conditions: volume high enough to cover the team's cost, patience to fund the effort past the first one or two quarters, and an engineering bench that can build and maintain the system without pulling focus from the core business. Building is only the right call when all three hold at once.
What happens if only two of the three conditions are met?
Each missing condition produces a specific, predictable failure: no bench stalls the project into a half-built backlog; no patience gets it killed in the third or fourth quarter, just before it turns the corner; no volume means it works but never pays back the investment.
Should undifferentiated work like invoice matching or helpdesk requests ever be built in-house?
Generally no. A vendor's product has already been proven on volumes most companies will never see, and buying it does not cost more than the company could otherwise sell. The exception is differentiated work: pricing logic, quality judgment, scheduling rules, where owning the data and evaluation layer still matters even at low volume.
Talk to us
Mirai360 AI helps manufacturing leaders run this test before they commit budget or headcount to a build. Tell us the workload you are weighing. Message us on WhatsApp: Tell me more.
FAQ
- How do I know if my company should build an AI capability in-house?
- Check the workload against three conditions: volume high enough to cover the team's cost, patience to fund the effort past the first one or two quarters, and an engineering bench that can build and maintain the system without pulling focus from the core business. Building is only the right call when all three hold at once.
- What happens if only two of the three conditions are met?
- Each missing condition produces a specific, predictable failure: no bench stalls the project into a half-built backlog; no patience gets it killed in the third or fourth quarter, just before it turns the corner; no volume means it works but never pays back the investment.
- Should undifferentiated work like invoice matching or helpdesk requests ever be built in-house?
- Generally no. A vendor's product has already been proven on volumes most companies will never see, and buying it does not cost more than the company could otherwise sell. The exception is differentiated work: pricing logic, quality judgment, scheduling rules, where owning the data and evaluation layer still matters even at low volume.