process
The Five Gates to Production
A demo has to be useful. Production AI has to pass all five gates.
Structure
- Gate 1: Useful
- Does it do something a person actually needs done? Dies here: a solution looking for a problem.
- Gate 2: Reliable
- Does it behave the same way on Tuesday as it did in the demo? Dies here: non-determinism nobody budgeted for.
- Gate 3: Governed
- Who owns it, who approved it, what may it touch, who is accountable when it is wrong? Dies here: nobody will sign.
- Gate 4: Integrated
- Does it live inside the systems where work happens, or does it require going somewhere else? Dies here: the tool nobody opens.
- Gate 5: Valuable
- Did a measurable operational outcome change? Dies here: it cannot be defended in a budget hearing.
A demo has to clear one gate: useful. Production AI has to clear all five, in order, and most of the mortality happens where nobody is watching for it.
Most pilots die between gate two and gate three, and it is rarely a technology problem. A model that works is not the same thing as a model somebody is willing to be accountable for. Pilots that are called successful but never scale usually died at gate four: they never lived inside the systems where the work actually happens, so nobody opens them after the pilot period ends.
Deployed AI that quietly disappears the next budget cycle died at gate five. Gate three used to be treated as the hard gate, because governance was the unfamiliar step. It no longer is. Governance templates are now free and repeatable. A defensible answer to what operationally changed is not, which is why gate five, not gate three, is now the gate that decides whether something survives.
GAO reported that of 282 federal generative AI use cases logged for 2024, only 41 percent were implemented or operational; 56 percent remained in acquisition or development.