Glossary
Production AI
Production AI is a system people depend on to do their job, which means it carries expectations about reliability, support and predictable cost that no trial is ever asked to meet.
In plain terms
At some point people stop trying the tool and start relying on it. Their work does not get done if it is unavailable, and they notice immediately when it behaves differently. That is production, and the distinguishing feature is dependence rather than any technical property. A tool can be in production on a laptop and out of production on expensive infrastructure; what decides it is whether somebody is stuck without it.
Why it matters
Because dependence arrives quietly and the obligations arrive with it whether or not anybody has accepted them. Nobody sends a note saying the pilot is over, so a tool crosses into production while it is still being treated as an experiment: no owner, no support route, no plan for the day the vendor changes something. The gap between when people start depending on a thing and when the organisation starts supporting it is where most of the unpleasant surprises in this subject live.
How it works
Somebody has to own it, by name. Not the team that championed it and not the department using it, but a person who is accountable when it stops working and who is expected to notice a change before the users do. An unowned tool that people depend on is the most common shape of this problem, and it stays invisible until the first time it fails.
There has to be a route for when it goes wrong that is not asking around. Where a user reports a problem, who looks at it, what happens if the answer needs the vendor. That route is cheap to establish while nothing is broken and impossible to invent calmly at the moment it is needed.
The behaviour will change underneath you, which is the property that distinguishes this from most software. Vendors update models, adjust safety behaviour and retire versions, and output that was fine last month can shift without anybody on your side changing anything. Production means somebody would notice, which in practice means holding a small set of examples you check rather than waiting for a user to report that it feels worse.
Cost becomes a running concern rather than a purchase. Usage rises as adoption spreads, a heavier feature changes the rate, and a bill that was comfortable during a trial with a few people is a different number across a department. Knowing what drives the figure before it grows is the difference between managing it and reacting to it.
The quiet requirement is a way back. What happens if this is withdrawn, becomes unaffordable or stops suiting the work: whether the material can be exported, whether the process can be done another way, how long it would take. Production without an exit is a dependency on somebody else's decisions, and the cost of that is invisible right up until it is the only thing that matters.
When the obligations start
Seen in the wild
An assistant that a team now drafts everything in, where an unavailable morning means work stops rather than slows.
ChatGPTInternal search that people rely on to find answers, where a bad result is a wrong answer given to a customer rather than a mild inconvenience.
GleanAn automation running unattended over incoming records, where a silent failure is discovered by whoever was expecting the output.
n8n
Common misconceptions
People assume
It is in production when we announce it.
In fact
It is in production when somebody is stuck without it, which is usually well before any announcement. The announcement is when support was arranged; dependence started earlier. The interval between the two is exactly the period in which the organisation carried an obligation nobody had accepted.
People assume
Production is about scale.
In fact
It is about dependence. Three people who cannot do their job without a tool is production; three hundred people who would shrug and use something else is not. Volume changes what production costs and what it requires technically, but it is not what makes something production in the first place.
People assume
It works, so the hard part is done.
In fact
Working is the entry condition rather than the finish. What follows is keeping it working while the vendor changes it, while usage grows, while the people who set it up move on. Most of the effort in production AI is spent after the point where somebody first said it works.
Telling them apart
Production vs Pilot
Production
People depend on it. Owned, supported, watched, with a way back.
Real work under a bound, with the option to stop still intact.
Ask what happens if it is unavailable tomorrow. An inconvenience is a pilot; stopped work is production.
Questions
- How do we know we have crossed into production?
- Ask what happens if the tool is unavailable tomorrow morning. If the answer is that people work around it, you are still in a pilot; if work stops or a customer waits, you are in production regardless of what the project is called. The label follows dependence rather than deciding it.
- What is the minimum a production tool needs?
- A named owner, a route for reporting that it is broken, somebody who would notice if the output quality changed, and a rough answer to how you would leave. Those four are modest and they are what separates a supported tool from a popular one, which is the distinction that matters when something fails.
- Why does output change when we changed nothing?
- Because the vendor changed something. Models are updated, retired and adjusted on the vendor's schedule, and behaviour shifts underneath you in ways release notes do not always describe. Keeping a small set of examples you re-run periodically is the cheapest way to see it happen rather than hear about it from a user.
- Does production mean self-hosting?
- No. A hosted tool can be production and a self-hosted one can be an experiment. What self-hosting changes is who controls the timing of changes and who carries the operational work, which are real trade-offs, but dependence is what defines production and it is independent of where the thing runs.
Key takeaways
- Dependence defines production, not scale and not an announcement.
- It crosses over quietly, so the obligations usually arrive before anybody accepts them.
- Behaviour changes underneath you; somebody has to be positioned to notice.
- Cost becomes a running concern once adoption spreads past the trial group.
- Production without a way back is a dependency on somebody else's decisions.
Last checked July 2026