The maintenance is the work — why AI is the first system that drifts under you
There is a category of software that, once shipped, sits there. The accounts-payable workflow you wrote in 2018 still runs the same way today. The Stripe webhook handler from 2021 still does what it did at deploy time. Software, in the traditional sense, is stable until somebody changes it.
AI systems are not like that. They drift on their own.
Three places the drift comes from
The first is the model itself. The vendor ships a new version, your AI agent picks it up at the next API call, and the new version handles edge cases differently than the old one. A prompt that reliably extracted the customer's order number from a chat message now occasionally returns the order number plus a polite explanation about why it's confident — and the downstream system that expected a clean string breaks. Not catastrophically. Quietly. The error rate just creeps up by two percentage points and nobody notices for a month.
The second is the data. Your customers start using new vocabulary. Your team adds a new product SKU. A new business line opens that the AI was never trained on. The model itself has not changed; the world it is reading has. The agent's accuracy on the new shape of inputs degrades because nothing in the system was built to notice the shape of inputs is different.
The third is the prompt. Your operations team learns the AI well enough to know what it gets wrong, and they want to add a few clauses to handle the new edge cases. The prompt grows. The original three sentences become twelve. Eventually the prompt is fighting itself, and the agent that was making one specific decision well is now making three loosely-related decisions poorly.
None of these failure modes show up at deploy time. They all show up two to six months in, after the team has stopped looking at the agent's outputs because it was working.
Why the relationship has to change shape
The traditional consulting shape is: scope, build, ship, walk away. That shape works for a static system. It does not work for a system that drifts.
For AI work, the right shape is build plus operate. The build is the upfront engagement: scope the problem, integrate the data, ship the agent, measure adoption. The operate phase is the relationship that keeps the agent healthy: monthly accuracy checks against a held-out evaluation set, model swaps when the vendor ships an upgrade that changes behavior, prompt rewrites when the data drifts, occasional retraining when the prompt has accumulated too much drift to be salvaged.
This is not a make-work retainer. It is the difference between an agent that performs at deployment quality two years later and an agent that quietly stops working in month three.
What the operate phase actually looks like
A weekly automated check that runs the agent against a fixed set of historical inputs and flags any output that has drifted from the previous run. A monthly human review of a small random sample of real production outputs to catch the failure modes the automated check misses. A quarterly conversation with the team that uses the agent to surface the new edge cases they have started seeing. A regression test suite that runs every time the underlying model is updated.
These are unglamorous. They are also the difference between an AI program that compounds value over years and an AI program that becomes a maintenance burden the team eventually shuts off.
The economics make this obvious
The unit economics of AI work favor the operate phase, not the build. The build is bounded — it ships once. The operate phase is where the value compounds, because every month the agent stays accurate is another month of recovered hours, faster response times, fewer errors. An agent that runs at 95% accuracy for two years generates roughly 24 times the value of an agent that runs at 95% accuracy for one month before drifting to 60% and getting silently abandoned.
The right business model reflects this. The build phase is fixed-quote, because the scope is bounded. The operate phase is monthly retainer, because the work is open-ended and the value is recurring.
The shift in expectation
The teams that get the most out of AI in 2026 and beyond will be the teams that stop treating it as a project and start treating it as an ongoing relationship with a system that drifts. Not because they want to spend more on AI, but because they have already learned, through a few quietly-broken agents, that the alternative is worse.
The maintenance is the work. Build the agent. Then operate it.