Conversational AI degrades when nobody maintains it.

A bot in production is not "finished". It is only aligned with a given moment. Then customers change their words, offers move, exceptions multiply, and quality drops off.

Conversational AI is a living asset. It feeds on new examples, corrections and business follow-up. Without maintenance, the project does not fail all at once. It wears down, then gradually becomes less useful.

Go-live day proves nothing about robustness

A bot can succeed at launch and degrade within the following weeks. The first conversations often look like the test scenarios. Then come the real phrasings, the shortcuts, the typos, the new products and the hybrid requests.

The bot keeps answering. It simply understands less well. The difference between "it runs" and "it helps" shows up in escalations, rephrasings and abandons. In sectors where vocabulary moves fast, this drift appears very early.

Ask: "What is the understanding rate per intent over the last three months, and which way is it trending?" Also require tracking of the escalation rate to an agent, intent by intent. If understanding falls while transfers rise, the bot is not ageing well.

Check a simple contractual criterion: a monthly performance report with history, not an isolated score. Common trap: a flattering global dashboard hides a critical intent in free fall, with the classic symptom of repeated generic answers.

Customer language changes faster than project plans

Customers do not use the internal taxonomy. They pick up the words of campaigns, social media, customer service and the irritants of the moment. They abbreviate, mix, twist, then invent phrasings the project committee never wrote.

An unmaintained model ends up recognizing the structure of the sentence, but no longer the real intent. The drift is slow. It does not interrupt the service. It increases rephrasings, catch-all answers and the fatigue of a customer who has to repeat the request.

Ask: "Can you show me the misunderstood verbatims from the last 30 days, ranked by frequency?" Then require that those exact sentences feed the training set. Project-written rephrasings are not enough, because they smooth out real usage.

Verifiable criterion: a monthly review loop of emerging expressions, with a handling decision and an update date. Precise trap: if agents say "customers don't talk the way they used to" and nothing is added to the corpus, the bot piles up silent false negatives.

The offer changes, the intents must follow

Every change to the offer creates blind spots. A new guarantee, a modified journey, a pricing condition, and the intent tree becomes incomplete. The bot then answers with old logic, sometimes coherent on the surface, but wrong on the substance.

The risk is not limited to a missing intent. It also exists when two close intents are badly prioritized. A customer who wants to amend a contract can be sent to an obsolete FAQ, simply because the model was not revised after the business release.

Ask: "Who validates the conversational update after each change to the offer, and within what timeframe?" Require a target delay between the business publication and the bot update. Without that delay, the assistant becomes a lagging mirror of the catalogue.

Check a technical criterion: every change to the offer must trigger a review of the impacted intents, with traceability. Common trap: a new pricing rule published in production without a bot update, visible as answers that are right on the commercial substance but wrong on eligibility.

An unmaintained bot does not break down. It gradually becomes less accurate.

Maintenance must be budgeted as a recurring capability

The real issue is not the launch cost. It is the cost of sustaining performance. Annotation, intent correction, example enrichment, retraining, quality control and business arbitration all take time, tools and available people.

Without a recurring budget line, these tasks come after the visible emergencies. The bot is then only corrected after an alert, once the drift has already done its damage. Maintenance becomes reactive, therefore more expensive and less effective.

Ask: "What annual budget is reserved for maintenance, separate from the initial build?" Also ask who decides between business correction, data adjustment and operational priorities. Without a clear owner, anomalies stay in the queue.

Useful contractual criterion: a fixed fee or a volume of hours dedicated to conversational maintenance, with a defined retraining frequency. Precise trap: if everything sits in a single project envelope, the first budget cut removes maintenance, then quality degrades without a sound.

The right indicators are simple and unambiguous

Understanding rate and escalation rate must be tracked over time. Not only on average. Per intent, per channel, per customer segment and on emerging phrasings, that is where drift becomes visible.

A dashboard centred on conversation volume is falsely reassuring. Volume can stay stable while relevance collapses. You also need to look at successive rephrasings, abandons, off-topic answers flagged by agents and the intents that concentrate the failures.

Ask: "What threshold triggers a review, and who receives the alert?" Require an explicit mechanism, for example a continuous drop in understanding over two measurement periods or a sustained rise in escalations on a critical intent.

Check a technical criterion: the alert must be automatic, timestamped and retained. Common trap: a threshold exists on paper, but nobody looks at it. The symptom is clear: the numbers degrade for weeks before a human notices.