HomeinsightsApplication Performance: Data First, AI Second
data & AIquality assurance & testing

Application Performance: Data First, AI Second

Didier Nizard, Directeur d'Itecor Paris et Franck Jarre, QAT Practice Representative · November 19, 2000

What if the main challenge of applying AI to performance were not the algorithm, but the quality of the data fed into it?

THE KEY IDEA

Value does not depend on the volume of data collected. It stems from reliable, contextualised and correlated data capable of informing a decision. AI speeds up analysis; it does not replace this foundation.

Response times, errors, CPU, memory, Kubernetes metrics, logs, traces, deployment events: information systems already generate a vast amount of data. The challenge lies in making sense of it. A transaction takes 2.5 seconds instead of 800 ms during a load test. The slowdown is measured, but its cause remains unknown: is it CPU saturation, the database, an external dependency, or a regression introduced by the latest version? To understand this, the data must be cross-referenced. This is where NeoLoad and the Grafana ecosystem complement each other: one measures performance under controlled conditions, whilst the other provides the context needed for diagnosis.

NeoLoad: measuring under controlled conditions

With NeoLoad, we don’t just observe the system: we define the test conditions. Concurrent users, load scaling, endurance, stress testing, business workflows, data volumes and critical transactions all form part of reproducible test campaigns. Their history enables us to answer specific questions: is this version slower? At what load level does performance degradation occur? Has an optimisation actually worked?

However, you must compare like with like. Rule number one: examine the p95 and p99 percentiles, not just the average. It is often the highest response times that compromise a service objective. Rule number two: maintain the same dataset, the same platform state and the same warm-up phase, then repeat the tests to measure natural variability. Two isolated runs with a 10 per cent deviation are not enough to draw a conclusion; three stable runs followed by a comparable deviation constitute a much stronger signal of regression.

When integrated into the CI/CD pipeline as quality gates, these metrics become a delivery criterion: a release that fails to meet the service level objectives (SLOs) for a critical transaction will not proceed to production.
Representativeness remains essential. This applies not only to scenarios but also to the scale of the platform, its data and its dependencies, including simulated ones. Without a credible environment, results are difficult to translate into production.

Grafana: understanding what lies behind the figures

NeoLoad indicates when performance is deteriorating. Grafana helps to understand what is happening at the same time: metrics in Mimir, logs in Loki and traces in Tempo, all brought together in a single interface. An increase in response times, coupled with the saturation of a component or an abnormally slow dependency, becomes a useful diagnostic lead.
Endurance tests illustrate this benefit well. A memory leak or a garbage collector drift may only become apparent after several hours of stable load. It is the trend in the graphs, rather than an isolated measurement, that reveals this.

For this analysis to be useful, three conditions are essential. Firstly, instrument the test platform with the same rigour as the production environment; you can even retain all traces there when the production environment only samples them. Secondly, collect application metrics (connection pools, threads, the garbage collector, queues) beyond just CPU and memory. Finally, adopt a common language: OpenTelemetry enables you to track a transaction across services without relying on a proprietary format.

Value is created through correlation

QA, development, infrastructure, SRE and business teams each hold part of the explanation. It is not a matter of adding a dashboard, but of linking signals around a single event, from the load injected right through to the business impact.

THE ANALYSIS CHAIN

NeoLoad load → business workflow → response time → metrics → logs → traces → anomaly → probable cause → business impact. Measurement and observability then inform human analysis, potentially assisted by AI.

Four prerequisites make this chain possible: synchronising the clocks of the load generators and the platform; annotating campaigns, their stages and the versions tested in Grafana; exporting NeoLoad data series (active users, throughput, response times) to Mimir, via API or export; and finally, propagating the trace context. By injecting W3C Trace Context headers, NeoLoad enables simulated calls and their identifiers to be retrieved in Tempo.

The consequence then becomes clear: a shopping basket that cannot be confirmed, a delayed payment or processing outside the SLO. We move from a technical symptom to an identifiable business impact.

ILLUSTRATIVE EXAMPLE

NeoLoad gradually scales up to 3,000 users. At 2,200 users, the p95 for ‘basket validation’ rises from 600 ms to 1.8 s, whilst the other stages remain stable. In Grafana, this plateau corresponds to 100 per cent saturation of the database connection pool. Tempo traces show that slow transactions are waiting more than one second for a connection. After resizing the pool and rewriting a query, a new test measures 650 ms at the 95th percentile with 3,000 users. This correlation guided the investigation towards the correct cause, rather than blaming the database as a whole.

AI depends first and foremost on data quality

AI can analyse telemetry, detect anomalies, correlate signals and speed up investigations. However, it cannot reconstruct information that has never been collected. Unrepresentative scenarios, poorly defined transactions or traces lacking context simply lead to misleading analyses being produced more quickly.

An AI strategy applied to performance therefore begins with decisions about data: what data to collect, at what level of detail, for how long, and using which user experience metrics? On this basis, tools such as Grafana Assistant and Assistant Investigations, within Grafana Cloud, make it easier to query telemetry using natural language. The expert remains responsible for the context, causal interpretation and decision-making.

Our belief

The future of performance testing is not simply a matter of injecting more load or multiplying the number of dashboards. It rests on the ability to make the data speak for itself. Itecor Paris supports its clients throughout this entire process: relevant scenarios and representative environments, reliable measurements with NeoLoad, structured observability with Grafana and OpenTelemetry, signal correlation and AI-assisted use case preparation.
Three questions for your organisation

Three ‘yes’ answers: you have a solid foundation for leveraging AI. One ‘no’ answer: you have identified a concrete priority.

IN A NUTSHELL

Before asking AI why an application is slowing down, let’s give it the right data, at the right time and in the right context. AI can become the co-pilot; data remains the fuel.

Would you like to assess the maturity of your performance and observability pipeline? Our NeoLoad and Grafana experts are offering a 30-minute consultation.

More insights

Tricentis Transform London 2026: Agentic AI is redefining the boundaries of Quality Engineering


company newsdata & AIquality assurance & testing

October 30, 2026

Enjeux et bénéfices liés à la mise en œuvre d’un assistant IA conversationnel


data & AIdigital solutions

October 13, 2026

Upcoming Itecor Testing Breakfast – AI for Software Quality and Value Creation


data & AIquality assurance & testing

October 12, 2026