Methodology

Methodology

Through a well-established annotation methodology, continuous upskilling of our teams, and enhanced cybersecurity, we protect your data while maximizing its value for training your AI models.

EXpertise

6 Steps to Power Your AI Model Training

At Infoscribe.ai, we apply a strict process to turn your raw data into high-quality, consistent, and immediately usable AI training datasets.

Human in the Loop

We place humans at the heart of our approach: our annotators validate, correct, or reject your AI’s predictions to enhance its accuracy, robustness, and operational reliability.

We design custom workflows and APIs, integrated with your tools, to provide real-time access to controlled, documented, and traceable data streams—perfectly tailored to the requirements of your use cases and your product or data teams.

This oversight is performed within very short timeframes; depending on the tasks, we can commit to reviewing a batch in as little as 5 minutes.

Our Quality Approach : 3 Levels of Engagement

We tailor our quality framework to your priorities, with three clear levels of commitment to balance accuracy, budget, and project timelines.

Quality target

95%

1. launch

2. Production

3. Continuous Quality Control:

Quality target

97%

1. launch

2. Production

3. Continuous Quality Control:

Quality target

99%

1. launch

2. Production

3. Continuous Quality Control:

Annotator Training and Upskilling

We support our annotators with regular, targeted training to maintain a high level of quality on every project.

Initial Onboarding and Training for Image Annotators

From the very start of the project, our operators are trained in sessions led by our project managers, based on the guidelines and examples you provide.

If needed, your experts can also participate directly to train our project managers via video conference, ensuring full alignment on business requirements.

At the end of these workshops, we conduct a live annotation session: annotators work on real cases, ask questions in real time, and immediately correct errors with the support of the project manager.

Continuous Training

Once production is underway, we monitor the quality of each annotator daily using a skills matrix that allows us to quickly identify operators with weaknesses in accuracy or speed.

Identified profiles are supported by more experienced annotators through a mentorship and collaborative learning approach. Project managers remain available to answer questions.

Any client feedback or guideline updates trigger a new training session, followed by a group annotation session to ensure that the updated rules are fully understood and correctly applied by all.

FAQ

Frequently Asked Questions

Every project starts with a pilot using a representative sample of your data. During this stage, we review the dataset and annotation guidelines, clarify your quality objectives, identify ambiguous cases, consult your teams to refine the rules where needed, and agree on communication channels and day-to-day project operations.

The pilot allows us to design and test the annotation and quality-control workflow on real examples, measure throughput, identify common sources of error and validate the guidelines with you before scaling up.

Once the pilot is complete, we have the information needed to provide a firm price, a production schedule and a team sized to your volumes. The production team is then trained on the guidelines and lessons learned during the pilot before work begins.

We offer three contractual quality levels: 95%, 97% and 99%. The appropriate target is agreed with you based on project requirements, annotation complexity and the level of review needed.

Production is checked at several levels. Annotations first undergo quality control within the production workflow. This can be supplemented by independent review by a second team and an additional check by the project manager. Sampling rates across these checks increase according to the selected contractual quality level.

Quality is measured regularly using reviewed samples and tracked throughout production. Identified errors are documented and analyzed, then used to refine team guidelines and checklists.

If measured quality falls below the agreed 95%, 97% or 99% threshold, the affected annotations are corrected or redone at no additional charge until the required standard is achieved.

Our aim is to meet the agreed quality standard on the first delivery, sparing clients time-consuming rounds of feedback and rework.

If a quality check finds results below the contractual threshold of 95%, 97% or 99%, we do more than fix individual annotations. We investigate the root causes, determine why errors passed through the review process and implement corrective actions to prevent them from recurring in subsequent batches.

Affected data is then corrected or reannotated and reviewed again before delivery. All such rework is performed at no additional cost. Our commitment covers both achieving the contractual quality target and addressing the underlying causes of errors.

Yes. Infoscribe adapts to your technical environment and the sensitivity of your data. Our teams can work directly in your existing annotation platforms, including CVAT, Label Studio, Supervisely, SuperAnnotate, V7 and proprietary tools, without requiring you to migrate to a particular solution.

When data must remain within the European Union, we can host the working environment on our infrastructure in France and arrange team access in line with your security requirements.

For highly sensitive datasets that cannot leave your own servers, we can help install and configure annotation tools directly on your infrastructure. Our teams then connect remotely to that secure environment to annotate and review the data, without transferring it to our servers.

Architecture, tools, access rights, roles, and production and quality-control workflows are agreed during the pilot to meet your security, confidentiality and data residency requirements.

Infoscribe is ISO 27001:2022 certified. Data security is embedded throughout the project, from receipt through delivery and, where applicable, deletion at the end of the engagement.

Access to datasets and tools is granted according to project needs and restricted to authorized personnel. Data transfers are secure and encrypted, and file integrity can be verified using checksums.

Where confidentiality or data residency requirements apply, we adapt the project architecture: processing on our infrastructure in France when data must remain in the EU; HDS-certified hosting for health data when required; or working directly within the client’s infrastructure when datasets cannot leave its servers.

For certain sensitive datasets, particularly medical data, a de-identification stage can also be completed before annotation teams are given access.

Traceability requirements are defined during the pilot and maintained throughout production. Depending on the project, we can retain version histories of annotation guidelines, processed batches, quality checks, measured compliance rates, identified nonconformities, and corrections or reannotations performed.

Significant changes to annotation rules are documented so it remains clear which guidelines applied at each stage. Regular reporting also provides visibility into output volumes, quality results and progress against the delivery schedule.

Where required, these records can be consolidated into a data governance file at project closeout. This gives clients structured documentation of work performed on their dataset and can support the data governance practices expected of systems covered by Article 10 of the EU AI Act.

For relevant high-risk AI systems, Article 10 of the EU AI Act sets out requirements for the governance and management of training, validation and testing data. These cover, among other things, data origins and collection processes; preparation activities such as annotation, labeling, cleaning and enrichment; dataset suitability and representativeness; identifying bias; and addressing potential gaps.

Infoscribe can incorporate these considerations from the pilot stage and maintain traceability of operations throughout production. Depending on the project scope and information available to us, we can document data provenance, versions of annotation guidelines, rules applied, data-processing activities, quality checks, measured compliance rates, nonconformities, corrections and reannotations.

These records can be consolidated into a data governance file delivered alongside the dataset. This provides structured, usable documentation of the work performed, rather than requiring your team to reconstruct the history later.

This documentation can support your preparation for the EU AI Act, but does not in itself certify that your AI system is compliant. Compliance depends on the system as a whole and the obligations applicable to its provider.

The timeline depends on the type of system. For high-risk AI systems listed in Annex III, including certain uses in employment, education, biometrics, critical infrastructure, access to essential services, justice and migration, the relevant obligations apply from 2 December 2027.

For high-risk AI systems embedded in regulated products covered by Annex I, including certain medical devices, machinery and other products subject to specific EU legislation, the deadline is 2 August 2028.

For data governance, it is worth preparing well ahead of these dates. Data provenance, guideline versions, annotation activities, quality checks and correction histories are far easier to document as they are produced than months or years later.