You want the fair dialogue familiar with the Machine Shape it, instead of just experiencing it. Many companies face concrete problems: lack of control, loss of customer trust, uncertainty regarding regulations, and a lack of practical solutions. Here you'll learn in a nutshell how you, as a... Company You create clear rules, processes and benefits – quickly implementable and economically relevant.
Seize the opportunity instead of risking falling behind. Practical tips will help you to introduce technology responsibly, bring your team and customers along, and thus build trust and competitive advantages in South Tyrol, Bolzano, or the DACH region.
AI governance as a competitive advantage: How to build fair, scalable processes from use case to rollout
Strength AI governance It gives you speed and confidence at the same time. Build a continuous, repeatable path from Use Case to Rollout with clear Stage GatesIdea gathering, prioritization, risk/value assessment, pilot, production release. Use a streamlined scorecard (business value, complexity, risk, data maturity) to make go/no-go decisions transparent. This way you prioritize fairly and scalable – for example, if a sales team only pilots initiatives that promise measurable benefits within 8 weeks according to the scorecard.
Make governance operational: Standardize artifacts and automate controls so that rules accelerate rather than hinder. Rely on templates for problem definition, assumptions, risk/impact logs, decision protocols, and a central repository. Model catalog (Registry) for versions, prompts, and training sources. Link these to your delivery pipelines (MLOps/GenAI-Ops): automated checks for documentation requirements, approvals, and traceability before the Rollout. Miss governance KPIs such as time-to-approval, component reuse rate and percentage of fully documented models to specifically address bottlenecks.
Scale without bureaucracy by making responsibilities and processes easily discoverable. Establish a lightweight governance playbook, a central use case board with visibility into status and risks, and firm SLAs for reviews. Empower teams with self-service guardrails (e.g., pre-approved building blocks, verified data access) and offer regular governance clinics where departments can clarify questions and accelerate decisions. This will AI governance from controller to enabler, creating a fair, repeatable path from idea to value creation.
Quick Wins
- One-sided AI canvas per Use Case (Goal, users, risks, metrics) – mandatory before every kick-off.
- Standard scorecard with 5 criteria for prioritization; record decisions in the decision log.
- Central registry for models/prompts with ownership, versions and release status.
- Release SLA (e.g. 5 working days) and “fast track” for low-risk pilots.
- Governance dashboard: Time-to-approval, documentation rate, reuse – visible weekly.
EU AI Act in practice: Efficiently implementing risk classes, documentation and audits
The EU AI Act It becomes practical if you can immediately assign each use case to a specific application. Risk class assign and trigger the appropriate obligations. Implement a 5-question triage: Does the use case affect access to education, jobs, loans, or critical infrastructure? Does it use biometric recognition/categorization? Does the AI interact directly with people? Is content synthetically generated? Examples: Application screening is high risk (Human Oversight, Conformity Assessment); Service Chatbot is limited risk (Transparency notice “You are talking to AI”); emotion recognition in the work context regularly verbotenRecord the classification in the decision log and link it to your approval process – this way you can provide proof of compliance at any time.
For Documentation and Proof The following applies: Collect evidence where it is generated. Build a lean "Technical File" for each system: problem definition, training/evaluation data, data origin and rights, model/prompt versions, risk analysis. Human OversightConcept, performance and bias tests, logging and security measures. Automate the population via your MLOps/GenAI Ops pipeline: automatic exports of experiments, data snapshots, test reports, red teaming results, monitoring dashboards; generated media with DeepfakeLabels and watermarks. Define retention periods and a complete audit trail leading back to the production logs – this makes audits routine instead of requiring firefighting.
Audits and conformity assessment Manage efficiently: Plan "audit readiness by design" for high-risk systems. Establish a quality-assured management system (risk management, testing standards, change and incident processes) and conduct an internal preliminary review before involving an external party. Monitor after go-live using Post-market monitoring Drift, errors, and complaints; report serious incidents and implement corrective actions in a controlled manner. Practical example: A scoring system for credit decisions uses defined acceptance criteria, human review in borderline cases, monthly bias reviews, and an escalation path – the audit evidence is continuously generated from logs, samples, and retrain documents.
Quick Wins
- Compact risk decision tree with 5 questions; classification and justification in the decision log.
- Transparency building blocks “out of the box”: Explanatory texts for chatbots, Deepfake-Labels, Watermark Policy.
- Standard templates: Model Card, Data Sheet, Risk/Impact Log, Human Oversight-Directions.
- Automatic evidence archive from the pipeline: version states, data snapshots, test and red team reports.
- Audit calendar with risk-based frequency; internal preliminary review before external audit conformity assessment.
- Incident register with clear reporting channels and playbook for corrective actions.
- Contract clauses for suppliers: documentation obligations, data origin, IP assurances, security standards.
- Optional: FRIA template (Fundamental Rights Impact Assessment) for relevant high risk-cases.
Actively reduce bias: Data quality, test sets and continuous monitoring for reliable models
bias It starts with the data – and that's where you can reduce it most effectively. Secure Data quality through clear goal definition, representative samples across all relevant areas Subgroups, duplicate and outlier checks as well as consistent Labels (Guidelines, double labeling, disagreement analysis). Address class imbalances with Re-Weighting or group-conscious sampling instead of "flying blind". Collect – legally sound – sensitive attributes for Fairness analysesWithout them, bias remains invisible. In practice: When scoring applications, you reduce distortions by contextualizing career-related gaps, checking proxy features, and explaining high-contributing characteristics for each group.
clean Test sets Make bias measurable. Create a stratified holdout that covers scenarios, languages, dialects, and edge cases; define hard biases. Acceptance criteria for accuracy, error rates and Fairness metrics such as equal opportunity, calibration gap, or false-positive rate per group. Add to that. Counterfactual Tests (e.g., swapping gender-coded terms, identical semantics) and stress tests for long texts, colloquial language, or poor audio quality. For Generative AICheck for hallucinations, toxicity, and stylistic fidelity for each language; use blocklists and safe completion prompts in the tests. Release decision: The model only goes live if performance and fairness meet the thresholds for each group.
After the go-live continuous monitoring Your models are reliable. Monitor. Data Drift (Feature distributions), Concept Drift (Performance drop) and fairness per subgroup with alerts and clear escalation paths; use Shadow Mode and Canary Releases For secure updates, set up feedback loops: user corrections, sample reviews, and bug labels flow into retraining backlogs; document changes and compare before/after metrics. In practice: A service chatbot tracks rejection and escalation rates per language, triggers data expansion for weak segments, and only rolls out an update after passing bias gates.
Quick-Wins Bias Reduction
- Data profiling Per feature: Identify missing rates, distributions, leaks, and proxies.
- Group-conscious splitTrain/Val/Test stratified; no individuals/organizations across splits.
- Fixed Fairness metrics and thresholds per use case; results matrix per subgroup.
- Counterfactual and integrate adversarial tests into the pipeline (e.g. language/dialect variants).
- Pre-Release Bias-GateDeployment only if fairness criteria are met.
- ProductionMonitoring with subgroup dashboards, drift alerts and monthly reviews.
- Rollback plan and Canary strategy for model updates; Shadow comparison before switching.
- Feedback collection point: Mark corrections, complaints, and edge cases as training candidates.
Properly organizing the human-in-the-loop: Roles, approvals, and escalation paths for secure automation
Organize yours Human-in-the-Loop like a team sport: clear RolesResponsibilities and decision-making processes. Define a lightweight RACI (e.g. Use Case Owner = Responsible, Model Owner = Accountable, Reviewer/Approver = Consulted, Legal/InfoSec = Informed) and that Four-eyes principle for high-risk decisions. Establish a central Review queue with SLAs for response times and document each release in Audit trail with justification or OverrideIn practice: A loan workflow automatically grants adjustments up to a ±10% limit; above that limit, a trained reviewer checks using a checklist and approves or rejects the application.
Tax Approvals and Escalation pathways Risk-based: three levels – Auto (green), Assisted (yellow), Manual (red). Route to Confidence Score, Thresholds and Risk grading (Impact on customer, regulation, brand). Define hard escalation triggers (e.g., sensitive content, legal claims, PII, safety risks) including On-call-Role, Kill Switch and fallback into the manual process. In practice: In content moderation, clear-cut cases are automatically published, close cases are referred to experts, and sensitive claims are immediately escalated to legal/compliance.
Hold that Quality control Human reviews are lean and effective. Use standardized tools. playbooks and decision trees, regular Calibration the reviewer (comparative assessments, edge cases) and random second reviews. Miss Metrics Such as approval rate, reversal rate, time-to-approval, escalation rate, and reasons for overrides – these inform training, coaching, and improvements to prompts and policies. In practice: A service chatbot hands over to a human live when the customer is of high value but has a negative sentiment; all takeovers are categorized to allow for finer adjustment of the thresholds later.
Quick‑Wins Human‑in‑the‑Loop
- Role card On one page: Owner, Reviewer, Approver, Escalation Manager with deputies.
- Three-stage routingGreen (auto), yellow (assisted), red (manual) based on clear thresholds.
- Checklists per decision type; maximum 8-10 points, including permissible justifications.
- Standard escalationOn-call roster, response SLA, decision rights, Kill Switch.
- Audit trail By default: every release/change includes context, screenshot/proof, and responsible party.
- Calibration round Once a month: Evaluate 20 samples together, address deviations.
- dashboards for review turnaround times, reversal rate and most frequent reasons for escalation.
- Fallback plan: If queue size increases or quality decreases, temporarily adjust thresholds by adjusting more manual intervention.
Clarifying data strategy and rights: consent, copyright, and usage rules for generative AI
Build a sustainable Data StrategyCreate a current Data inventory, classify personal data, trade secrets and copyrighted content and link each use case to a clear Legal basis according to GDPR (Consent, contract, legitimate interest). Hold consent demonstrably fixed (purpose, duration, revocation) and separate training data, Context data (RAG) and Feedback data Both technically and organizationally. Minimize data and remove PII through redaction/pseudonymization before they reach models; set Retention- and Opt-out-Flags by design. In practice: A service team uses ticket texts only as context, does not store them with the model provider, and only includes them in the training pool after active opt-in and anonymization.
Clarify the Copyrights along the entire chain – source, model, output – and document Licenses central. Check for sources whether the use is for Training and Generation What is permitted (terms and conditions, TDM exceptions, opt-out, database rights) and document the commercial use. Respect model licenses and API conditions (no covert data sharing, no prohibited advanced training); organize the Usage rights for the output to the company and regulate personality, trademark, and style rights. In practice: The marketing department creates texts/images only from licensed sources and internal knowledge bases, and marks generated assets accordingly. Origin/Watermark and archives source, prompt and license in the DAM.
Quick Wins Data Strategy & Rights
- Data card & classification: PII, secrets, IP; training release flag and TTL per data record.
- Consent Flow: Granular consents (purpose, channel), double opt-in, revocation via self-service; audit log.
- Prompt guidelines: No sensitive data in the prompt; secret scanning and DLP for uploads.
- Redaction & Pseudonymization Store reversible assignments separately before accessing the model (names, IDs, free text fields).
- License Register: Source, license type, permitted use (train, generate, publish), opt-out status, evidence.
- Model/Provider PolicyAllowed models, region/data storage, logging options, fine-tuning conditions.
- Output Policy: Labeling “AI-generated”, release criteria, C2PA/watermarks, prohibitions (images of people, logos, styles without rights).
- Scraping rulesRespect robots.txt and terms and conditions, observe TDM opt-out, rate limits, and use only legally compliant sources.
- Contracts & Company Agreements: IP transmission at output, confidentiality, employee prompts as company data.
- Incident Playbook: Rights Claim/Takedown, Kill Switch, Content Removal, Retraining/Unlearning and Notification Path.
Questions at a glance
What does "fair dialogue with the machine" mean in concrete terms within a company?
Fair means: Your AI supports people, is transparent, avoids preventable errors, and respects rights. In practical terms, this means clean data origins, clear usage rules, visible boundaries (e.g., indications of AI-generated content), documented decisions, and transparent models. Dialogue is fair when those affected know that AI is involved, receive an understandable explanation, can object, and a human takes responsibility. This is how you gain trust both internally and externally and reduce compliance, reputational, and liability risks.
How do you build AI governance as a competitive advantage?
Start with a lean but binding framework: principles (safety, fairness, transparency), roles, and processes from concept to operation. Establish an AI Steering Circle (business unit, IT, legal/data protection, security, compliance), define approval thresholds for each risk, and document decisions in a central AI register. Define standard building blocks such as templates for use case assessments, data profiles, model maps, and monitoring standards. Benefits: faster approvals, less friction, reusable building blocks – and measurably higher quality while maintaining compliance.
What roles do you need for effective AI governance?
Designate a Product Owner for each use case, a Technical Lead (Model Lead), a Business Risk Owner, and an independent AI Review function. Add an AI Steward to maintain templates, the AI registry, and metric standards. For generative AI in business units, champions serve as the first point of contact. Clear RACI mapping prevents ambiguities: who decides, who reviews, who approves, and who informs.
What does the EU AI Act require, and when does each requirement come into effect?
The EU AI Act has been in force since 2024 and is being phased in: prohibitions on impermissible practices take effect first, followed by obligations for general-purpose/foundation models and authorities; extensive requirements for high-risk systems apply later. The core principle is risk-based. Minimal-risk systems are freely usable, limited-risk systems require transparency notices (e.g., "You are speaking to an AI"), and high-risk systems are subject to strict requirements such as risk management, data and model quality, documentation, logging, post-market monitoring, conformity assessment, and CE marking. Check current deadlines with the European Commission and your association, but plan processes and artifacts now to ensure you are audit-proof later.
Which use cases are typically high-risk?
High-risk applications include those listed in Schedule III, such as recruitment and assessment, access to education, credit and insurance scoring, access to critical private or public services, use in medicine (via existing product rules), critical infrastructure, and certain police/border control applications. For example, a job application screening tool is high-risk, while an internal chatbot used for knowledge retrieval is considered low-risk. Purpose, context, and impact are crucial – document the classification and keep supporting evidence readily available.
How do you efficiently prepare technical documentation and audits?
Work with reusable templates: datasheet for training and test data (source, rights, quality, preprocessing), model map (purpose, assumptions, limits, metrics), risk file (failure modes, impacts, controls), operations log (monitoring, alerting, escalation), and user instructions. Version everything in the repository, automatically generate artifacts from MLOps pipelines (e.g., metrics, data trees, hyperparameters), and maintain an AI register for each use case. For audits: evidence chain from business problem to decision template, including bias and robustness tests, release log, and change log.
How do you measurably reduce bias across the entire product lifecycle?
Start with a clear definition of fairness for your use case (e.g., equal error rates across groups or equal access quotas). Ensure data quality through balanced sampling, controlled feature selection, and documented exclusions. Test systematically with subgroup analyses and counterexamples, define fairness metrics such as demographic parity or equalized odds, and establish thresholds. Utilize counterfactual data enrichment, bias correction during training, and human review for borderline cases. During operation, monitor drift, conduct A/B comparisons, and automatically switch to safe defaults or manual review in case of deviations.
How do you effectively organize a "person-in-the-loop" approach?
Define when a human decides, when they confirm decisions, and when they are merely informed. Set clear thresholds based on risk and uncertainty, for example: low confidence, sensitive attributes, or high impact necessitate manual approval. Provide reviewers with understandable evidence such as input, explanatory features, alternatives, and justifications. Document decisions with brief explanations so you can learn and refine your rules. Ensure backup, target times, and escalation procedures to maintain operational stability.
How do you define approvals and escalation paths in your company?
Work with quality gates: before go-live (model quality, fairness, safety), after shadow operation (real-world performance), and during regular operation (service level, drift). Define alarm rules, for example, in case of significant data shifts, increased error rates, or complaints. Establish escalations based on severity: automatic deactivation of individual functions, switching to manual operation, informing the crisis team. Maintain clear responsibilities and contact information, and document every action in the operations log.
What data strategy do you need for generative AI?
Define which data you are permitted and willing to use for prompting, RAG, and fine-tuning, including storage locations, retention periods, and deletion deadlines. Clarify the legal basis (contract, consent, legitimate interest), implement data minimization, and remove personal data where not absolutely necessary. Establish usage rules: no sensitive data in public models, red lists for prompts, permissions for knowledge sources, and labeling of generated content. Develop a rights concept for training and knowledge data, as well as governance for source quality.
Are you allowed to use internet data for training, and how do you handle copyright?
EU law includes exceptions for text and data mining with opt-out options for rights holders. Therefore, you must check rights and terms of use, respect opt-outs, and document sources. For commercial purposes, licensed datasets, your own content, or providers with contractually agreed rights are safer. For generative AI: specify source preferences, use retrieval instead of full training, respect trademark and design rights, and conduct a legal review if there is a risk to the output. Maintain a TDM policy and a record of licenses.
How do you implement the EU AI Act in practice without becoming too bureaucratic?
Build a two-stage process: a quick preliminary review (use case canvas including purpose, risk, data, and affected parties) and a more in-depth review only for higher risks. Standardize artifacts and automate data capture in CI/CD pipelines. Use pilot sandboxes with logging and clearly defined boundaries, conduct shadow operations, and collect evidence. Maintain an AI register and link it to your asset and risk management. This way, you fulfill transparency, documentation, and monitoring obligations with minimal additional effort.
What transparency and labelling obligations apply to generative AI?
Users should be able to recognize when content is AI-generated and when they are interacting with a system. For images, audio, and video, clear labeling and preferably technical tagging are advisable. In internal tools, a clear notice and a link to usage guidelines are sufficient; externally, you should embed labels in the output. Define in style guides how generated content should be reviewed and approved before publication.
How do you build audit-proof data and model documentation?
Create a unique ID for each use case, link datasets, versions, and model artifacts, and document changes with timestamps. Maintain traceable records of training and test splits, data sources, cleanups, and feature engineering. Explain model selection, parameters, evaluation metrics, and trade-offs. Describe limitations, known weaknesses, and user feedback. Secure the chain from business need to decision-making through tickets and release protocols.
How do you measure the benefits, quality, and ROI of AI projects?
Define measurable target metrics in advance, such as throughput time, conversion rate, error rate, manual effort, and customer satisfaction. Establish baseline values, plan A/B or before-and-after tests, and track costs for development, operation, and corrections. Include risk and compliance costs, such as audit expenses or liability risks, and compare them with savings and additional revenue. Communicate results regularly, and if there is no impact, adjust accordingly or discontinue the project.
How do you scale from pilot to rollout?
Let pilot projects mature in a controlled sandbox, conduct shadow operations in the real data stream, and define entry criteria for production. Standardize infrastructure, monitoring, observability, and logging so that every new use case follows the same path. Create self-service building blocks such as prompt catalogs, RAG blueprints, evaluation sets, and release checklists. This way, your portfolio grows smoothly and with calculable risk.
What monitoring practices make models reliable?
Continuously monitor input drift, output quality, latency, and error rates. Supplement user feedback and expert samples, define thresholds and automated responses. For generative AI, implement quality assessments with reference responses or human feedback and securely log prompts. Regularly review subgroup performance and bias metrics, plan retraining cycles, and maintain a post-market monitoring log of incidents and corrective actions.
How do you handle suppliers, open source, and model purchases?
Request model maps, evaluation results, training data summaries, security practices, and audit support. Secure contractual rights, obligations, update cycles, incident reporting, and export restrictions. Review open-source licenses for usage limits and liability, document changes, and evaluate security risks. Maintain a release list of approved models and an update procedure, including a fallback plan for recalls or vulnerabilities.
How do you prepare for internal and external audits?
Have an up-to-date AI registry, the associated artifacts, and approvals readily available, and simulate an audit run with random sampling. Train the teams in audit procedures, designate contact persons, and ensure clear and consistent filing systems. Show not only documentation but also operational evidence: logs, alerts, responses, and improvements. Prepare a short narrative for each use case: purpose, risk, control, and impact – fact-based and comprehensible.
Is a data protection impact assessment (DPIA) required and how does the GDPR apply?
If an operation is likely to pose a high risk to rights and freedoms, a Data Protection Impact Assessment (DPIA) is mandatory under the GDPR, for example, in the case of large-scale assessments of individuals. Combine the DPIA and the EU AI Act risk file to avoid duplication of effort. Clarify the legal basis, purpose limitation, data minimization, retention periods, and data subject rights; implement privacy by design (e.g., pseudonymization, access controls); and keep contracts with data processors up to date. Document decisions and provide a contact point for data subject access requests.
How do you train teams and strengthen acceptance?
Offer role-specific learning paths: basics for everyone, deep dives for developers, and legal and ethical aspects for decision-makers. Use real-world use cases, demonstrate limitations, error patterns, and best practices. Set up a feedback channel, maintain a prompt and example catalog, and celebrate measurable successes. Acceptance increases when benefits become tangible, risks are addressed, and no one feels like they're being controlled by black boxes.
What security risks are particularly relevant in AI and how do you mitigate them?
Address prompt injection, data leaks, model theft, poisoned training data, and hallucinations. Utilize inbound and outbound filters, RAG with source citation, strict segregation of context, secrets management, and role-based access control. Examine supply chain dependencies, track model lineage, and scan artifacts for known vulnerabilities. Test with red teaming and maintain incident playbooks, including rapid shutdown, notification, and follow-up.
How do you deal with hallucinations and errors in generative AI?
Limit tasks to knowledge domains with reliable sources, use retrieval with curated documents, and require citations. Implement quality metrics and human reviews for critical content and train users in proper prompting. Communicate uncertainties openly, choose clear output formats, and provide easy ways to report errors. Update knowledge repositories and prompts regularly to prevent known errors from recurring.
What are fast, value-creating AI use cases with low risk?
Suitable methods include knowledge retrieval using internal documents, assistance with standard texts requiring human approval, classification of simple queries, and automated logging and summaries. These deliver rapid efficiency gains, are easily measurable, and operate with streamlined governance rules. Ensure clear labeling, avoid sensitive data, and implement a simple fallback plan.
How do you define "good" explainability in practice?
Explanations should be understandable, useful, and truthful for the target audience. For business users, the most important influencing factors, comparative cases, and an actionable recommendation are often sufficient. For auditors, you additionally need data origins, model assumptions, threshold values, and tests. Avoid specious justifications, test explanations with real users, and verify whether decisions demonstrably improve as a result.
How do you combine sustainability and AI operations?
Focus on compute and energy costs, avoid over-engineering, and choose models based on "as small as possible, as large as necessary." Utilize distillation, quantization, and caching; schedule retraining based on data demand rather than a calendar. For generative AI, retrieval is often more effective than fine-tuning. Transparent metrics on cost per request foster awareness and discipline.
What three steps can you confidently implement in 90 days?
First: a streamlined AI policy set with roles, an AI register, approval checks, and usage rules for generative AI. Second: two pilot use cases in a sandbox environment with logging, quality metrics, bias testing, and human oversight. Third: a reusable toolkit consisting of a data profile, model map, risk file, and operations log, embedded in your DevOps pipeline. Afterward, you'll be faster, more auditable, and able to roll out AI at scale.
What practices help to avoid content and brand risks?
Establish style guides for AI-generated texts, templates for tone of voice, and clear no-go areas. Maintain a whitelist of permissible sources, require citations for factual content, and allow publications to be approved after a brief four-eyes review. Utilize plagiarism and trademark checks, save prompts for traceability, and train teams in research standards. This will ensure your brand remains consistent and legally compliant.
How do you ensure that AI outputs are used in compliance with the law?
Label AI-generated content, clarify usage rights, and check whether industry standards require specific disclaimers. Establish approval limits, such as requiring human review of contracts, medical, or legal information. Document approvals, secure evidence, and maintain a contact point for complaints. For external use: retain extracts of the sources and establish correction processes.
What are some typical pitfalls you should avoid?
Unclear goal definitions, missing data rights, overly large models without actual need, no measurement of real impact, legal and specialist departments involved too late, and inadequate monitoring are classic pitfalls. Avoid isolated solutions by relying on standards, reusable building blocks, and clear responsibilities. Start small, measure, learn, and scale in a controlled manner – with a focus on both benefits and security.
Concluding Remarks
In short, the three most important findings: First, dialogue with machines needs clear, human guidelines – Transparency It determines trust. Secondly, participatory processes and training are necessary for technology to truly be beneficial – co-determination Thirdly, it fosters acceptance. Technological implementation must be accompanied by clear roles, ongoing monitoring, and take on They must be secured so that benefits and risks remain balanced.
Recommendations and outlook: Start with a small, cross-functional pilot project, define measurable goals and feedback loops, and establish governance and training from the outset. Especially with digitalization, AI solutions, automation, and process optimization, an iterative approach pays off: learn quickly, scale gradually, and adapt guidelines to regulatory and ethical developments.
Get started and actively shape the future: Take concrete steps, gain experience, and get the organization on board. If you're looking for support with implementation in the DACH region, Berger+Team, as a partner for digitalization, AI, and marketing, can help you plan and implement concrete measures – pragmatically, hands-on, and with a future-oriented approach.
