Data & AI trends

In HealthTech Magazine, Stanford Health Care explains how it worked with Microsoft to build an agentic AI application for thoracic tumor boards, using the healthcare agent orchestrator. CTO Christian Lindmark and chief enterprise architect Anurang Revri stress that such applications only work once a reliable data foundation is in place, with clear governance over which data is available and who is accountable for it. Lindmark notes that AI can improve efficiency, but the real aim is productivity across the organisation. For organisations considering agents, this shows that sequencing matters: data architecture and ownership determine how much an AI agent can actually do.

Binnenlands Bestuur reports that conversations civil servants have with AI tools such as ChatGPT and Copilot can be disclosed under the Dutch Open Government Act (Woo). The Archives Act and the GDPR apply as well. The real challenge is not new legislation but applying existing obligations in daily work: civil servants need to know when an AI conversation must be retained, and their organisations must make that technically and organisationally possible. Any Copilot rollout therefore comes with concrete requirements for logging, retention periods and data access. For public bodies and other accountable organisations, governance and archiving are part of how such tools are set up rather than something to arrange afterwards.

Microsoft charges 26 euros per user per month for Microsoft 365 Copilot on an annual plan, and 18.20 euros for the Business edition, excluding VAT and on top of a qualifying Microsoft 365 subscription. According to an analysis by Crux Digits, the licence is not what decides the outcome: Copilot only surfaces what a user is already allowed to open, so folders that have been shared too broadly become visible the moment the tool is switched on. The recommended sequence is to fix permissions and sensitivity labels first, then license a pilot group, and only then measure adoption. For organisations, this means a Copilot rollout is primarily a question of data access and governance, with licences and training coming afterwards.

Microsoft reports growth in business use of Copilot, pointing to higher daily usage, large licensing customers and a rise in consumption-based billing. The figures suggest that Copilot is becoming part of routine work at a growing number of organisations, and that part of the cost now moves with actual use. That shifts the question from running a pilot to embedding the tool for good: who can access which data, how usage is tracked and what rules apply to staff. For organisations this means a Copilot rollout is not just a licensing decision but also a decision about governance, data access and adoption that needs attention beforehand.

Consultancy.nl reports on the view of advisory firm Dux Group that poor data management quietly inflates the cost of technology investments. An error in a source system can spread to other applications and eventually end up in the reports on which decisions are based. A rule of thumb from data quality management applies here: the later an error is discovered, the more it costs to fix. Dux Group therefore treats data governance as a continuous improvement process rather than a one-off IT project, so that new technology does not amplify existing data problems. For organisations, this means that ownership and processes around data quality determine whether investments in AI and analytics deliver the intended return.

A new video sets out five KPIs for measuring adoption of the paid version of Microsoft 365 Copilot, including active use and how often each user turns to the tool. The problem it addresses is familiar: organisations buy licences but often have little insight into who is actually using Copilot and how frequently. Without those figures it is impossible to tell whether the investment pays off, where training is needed and which departments are lagging behind. For organisations rolling out Copilot, this means usage measurement belongs in the deployment plan from day one, alongside governance and data access.

Microsoft has updated its Copilot adoption and onboarding guide for IT administrators on Microsoft Learn. The guide walks through the steps for enabling Copilot in an organisation and stresses preparing a secure, governed data foundation first. It points administrators to features included in Microsoft 365 subscriptions and to SharePoint Advanced Management for tidying up access to sites and documents before Copilot starts drawing on that content. It also links to the Microsoft Adoption site, which offers material for guiding employees through the change. For organisations, the practical implication is that a Copilot rollout starts with the state of their own data and permission structure, not with the licence.

A Swedish government agency has restricted the use of AI services that are not covered by a formal contract. The decision follows several recorded cases of improper AI use, in which staff entered sensitive information into public AI tools and processed the generated answers without clear rules. Consumer AI tools often forward user data and prompts to the provider’s servers, which can affect information security. By limiting AI use to contracted services, the agency aims to reduce the risk of security breaches. For organisations, the case shows that rolling out AI assistants depends less on the capabilities of the tool than on governance arrangements and controlled access to data.

TechRound explains why FinOps on its own cannot keep the costs of modern data platforms under control. Spending is no longer driven by static infrastructure uptime but by query execution, dependencies between pipelines and token consumption. Monthly spend dashboards and warehouse tagging show how much was spent, yet they cannot isolate the bad query logic or the upstream architectural choice that inflates the bill. The article therefore argues for active architectural governance across the software delivery lifecycle, linking financial telemetry to everyday design and deployment decisions. For organisations, this means that how manageable platform costs are is largely determined by the choices made in the data foundation itself.

Trade magazine c’t describes how IT consultancies help organisations move from AI ambition to working infrastructure. The GDPR applies even when a model runs locally, and the EU AI Act takes effect in stages: the AI literacy obligation has applied since 2 February 2025, and transparency rules for chatbots since 2 August 2026. AI also introduces new attack surfaces, such as prompt injection and assistants that retrieve more data than the user is allowed to see. Consultants’ rule of thumb: enforce authorisation outside the language model. For organisations rolling out an AI assistant, access rights and data governance must therefore be in order before the model uses them.

Diginomica reports that many organisations migrating their virtual machines to a new platform still approach the move as a question of compute. The article argues it is really a data architecture decision: how data is stored, protected and moved determines whether the migration removes silos and fragmentation or simply recreates them. A shared data platform can reduce the number of tools in use while still delivering the performance and resilience that critical workloads require. The new environment also has to leave room for AI initiatives rather than only existing applications. For organisations, this means an infrastructure migration is also a moment to reassess the data foundation underneath it.

HackerNoon has published an analysis of how Fortune 500 companies build their data platforms. The common thread is a layered design, often called the medallion architecture, in which raw, cleaned and validated data each get their own layer, owner and audience. The author notes that the diagrams show the version that went well and leave failed migrations out of sight. The value lies in four questions: where raw data lands, who owns each dataset, what stops bad data from reaching a dashboard, and how a number would be explained to an auditor. For organisations building reporting and AI on their data, these choices determine whether figures can still be traced later.

Research by Everpure shows that almost six in ten large companies have no data sovereignty strategy, while 88 percent of senior executives fear that failures in this area could cost them their jobs. Everpure recommends becoming ‘sovereign by design’, with controls and governance tailored to the risk of each dataset, application and workload. The focus shifts from national sovereignty to the risk appetite of the organisation itself, with attention to data quality and data management policy. For organisations, this means that decisions about where data is stored and who can access it become part of the data architecture rather than a question raised only once an AI project starts.

Microsoft is changing how Copilot Business is billed, ICT Magazine reports. Alongside the fixed per-user licence, a usage-based charge is being introduced, so costs will depend on how often and how heavily employees actually use Copilot rather than solely on the number of licences assigned. According to the report, organisations that have not set budgets and consumption limits in advance could see their monthly bill quietly climb to thousands of euros. For organisations, this means rolling out Copilot involves more than licences, data access and adoption: it also requires cost control and clear agreements on who may use which features.

Berenschot’s AI Trend Survey 2026 finds that many organisations remain stuck in small-scale pilots and struggle to embed AI in a structural way. Digital security tops the list of obstacles, ranging from employee privacy to the handling of confidential client information, followed by AI literacy, the accuracy of output, data quality and staff skills. Notably, 37 percent of organisations offer no AI training at all, even though the EU AI Act has required adequate AI literacy among staff since February 2025. For organisations, this shows that rolling out generative AI tools such as Copilot depends on preconditions settled in advance: rules on data access, checks on output and training for employees.

SiliconANGLE argues that most enterprise AI budgets rest on one assumption: data will move to wherever the new platform lives. At enterprise scale that rarely holds. Data has gravity: regulations dictate where certain records may be stored, sovereignty rules keep datasets inside a jurisdiction, and moving volumes costs time and money. Data architecture therefore becomes a hidden tax that surfaces after the compute deal has closed. The article calls for treating architecture choices and vendor selection as a first-class line item in the AI business case. For organisations, this means the return on AI depends on decisions about the data foundation that come before the choice of models or hardware.

A benchmark study by Virtuous finds that 92 percent of nonprofits already use AI in some form, yet only 7 percent report an impact at the level of strategy. According to The NonProfit Times, the reason lies in the data: AI amplifies whatever it is fed, including duplicate records, inconsistent formatting and outdated information about constituents. A 2026 survey by Drexel University adds that organizations with a functioning data governance program report high trust in their data at a rate more than 20 points higher. For organizations, this means the return on AI depends largely on ownership and standing processes around data quality rather than on the tools themselves.

Research by ITDS, reported by Consultancy.nl, shows that financial institutions are investing heavily in AI while their data management lags behind. Many organisations have appointed data owners and data stewards, but these roles are not yet filled effectively in daily operations. As a result, data quality, accessibility and usability fall short of what AI applications require. The researchers recommend a pragmatic approach in which concrete AI use cases and improvements in data quality reinforce each other, rather than waiting until the entire data landscape has been optimised. For organisations, this means that ownership of data and the processes around it determine whether AI investments actually deliver results.

Berenschot’s AI Trend Survey 2026 shows that many organisations have yet to embed generative AI in their day-to-day work. As long as the basic conditions are not in place, most stick to small-scale pilots. Digital security tops the list of obstacles, ranging from employee privacy to the handling of confidential customer information, followed by AI literacy, output accuracy and data quality. Notably, 37% of organisations offer no AI training at all, even though the EU AI Act has required adequate AI literacy among staff since February 2025. For organisations, this means that rolling out AI tools is first a matter of governance, data access and training, and only then of technology.

A survey of Dutch employees shows that 68 percent have yet to see any benefit from AI in their own work. Among those who used AI in the past year, 22 percent use generative AI daily, for tasks such as summarising documents or drafting emails. Daily use of AI agents within that group stands at 7 percent, mostly in specific processes such as recruitment. The gap between occasional use and an established routine is wide: opening a tool now and then does not change how the work itself is done. For organisations, this means that rolling out tools like Copilot only pays off when usage becomes embedded in everyday work processes.

Market research agency IntoTheMinds has published a study on how companies use generative AI. Its main finding: shadow AI is already a reality. Employees use generative AI tools every day, while organisations exercise little or no oversight of that use. The study also identifies a generational gap in adoption and points to consequences for HR. The researchers argue that a policy framework is needed, but warn that it must be weighed carefully: overly strict rules affect an employer’s attractiveness and, with it, recruitment. For organisations this means generative AI is in use before governance, data access and responsible use have been defined, and that a structured rollout of a business alternative exposes that gap.

Precisely has announced a Unified Data Management Platform intended to make enterprise data ready for AI. It covers six areas: data integration, data quality, data governance, location intelligence, master data management and customer communication. According to the company, data remains spread across cloud environments, lakehouses, on-premises systems, mainframes and SAP, so the platform manages data where it lives rather than requiring migration into a single ecosystem. Within pipelines it profiles, validates, standardises, matches and reconciles data, enforces governance policies and tracks lineage. For organisations, this shows that governing data in place or centralising it is an architectural choice that shapes how usable the data becomes for AI.

Research firm Barc surveyed 234 leaders in data, AI, IT and business for its report The State of Data and AI Governance. Two in three organisations have already experienced an AI incident. The data layer is reasonably well covered: 58 percent manage data quality and 62 percent apply governance controls. Models and agents are a different story. Only 26 percent use monitoring and observability for models, 23 percent keep a model register or inventory, 20 percent run a dedicated AI governance platform and 10 percent have a solution for agent governance. For organisations this means that managing AI systems in production, including monitoring and maintaining a current overview of models, lags behind the way data itself is managed.

In a KPN Zakelijk blog, Microsoft’s Michel Bouman describes AI adoption in the workplace as a three-stage process. The first stage is Copilot as an assistant: meeting notes, Excel formulas, presentations and searching across files, so staff spend less time on routine work. The later stages build on that foundation. One condition runs through all of them: the more business data Copilot can reach, from Microsoft 365 and connected SaaS services, the more useful its output becomes. That shifts attention to data access and permissions before rollout. For organisations, the value of Copilot depends largely on the condition of their own data and on how access to it is arranged.

In a new analysis, IBM identifies the layer that is missing from the data architecture of many organisations that want to deploy AI agents: a real-time operational layer. Much of their data still moves through batch pipelines into downstream systems, so agents end up acting on information that is already out of date. A streaming architecture captures events such as transactions and inventory updates as they happen, enriches them with context and applies governance before applications consume them. In a survey of more than 4,500 IT leaders, 88 percent said data streaming can unblock progress on agentic AI. For organisations, the choice between batch and streaming determines how current and reliable the data is that AI decisions rest on.

In many organisations Microsoft Copilot is still used mainly as a smart text generator, writes adoption specialist Philippe Dingemans in an opinion piece on ITdaily. That saves time, but the strategic value stays out of reach as long as Copilot has no access to the organisation’s knowledge, processes and expertise. According to Dingemans, the organisations getting the most out of Copilot invest heavily in knowledge sharing, practical examples and internal champions. The real question is not whether an employee can write a prompt, but whether they understand how AI fits into their own work. For organisations, this means a Copilot rollout involves more than licences: it also requires decisions on data access, governance and guidance.

Research among organisations in the Benelux shows that nearly a quarter of recommendations produced by AI systems are either adjusted or ignored. Organisations are becoming more selective about deploying autonomous AI and are first trying to strengthen oversight and accountability. Data quality and governance are now cited by 63 percent as a critical factor for reliable AI, up from 47 percent a year earlier. Scores for data quality, oversight and explainability improved, yet the region still trails global benchmarks, and the measured business impact barely changed. For organisations, this indicates that governance and data quality only pay off once data ownership and the use of AI output are embedded in everyday working processes.

Microsoft has expanded its Copilot offering in an effort to close the gap between the reach of Office and actual use of its AI assistant. Invezz reports that Microsoft has more than 450 million commercial Office 365 licences worldwide, yet fewer than 7 percent of them include the Copilot add-on. Analysts identify stalled adoption as the main risk, pointing to delays caused by security, compliance and governance requirements inside enterprises, as well as weak user demand. For organisations, this shows that rolling out Copilot depends less on the licence and more on how data access, governance and day-to-day use by employees are organised.

On 25 September Microsoft announced a set of new Copilot features. Word, Excel and PowerPoint are being built directly into Copilot, so users can work on documents without switching applications. Later this month Microsoft will also start limited testing of Autopilot, an agent with its own identity that carries out tasks independently, alongside tools that let employees track the cost of their AI usage. Copilot is thereby moving from a chat assistant to a work platform in which agents gain access to documents, data and applications. For organisations, this means questions about data access, permissions and governance need answers before rollout, since an agent with its own identity can only operate responsibly within clearly defined boundaries.

Enforcement of Article 4 of the EU AI Act has applied since August 2026. The article requires providers and deployers to take measures supporting the AI literacy of staff working with AI systems; the obligation has existed since February 2025, but the amended article no longer sets a specific individual level. Writing on CFO.nl, Jennifer Kuiper argues that uniform training adds little: AI literacy only pays off when employees know in practice when AI is useful, what needs checking, where the limits lie and when to escalate. For organisations, AI adoption is largely an organisational task: roles, processes and ownership decide whether the duty takes hold in daily work.