The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Businesses were eager to use generative AI, but experimentation was moving faster than production readiness. A Forrester Consulting study commissioned by Dataiku found that inadequate data infrastructure, difficult integration, governance, privacy, skills, computing capacity and trust issues were holding companies back. The findings describe conditions measured in November 2023—not a 2026 ranking—but they explain why many enterprise AI pilots struggled to become dependable business systems.
What the Forrester study actually measured
The headline comes from a VentureBeat article published January 3, 2024, summarizing research conducted by Forrester Consulting on behalf of Dataiku. Dataiku says the survey questioned 220 AI decision-makers at large companies in North America in November 2023 (Dataiku’s study description).
That sponsorship matters. The respondents’ reported obstacles and Forrester’s analysis should be distinguished from Dataiku’s recommendations for using its own platform. The sample also represents large North American companies and AI decision-makers; it should not be treated as a survey of every business, industry or geography.
The adoption paradox: broad interest, limited readiness
Some 83% of respondents were exploring or experimenting with generative AI, and more than half had identified multiple potential use cases. The VentureBeat summary also reported that more than 60% considered generative AI critically or highly important to business strategy. Customer experience, product development, self-service analytics and knowledge management were among the areas being considered.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
#1 Best Overall
Exploration is not deployment. A prototype can answer questions from an uploaded file, while a production system must use current, permissioned data, fit an existing workflow, withstand testing and produce an auditable result. The gap between those two states is the central finding.
How the barriers were reported
The percentages below come from the source summaries. They describe reported barriers and concerns, not a statistically validated league table: the available material does not provide question wording, margins of error or evidence that a 35% result is meaningfully greater than a 31% result.
| Barrier or concern | Share reported | What it means operationally |
|---|---|---|
| Data infrastructure for consuming, storing and sharing required data | 35% | Core data is not readily available, governed or usable by AI applications. |
| Integration with existing infrastructure | 35% | Models and prototypes do not connect reliably to business systems and controls. |
| Governance challenges | 35% | Organizations lack workable rules, ownership and operating controls. |
| Data protection and privacy | 31% | Data flows, retention, access and provider terms create compliance risk. |
| Lack of skills and governance capabilities | 31% | Teams lack the combined technical, domain, legal and operational expertise required. |
| Computational limitations | 27% | Capacity, latency or cost can constrain model use and scale. |
| Interpretability and explainability | 25% | Users cannot always understand, justify or audit an output. |
More than half of respondents were concerned that bias and hallucinations could damage output quality.
Why infrastructure—not just GPUs—is a bottleneck
The 35% data-infrastructure result covers much more than buying accelerators. An enterprise AI application may depend on warehouses and lakes, document repositories, APIs, identity systems, retrieval indexes, model hosting, monitoring, evaluation pipelines, network capacity and legacy applications. Computational limitations, reported by 27%, are only one part of that stack.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Rank #2
Integration was separately reported at 35%. A production assistant may need controlled access to CRM, ERP, HR, finance or customer-service records; a governed knowledge base; authentication and authorization; human approvals; audit logs; and incident-response processes. Manual copying between systems can expose confidential information, use stale material and make it impossible to establish which source version influenced an answer.
Data quality determines whether answers are useful
Data availability is not data usability, and model quality is not data quality. A capable model can still produce a plausible but wrong answer when its sources are incomplete, outdated, contradictory or inaccessible. Forrester later described data quality as a primary limiting factor for business-to-business generative-AI adoption and linked it with privacy and security concerns (Forrester’s analysis).
“Clean everything first” is rarely practical. A better approach is to identify the authoritative sources for a specific, valuable use case, assign owners, document permissions and test whether the information is current and complete enough for that workflow. Data should be fit for purpose, not perfect across the entire enterprise, as Dataiku’s later discussion notes (Dataiku’s implementation discussion).
Governance is an operating system for AI
A policy that merely says “use AI responsibly” does not resolve the 35% governance barrier. Practical governance must define:
Recommended Free Tools
- Approved, restricted and prohibited use cases.
- Which models may access which data, with role-based and least-privilege controls.
- Whether personal, confidential or customer data can be sent to an external provider.
- Retention, deletion, logging and audit requirements.
- Prompt, model, retrieval and application versioning.
- Testing before release and monitoring after release.
- Human review, escalation and accountability for consequential outputs.
- Vendor, copyright, confidentiality and regulatory risk controls.
Governance must continue after launch. Model changes, new data sources, altered prompts and changed workflows can change risk even when the user interface looks the same.
Privacy and data protection create concrete failure points
Thirty-one percent identified data-protection and privacy concerns as a barrier. Risks include sensitive information entering prompts, unauthorized retrieval of confidential documents, provider retention or reuse, personal information appearing in generated text, weak user-level permissions, untraceable deletion, cross-border transfers, unknown subprocessors and exposure through logs or debugging tools.
Generative AI does not automatically violate privacy law. It does, however, add data flows that must be mapped, limited and monitored. Organizations need to know what enters a system, who can retrieve it, how long it remains available, where it is processed and how an individual’s information can be corrected or deleted when required.
Trust: hallucinations, bias and explainability
Hallucinations
A hallucination is an answer that sounds credible but is unsupported or false. Retrieval from approved sources, citations, constrained workflows, automated checks and human review can reduce the risk, but no generic model should be assumed to be inherently factual.
Bias
Bias is systematic skew in generated or ranked outputs. Testing should use representative cases and examine outcomes for affected groups, not just average accuracy.
Explainability and provenance
Explainability concerns why a system produced an output and whether it can be justified. Provenance concerns what sources, model, prompt, retrieval context and workflow steps contributed to it. A low-risk brainstorming tool may need less explanation than a system supporting credit, employment, healthcare, legal, safety or financial decisions.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Skills and computing capacity are necessary, but insufficient
The reported skills obstacle is broader than a shortage of prompt writers. Production teams may need data engineering, model evaluation, security, privacy and legal review, domain expertise, product management, change management, governance and ongoing operations.
Likewise, more compute does not repair poor source data, weak access controls or an undefined workflow. Teams must balance model quality against latency, cost, scalability and the consequences of failure. A smaller or conventional system may be better than a large model when the task is deterministic and well specified.
Free tools Windows power users keep installed
One-click scans. No signup required.
A practical path from pilot to production
- Choose a measurable workflow. Define the baseline cost, time, quality or risk outcome, and compare generative AI with search, rules, analytics or ordinary automation.
- Classify the consequences of error. Decide whether the system is assistive or autonomous, whether it handles regulated data and when human approval is mandatory.
- Map data and permissions. Identify authoritative sources, owners, freshness requirements, access rights and prohibited data flows.
- Build representative evaluations. Test accuracy, groundedness, safety, bias, latency and cost, including unusual and adversarial inputs.
- Integrate with least privilege. Connect the systems the workflow actually needs, start with read access where possible and make actions reversible.
- Add review and traceability. Capture citations or source references, version the model and prompts, and record approvals and exceptions.
- Operate and monitor. Assign an owner, track quality and incidents, review drift and provider changes, and define a retirement process.
- Scale only on evidence. Expand after the application meets its business, risk, reliability and cost thresholds in real use.
What the 2023 findings mean in 2026
The original evidence reflects late-2023 conditions and was reported in January 2024. It should not be presented as a current 2026 ranking. Forrester’s broader 2024 generative-AI report (report page) and subsequent material continue to emphasize data quality, privacy, security and trust, but those later observations are separate from the 220-person survey.
The durable lesson is that access to a foundation model is not the same as enterprise readiness. Production success depends on trusted, fit-for-purpose data; secure integration; accountable governance; measurable workflows; evaluation; and people who can operate the system.
How to evaluate a platform—or decide not to buy one
A collaborative AI platform can combine development, integration and governance, but it is not the only route. Depending on existing architecture, an organization might use cloud model services, its current data platform, independently assembled retrieval and evaluation components, or software already used by the relevant business team. It may also decide that deterministic automation is safer.
Evaluate any option against data residency, provider choice, identity and policy controls, retrieval and evaluation, monitoring, auditability, integration, deployment model, portability, consumption pricing and implementation support. The common buying mistake is selecting a broad platform before defining the workflow, data owner, success metric and risk controls.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesQuick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
