63% of Corporate R&D Teams Admit Key Decisions Rested on Flawed Data

Sixty-three percent of corporate R&D professionals say their organization has at some point made an important R&D decision using data or evidence later found to be inaccurate, incomplete or outdated. That figure comes from 1,038 respondents at R&D-led corporations across 55 countries, surveyed by Elsevier and published Nov. 12 in its Corporate R&D of the Future report.

The number sits awkwardly beside another finding from the same survey: 76% of R&D leaders, managers and researchers say they are confident in the quality of their data. Confidence, in other words, is running well ahead of the track record. A further 51% of respondents say their organization has had to make an important R&D decision before all the evidence was in.

R&D managers reading these figures should note the method before treating them as gospel. The online survey ran by email invitation from August to October 2026. The figures are weighted, and the flawed-decision numbers exclude respondents who answered "don't know" or preferred not to say. There is also a commercial dimension: Elsevier sells evidence and AI tools to corporate R&D teams, and its LeapSpace research workspace, which went live in January and added claim-checking tools in June, stands to benefit from corporate anxiety about evidence quality.

What blocks access to good evidence?

Access is not the primary problem, according to the respondents. Seventy-eight percent say their teams can reach the third-party evidence they need. Yet 56% of that group still report at least one barrier. The obstacles are structural rather than technical:

• Scattered sources — 42% of respondents
• Slow finding and synthesis — 32%
• Uncertainty about what evidence exists — 30%

A European R&D leader quoted in the report put the mechanism plainly: "Fragmented data sources and disconnected systems create silos that slow down analysis and reduce transparency and weaken confidence in outcomes." For portfolio managers, that translates directly into slower cycle times and decisions taken on partial information.

How does AI change the risk?

AI adds a layer to the evidence-quality problem precisely because R&D teams are adopting tools that can generate inaccurate answers or unsupported claims. Nearly half of respondents (47%) say their organization's R&D function uses AI regularly, and another 40% say it is piloting the technology. Adoption is therefore close to universal across the surveyed population — regular use plus pilots covers 87% of respondents.

The tooling choices matter for anyone budgeting for evidence infrastructure. For literature searches, 48% of those answering use general-purpose tools such as ChatGPT or Copilot, while only 23% use specialist scientific AI tools. General-purpose chatbots outnumber domain-specific tools roughly two to one in one of the most evidence-sensitive tasks in the R&D workflow.

What do R&D teams want to fix first?

The survey asked respondents what their organizations should focus on over the next five years, and the answers sketch a priority stack:

• Building AI skills across the organization — 65%, the most-selected priority
• Turning data and evidence into actionable insights — 62%
• Tailoring AI to organizational knowledge and workflows — 59%

The ordering is telling. Skills outrank both tooling and data integration, which suggests respondents see capability gaps inside their teams rather than gaps in the vendor market as the binding constraint.

Is there a working model for reliable AI in R&D?

A Pistoia Alliance project offers one concrete approach. The Pistoia Alliance is a nonprofit that promotes collaboration in life-sciences R&D. Christian Baber, its chief portfolio officer, told R&D World in an Oct. 29 interview that the project covers large language models that build instrument data converters — software that translates instrument output between formats.

The models themselves are "not consistent enough" to perform that conversion directly, Baber said. So teams have the model write a Python script instead. "Once that Python script is written, it's testable and validatable," Baber said. "You use something unpredictable to generate something predictable." He added that the approach "is much more supported by regulators as well."

That construction — an unpredictable generator feeding a verifiable, auditable artifact — is the practical answer to the survey's central tension. It treats AI output as draft code to be validated, not as evidence to be trusted, and it aligns with regulatory expectations in a way direct model use does not.

The survey's own numbers make the case for the discipline: when 63% of organizations have already made consequential decisions on flawed evidence, adding a generative layer without validation steps would compound an existing failure mode rather than cure it. Expect validation-first AI workflows, of the kind the Pistoia Alliance is testing, to move from pilot projects into standard R&D practice over the next several budget cycles.