Most post-mortems on a stalled AI rollout point at the model, the vendor, or the budget. That’s rarely where the fault actually sits. Gartner’s own data-readiness research puts a number on this: organisations pursuing AI initiatives unsupported by AI-ready data are expected to abandon the majority of them, and separate IDC research finds that 94% of IT leaders now cite data quality, not model choice, as the primary factor determining whether an AI project succeeds (Gartner, Assess Data Readiness for AI) The tooling has never been the constraint. The data underneath it, and the people who have to work with that data, are.
The pattern shows up before the AI layer does
We run hands-on Excel, Power BI and SQL training for corporate teams, often brought in specifically because a company has just adopted, or is mid-rollout on, some flavour of AI copilot or analytics assistant. The brief is usually framed as “help our people use the new tool properly.” What we actually find, session after session, is that the tool is not the bottleneck. The spreadsheet underneath it is.
A team using Copilot in Excel to summarise a report will get confidently wrong answers if the underlying data has three different date formats, a merged-cell header row, or a “Region” column where “UK”, “United Kingdom” and “Uk ” are three separate values to anything doing string matching. The AI layer does not know that. It will happily aggregate all three as distinct categories, and the person reading the summary will not necessarily catch it, because the output looks clean. The failure is not visible until someone downstream acts on the wrong number.
Two literacy gaps, not one
Informatica’s 2026 CDO research describes this as a “trust paradox”: 65% of employees believe the data feeding their AI tools is solid, while 75% of the data leaders who actually own that data say those same employees need serious upskilling in data literacy (Informatica, CDO Insights 2026) That gap between confidence and competence is exactly what we see in the room. It splits into two separate problems that get treated as one.
The first is data literacy in the traditional sense: knowing what “clean” data actually looks like, why a pivot table breaks on inconsistent formatting, why VLOOKUP returns the wrong row when a source column is not sorted the way the function assumes. The second is a newer gap specific to AI tools: knowing how to prompt an assistant to interrogate its own output rather than accept it, and knowing which questions a spreadsheet-shaped AI answer cannot actually answer reliably given what is underneath it. Training programmes that only address the second gap, because it is the new, exciting part, leave the first one fully intact. The AI just makes the first gap’s consequences arrive faster and look more authoritative.
Where the ROI actually leaks
This is consistent with what the wider research is now converging on. Precisely’s 2026 State of Data Integrity study found a striking confidence gap: the large majority of leaders report having adequate data readiness for AI, yet a similarly large proportion separately admit data readiness is one of their biggest obstacles, in the same survey. Info-Tech Research Group’s Data Priorities 2026 report reaches a similar conclusion from a different angle: nearly 41% of IT leaders now rank improving data governance as a top priority for the year, ahead of AI-specific initiatives, because governance is the precondition, not a parallel workstream.
None of this is an argument against the AI investment. It’s an argument for sequencing it correctly. The organisations getting a return are not the ones with the newest model; they are the ones who treated “can our people actually work with this data” as a question to answer before rollout, not a support ticket to handle after go-live. In practical terms that means auditing what a “clean” version of the team’s core spreadsheets and reports actually looks like before the AI layer goes on top of them, and training people to check an AI-generated summary against the data the way they would check a junior analyst’s first draft, rather than treating it as ground truth because it arrived instantly.
What actually closes the gap
The fix we see work is not a one-off AI-tools workshop. It is going back one layer: making sure the team can genuinely handle the spreadsheet, the query, or the dataset the AI tool sits on top of, so that when the AI gets something subtly wrong, someone in the room notices. That’s a less exciting brief than “AI training,” and it’s the one that actually moves the ROI needle, because it fixes the thing the model cannot see for itself.
