Written by: Kacper Osiewalski, Lead Backend Engineer, Digital Colliers

A salesperson at a distributor spends ten minutes finding one part number. The data isn't missing. It's sitting in a PDF catalogue that's been on the shared drive for six years. The bottleneck isn't storage. It's the question format.

This is the pattern behind most of the AI work that actually ships. Same catalogue, same rows, same fields. What changes is that the person asking no longer has to translate their question into the shape the system expects. They ask in plain language, the system does the translation, and the answer comes back in three seconds instead of ten minutes.

The data was never the problem

Walk into any distributor, manufacturer, or logistics operator and you'll find the same thing. Product data in PDFs. Spec sheets in scanned images. Pricing rules in a spreadsheet someone maintains by hand. Historical orders in an ERP that requires four clicks and a transaction code to query.

The data exists. It's often clean enough. The reason nobody uses it at speed is that the interface to it is human patience.

When a sales rep needs a part number, they don't run SQL. They open a PDF, scroll, cross-reference a second document, and confirm with a colleague. Every step is cheap on its own. Stack twelve of them and you get ten minutes per query, forty queries per day, and a team that quietly accepts this as the cost of doing business.

That cost compounds. It shows up as slower quote turnaround, lost deals to faster competitors, and a training cliff every time someone new joins.

Why most AI pilots don't fix this

The failure numbers are brutal and consistent. 88% of AI proof-of-concepts never reach widescale deployment. More than 80% of AI projects fail, roughly twice the rate of conventional IT work. 95% of enterprise GenAI pilots deliver zero measurable P&L impact. In 2025, 42% of companies abandoned most of their AI initiatives, up from 17% the year before.

The common thread in the projects that die: they tried to change the data instead of the interface.

You see it in the shape of the pitch decks. Migrate the catalogue to a new system. Restructure the product taxonomy. Build a knowledge graph. All valid work, all expensive, all optional if the actual bottleneck is that a human can't ask the existing data a natural question.

The teams that ship in 2026 tend to invert this. They leave the PDF where it is. They leave the ERP where it is. They put a retrieval layer in front that speaks the language the salesperson already uses, and they measure one thing: time from question to answer.

The three-second version

A working version of this looks boring from the outside. The rep types "what's the equivalent to part X in the Y series, in stock, under Z price." The system parses the question, hits the same PDFs and tables that were already there, and returns a shortlist with source links.

The engineering underneath isn't glamorous. It's document parsing that actually handles tables. It's a retrieval layer that knows the difference between a spec value and a marketing claim. It's an evaluation harness so you know when the answers drift. It's access control so the system doesn't leak pricing to the wrong tenant.

The reason this ships when other pilots don't is that the scope is contained. One question type, one user, one measurable metric. No taxonomy rewrite. No data migration. No twelve-month roadmap.

The cost of leaving it alone

The temptation is to wait. Wait for the catalogue project. Wait for the ERP upgrade. Wait for the AI Act clauses to settle, given that Article 50 transparency obligations apply from 2 August 2026 and the high-risk regime kicks in on 2 December 2027.

Waiting has a price. Every quarter the interface stays broken, your reps process fewer queries per hour than a competitor who fixed theirs. Every new hire absorbs the ten-minute tax as normal. Every abandoned pilot in your sector, and there are many, makes the next budget conversation harder.

If your team runs manual lookups against documents you already own, the fix isn't more data. It's a better question format on top of the data you have. The operators who get this shipped in the next twelve months will be the ones who stopped confusing storage with access.

The data was never the problem

Why most AI pilots don't fix this

The three-second version

The cost of leaving it alone

Sources

Key-person risk is an AI use case

The questions your stack cannot answer

The take-home test is dead. What replaced it.