Key takeaways
AI can help indirect procurement teams work faster, but only if the catalogue data underneath it is clean, structured, and governed. Vague names, missing images, wrong manufacturer details, and uncontrolled access aren't small annoyances – they're the difference between AI that saves time and AI that quietly wastes it. Teams need to start with the catalogue data, only then think about AI.
A search that should take seconds
Picture a category manager who needs work boots for the warehouse team. She opens a general marketplace and types "black work boots". What comes back is a sea of options – black boots, black shoes, black trainers, yellow work boots, and dozens of things that aren't boots at all. She spends the next twenty minutes narrowing it down by hand, opening tabs, comparing specs that aren't listed the same way twice.
It's easy to dismiss this as a small irritation. You might think, it’s ‘only’ twenty minutes. But it rarely happens once. Multiply it across every buyer, every category and every week, and the cost shows up in places that are harder to see than a search bar.
What does a slow search actually cost the business?
Each unclear listing means another manual check. Is this the right specification? Is it from the supplier we have an agreement with? Does the price match what was negotiated? When the right item is hard to find, people buy the one they can find, and policy exceptions follow. Those exceptions then have to be approved, explained or corrected later, and every single one of them adds to the cost of the process.
The bigger loss is the least visible. Hours spent checking and correcting are hours not spent on negotiating, managing supplier risk or shaping category strategy. In a team that is already stretched, strategic capacity gives way to routine verification, and process costs rise without anyone approving a higher budget.
In a well-governed catalogue, the same search – height, material, safety rating, colour – returns a short, relevant list. Fewer results, but every one of them close to what she asked for.
How an integrated catalogue helped BIOTRONIK cut process time and costs
The team moved its indirect purchasing onto a single platform, with an integrated catalogue and its existing framework agreements linked. The results:
- Average process costs per order fell by 72%
- Maverick buying dropped from over 28% to under 5%
- Average order time went from two hours to 20 minutes
What do we mean by "data quality"?
Data quality is a phrase that gets used loosely, so it's worth being precise about what it covers. A well-governed catalogue entry gets a number of things right, consistently, across every article.
The first group is what lets a buyer – or an AI tool – compare like with like:
Classification – every article is in the right category under one consistent taxonomy, so the same type of product isn't filed three different ways.
Technical attributes – the specifications that tell similar items apart, such as height, material, and safety rating, held as structured fields rather than buried in a description.
Units of measure – order and content units that follow market standards, so a box, a pack, and a single piece are never confused.
Pack quantities – a clear statement of how many items come in each unit, so prices can be compared per piece as well as per pack.
Validated identifiers – manufacturer name, manufacturer number, and GTIN that are not just present but match the article and each other. A manufacturer number that belongs to a different product is arguably worse than no number at all, because it looks correct.
Price validity – prices that are still valid and reflect what was negotiated, not figures nobody has revisited since the last renegotiation.
Availability – whether an item can be delivered and how quickly, so a buyer isn't offered something that can't arrive when it's needed.
Provenance – who produced the item and where the information comes from. The supplier is not always the producer, and that distinction matters for compliance and sustainability reporting.
The second group is the content that lets a buyer trust the listing:
A clear, unique name – not a string of capital letters, not a name that could describe three different products.
At least one accurate image – not a placeholder, not a competitor's product, not a company logo standing in for the real thing.
A description that matches the article – written in the right language, free of pricing or supplier details that shouldn't be there and structured well enough to read easily.
Sensible keywords – market-standard terms rather than a pile of adjectives, and not so many that the signal gets lost.
Legally required documentation where it applies – safety data sheets for chemical products and compliance with regulations like RoHS or packaging law, kept current rather than outdated.
None of this is new. It's the reasonable level of information a buyer would want before clicking "add to basket" – and the basic information an AI tool needs before it can make a useful recommendation.
Victoria Folbigg, author of Unite's 2026 Expert Report, described one CPO’s experience: 81% of indirect materials were uncoded. Another leader reported that a delayed codification programme postponed marketplace adoption by months. These individual accounts show how data foundations can become an implementation constraint.
Catalogue governance is about more than clean data
Clean fields are one part of the picture. The other part is a question a lot of teams haven't formally answered: who is allowed to order what, from where?
Many procurement teams arrive with framework catalogues already in place – prices and terms agreed directly with a supplier, sometimes before they ever start working with a platform. These can be brought in and combined with the wider marketplace, but someone still has to decide the rules: if the team needs screws, do they have to buy through the approved catalogue, or can they buy anywhere? Who has the authority to make that call, and is it written down anywhere?
This is what catalogue governance is – a set of decisions about access, ownership and control, not just a set of clean product records. A catalogue can have perfect data and still be poorly governed, if nobody has defined who can buy from it, who looks after it and what happens when something changes. In practice, that comes down to five things:
Accountable owners – every catalogue has a named person or role responsible for it, not a shared inbox or a team that assumes someone else is watching.
Approval of catalogue changes – new articles, price updates, and removals pass through a defined approval step before they become orderable. This is what stops a handful of unreviewed additions from slipping in unnoticed.
Review cadence – a set rhythm for checking catalogues against what was negotiated and against the market, rather than waiting for a problem to surface. The right frequency depends on the category and how quickly it changes.
Auditability – a record of who changed what, when and why, so the team can trace an article or a price back to its source and answer questions from finance, compliance or auditors.
Escalation rules – a clear path for what happens when something doesn't fit, such as a price that looks wrong or a supplier adding articles outside the agreed scope. Who decides, and how quickly?
Without these, even a clean catalogue changes. Without these, even a clean catalogue drifts. With them, keeping it clean becomes routine.
When a good deal stops being a good deal
A supplier who agreed to a generous rate eighteen months ago might, without anyone noticing, adds a few extra articles that were never part of that deal – new variants, new pack sizes, small additions that sit just outside the original scope. If they become orderable without review, they create spend leakage that's invisible until someone goes looking for it.
This is something that might happen to catalogues left unchecked, and it's exactly why teams benefit from being able to challenge and benchmark their own catalogues – comparing what's negotiated against what's available on the wider marketplace, and deciding, with real numbers in front of them, whether the original deal still holds up.
AI can only work on clean data
AI is not a shortcut past the work described above. If you feed it a catalogue with missing images, mismatched manufacturer data, and vague descriptions, it won't fix any of that on its own. It will do what language models do with incomplete information – fill the gaps with something plausible, and hand it back with total confidence.
That's the part that catches teams by surprise. Bad data doesn't usually make AI fail loudly. It makes AI fail in the background, producing a recommendation or a search result that looks reasonable right up until someone checks it. Catching that after the fact will cost more time than the cleanup would have taken in the first place.
Training people to write better prompts as part of an AI onboarding programme is important too. "I'm looking for screws" will always return something less useful than a prompt with the material, thread size, and length specified. But no amount of prompting fixes a catalogue where the underlying product data was never right to begin with. That part is on the team, not the AI tool.
Orchestration without data is a promise, not a process
A growing number of procurement tools offer orchestration, guiding teams through the buying process – prompting the next step, flagging what's missing, and automating approvals. It's a useful idea.
But if the data behind those steps is scattered or ungoverned, the guidance may point a buyer towards a product with the wrong specifications, a supplier that isn't preferred, or a price that is out of date.
Orchestration can only guide a purchase as well as the product, supplier, contract, and policy data behind it allows, which is why that data needs to be reliable first.
Data comes before AI
Most organisations still have some way to go. At DPW Amsterdam 2026, HFS Research presented a survey of 150 procurement leaders in which only 4% said their data is unified, trusted and continuously feeding AI-driven decisions.
Read our key learnings from DPW Amsterdam.
None of this means every procurement team needs to rebuild its catalogue from scratch before it can think about AI. It means starting with a realistic look at what's there, using the benchmark comparison to decide what needs attention first.
Rather than aiming for a perfect catalogue across the whole organisation, start with a bounded pilot: one category or one use case, with data-readiness criteria agreed up front, such as complete attributes, validated identifiers and current prices. Fix what falls short of those criteria first, then widen the scope once the pilot meets them. Some teams can do this review internally. Others bring in outside expertise to challenge specific catalogues and flag what's changed, which is a reasonable option rather than a sign of falling behind.
Either way, the order doesn't change: start with the data. Then think about AI.
This was the second part of our AI in procurement series. For more insights, check out our previous article:
Six things procurement teams need before any AI tool
AI can lighten a heavy procurement workload, but only when the groundwork comes first. Here is what that looks like in practice.
Frequently Asked Questions
In a procurement catalogue, data quality means every article is recorded consistently and accurately enough for a buyer, or an AI tool, to compare like with like and trust the listing. That covers consistent classification, technical attributes held as structured fields, units of measure and pack quantities, validated identifiers such as manufacturer number and GTIN, and current prices. It also covers clear names, accurate images, matching descriptions and any legally required documentation.
Catalogue governance is the set of decisions about who owns a catalogue, who is allowed to order from it and what happens when something changes. In practice it comes down to five things: accountable owners, approval of catalogue changes, a review cadence, auditability and escalation rules. A catalogue can have clean data and still be poorly governed if nobody has defined these rules.
Procurement automation acts on whatever the catalogue contains. Tools that prompt the next step, flag what’s missing or automate approvals can only guide a purchase as well as the product, supplier, contract and policy data behind them allows. If that data is scattered or ungoverned, automated guidance may point a buyer towards the wrong specification, a supplier that isn’t preferred or an out-of-date price.
No. AI is not a shortcut past the clean-up. Given incomplete data, it fills the gaps with something plausible and returns it with confidence, so errors often surface only when someone checks. Better prompts help, but they can’t fix product data that was never right, and fixing it remains the team’s job.
Start with a bounded pilot: one category or one use case, with data-readiness criteria agreed up front, such as complete attributes, validated identifiers and current prices. Fix what falls short, then widen the scope once the pilot meets the criteria. Benchmarking existing catalogues against a wider, well-maintained marketplace helps decide what needs attention first, and teams can do the review internally or bring in outside expertise.