• Resources
• Events, Webinars

Data Quality & AI in Procurement: A 20-Year Journey and What’s Next

Procurement has known about its supplier data quality problem for more than 20 years. Now AI is raising the stakes. In this session, Bertrand Maltaverne, Lead Analyst for Upstream Procurement at The Hackett Group, and Jeremy Reeves, Senior Vice President of Product at HICX, explored what is genuinely new about this moment, and what procurement needs in place before AI agents become the next 20-year problem.

Watch the full session, or read on for the key takeaways, live poll results and questions.

Key takeaways

• Awareness was never the problem. Supplier data quality has been a top procurement challenge for two decades. What keeps it unsolved is ownership.
• AI amplifies what it reads. A small error rate in source master data can compound into a far larger share of wrong automated decisions.
• AI agents bring a new wave of fragmentation. Narrow agents from different vendors may each hold their own version of the same supplier.
• Humans stay in the loop, for now. Most attendees would let AI flag and route supplier data issues, but not act on them.
• Start in the next 90 days. Audit your supplier data against AI readiness, put a procurement cost on the gaps, and name an owner.

A 20-year problem, and why awareness was never the constraint

Bertrand opened with a timeline of Hackett Group research. The headline has barely changed in two decades. In 2005, around 40% of procurement organizations had spend analysis tools but lacked visibility. By 2010, fewer than 30% could track their savings pipeline. In 2017, data quality ranked as the number three challenge.

Today, 73% of organizations cite data quality as a moderate or major concern, particularly for their Gen AI programs. Different decade, different technology, same shape of problem.

Hackett’s maturity data shows the scale. Around 63% of procurement organizations describe their supplier data as fragmented or only partially connected across systems. Only 6% say it is fully unified. Master data management has sat for years in what Hackett calls the critical development quadrant: high importance, low maturity.

The cause, Bertrand argued, is structural. Supplier data is scattered across multiple ERPs, source-to-pay suites, portals, contract management and sourcing tools. Each wave of technology promised to fix it: first ERP consolidation, then procurement suites, and now AI. Yet nobody owns the whole supplier record.

Poll 1: Where is your supplier data today?

When attendees were asked how they would describe their supplier data across their systems, most said partly connected, with core systems mostly aligned. Two said fragmented, with systems that regularly disagree. Just one described their data as largely integrated, with one authoritative source.

Bertrand was not surprised. People who attend a session like this are often already making progress and want to go further. Jeremy was surprised more attendees did not say largely integrated, since integration is far easier than it was a decade ago.

He added a caution, though. Teams working inside their own silo often believe their data is integrated, until they compare it with another department’s version of the same supplier.

Why knowing hasn’t fixed it

If organizations have known about the problem for 20 years, why hasn’t awareness translated into action? Hackett research points to four governance challenges that recur year after year. The biggest is simple: nobody owns the supplier record.

Ownership is often disputed between IT and procurement. IT tends to end up owning the data because it lives in IT systems, but Bertrand’s view is that supplier data is a business asset. IT can enable the process through technology, but it cannot own the accountability that comes with the record.

It goes deeper than the record, too. A supplier record is made up of many data elements, with no single steward attached to each one. Without that, there is no real authority to enforce standards, and data discipline falls away the moment nobody is watching.

The payoff for getting governance right is significant. Hackett research shows organizations with strong governance in place are 4.8 times more likely to follow enterprise-level data policies, and run reporting and analytics 7.1 times more efficiently than their peers.

The lesson, Bertrand stressed, is that world-class teams are not better at data because they bought better tools. They are better because they put governance in place and sustained it long enough for it to compound.

Where bad supplier data actually costs money

AI’s rise up the procurement agenda is not happening in a vacuum. Hackett’s 2026 research shows procurement workload up around 8%, staffing down almost 1% and operating budgets roughly flat. Technology spend is rising to close the productivity gap.

As a result, deploying AI-enabled technology is now procurement’s number 3 priority for 2026, and a new entrant to the top three. Around 43% of organizations are actively pursuing AI deployment, roughly double the previous year, although many are still at the pilot stage.

One supplier, five answers

To make this concrete, Bertrand followed one supplier across five systems. Finance’s ERP shows payment terms of Net 45. Operations shows Net 30. The procurement suite holds two records with different legal entities. The supplier updated its bank details in the portal, but nobody picked up the change.

Every system, and every person using it, assumes their record is the truth. Nothing reconciles them. Bertrand’s conclusion was that the shift needed is bigger than fixing a few fields. It means moving from a system of record to a system of trust, built on shared meaning, explicit context and full traceability.

Four ways bad data turns into cost

Bertrand then set out four failure modes. Duplicate vendors are the classic example: “UPS”, “United Parcel Service” and “UPS, Inc.” read as three separate suppliers in many systems. This is not an edge case. It happens whenever organizations do not invest enough in deduplication.

Duplicates and outdated records lead to incorrect recommendations, where AI repeats the same mistake while producing outputs that sound credible. Missing risk attributes create permanent blind spots, because nobody reviews what they don’t know is there. And inconsistent contract, vendor and item data allows negotiated savings to leak away unseen.

Jeremy added that these errors often stay invisible until the point of payment. By then, it is far too late to fix them cheaply.

Poll 2: Which of these has cost you most in the last 12 months?

Answers were spread out, but the top response was contract terms not applied at the point of buying, followed by reporting nobody fully trusts. Supplier risk nobody saw coming, and duplicate or conflicting supplier records, drew one vote each.

Both speakers were surprised. Jeremy’s own pick was unseen supplier risk. At one conference, two different companies approached him on the same day to describe a supplier account being hacked, followed by a fraudulent payment. It is the risk that worries him most, because you only find out after the fact.

Bertrand would have chosen duplicate or conflicting supplier records, because duplicates are often the root cause of the other three.

The AI paradox

This brought the session to its central idea. AI is often sold as the fix for bad supplier data, but AI depends on the very thing it is supposed to fix. Deploying AI on top of poor legacy master data does not correct errors. It amplifies them.

Hackett research shows AI success is constrained primarily by the quality of the data, not by the AI technology itself. And every new AI use case adds another data dependency that may go unseen.

How a 2% error becomes a 25% problem

Bertrand illustrated how quickly errors compound. A 2% error rate in source master data can grow to affect 25% of automated decisions as it moves through the AI pipeline.

During ingestion, nulls become zeros and records get dropped. Transformation carries existing errors forward. Training bakes them into the model weights, one of the most expensive forms of technical debt there is. Inference adds more, and at the point of action, bad decisions start creating new bad data.

The result is AI that produces confidently wrong answers at scale, with nothing in the system aware that its input has degraded.

Fix the data first, or fix it as you go?

Here the two speakers offered usefully different perspectives. Bertrand’s research position is that improving master data is a prerequisite for AI: not something to do in parallel, or worse, later.

Jeremy agreed that data has to be at the core, but emphasized that organizations do not need perfect data before they start using AI. If processes are in place to make sure incoming records are accurate, teams can run projects to cleanse existing data and fix it as they go.

The session slides captured the balance: you do not need perfect supplier data before you deploy AI, but you do need a governed record that knows its own provenance and confidence. Bertrand’s closing remarks agreed. Don’t aim for 100% perfect, because what is good enough depends on the AI use case.

The blind spot: do you know what your AI is reading?

Bertrand described four blind spots in how most organizations’ AI uses data. They cannot see which data fields each AI model relies on, or trace AI outputs back to source records. They lack a framework linking data quality to model performance, and an audit trail showing how transformations affected outcomes.

This is not unique to procurement. Hackett research shows 69% of finance leaders cite data quality as a major or moderate concern in the age of AI.

What is different from 20 years ago is permanence. A flawed training data set gets baked into a model and affects every downstream decision. There is no error message, and nothing looks broken.

What AI-ready supplier data looks like

Of Hackett’s six dimensions of AI-ready data, Bertrand highlighted three. Accuracy is the most dangerous, because wrong values teach models wrong patterns that look legitimate and don’t get questioned. Consistency matters because conflicting values across systems send contradictory signals. Uniqueness matters because duplicate records skew what the model learns.

Governance also has to change, not just tighten. Traditional master data was built to store records for reporting cycles and ERP transactions. AI needs decision-grade data, with context, relationships and confidence levels, so recommendations are explainable, defensible and scalable.

Poll 3: Do you know what your AI is reading?

Asked whether they could name which supplier data fields their AI tools rely on, four of nine respondents said roughly, but not documented. Three said AI is not touching their supplier data yet. One said yes, it is mapped, and one said no.

Bertrand had expected more “no” answers, noting that many people using AI tools have little idea what data sits underneath the recommendations they receive. Jeremy challenged the “not yet” answer: suppliers are almost certainly already using AI to respond to RFIs and profile questionnaires.

He also predicted that questions about where AI touches supplier data, already standard in software RFPs, will increasingly shift to procurement departments themselves, as has happened with previous regulations.

What’s next: a new wave of fragmentation

Looking ahead, Bertrand warned that procurement is entering a new wave of fragmentation. He was not talking about chatbots, but about task-specific, autonomous AI agents that are already appearing in sourcing, risk management and other procurement tools, very often each built by a different vendor.

Hackett data shows this is already underway: 71% of procurement organizations have adopted Gen AI, and 56% have deployed agentic AI. Each agent may be trained on different data and hold its own version of the same supplier.

An invoice matching agent and a risk agent both need the same governed view of a supplier, or they will contradict each other. Neither one is technically wrong. Each is acting on what it knows. The question procurement needs to ask today is how these agents will share context and data.

A problem with a precedent

There is a precedent. In the 2000s, organizations faced a similar challenge with knowledge management, as they looked for ways to help people share and access knowledge. The same ingredients apply today, only now for people and agents together: a shared semantic layer, governance with authority, defined escalation paths and cultural change.

Jeremy noted that the organizations that got knowledge management right in 2005 did not do it with better technology. They put the right processes in place.

An agent-ready supplier foundation, as Bertrand described it, rests on semantic understanding, context-rich reasoning, and a continuous learning loop where data improves with use rather than degrading. Governance works in three layers, with agents resolving around 85% of routine issues at the operational level and escalating the rest.

Where does the human stay in the loop?

Governance also answers the question of where people stay involved. Bertrand described four levels, from operational stewards and buyers handling day-to-day work, up to executives who see only high-value exceptions, fewer than 5% of actions. The principle is simple: AI generates recommendations, and the system of record enforces controls.

Poll 4: Would you let an AI agent update a supplier record without human review?

Half of respondents said they would let an AI agent flag and route issues, but not act on them. A quarter said nothing should happen without review, and a quarter would allow AI to enrich or update non-critical fields. Nobody would let an agent update banking, tax or compliance data unsupervised.

Bertrand called flag and route the safest position today, given the maturity of both the technology and organizations. He expects the line to move slowly but steadily, as fewer issues need routing to a human and AI gains autonomy.

Jeremy was more cautious. His team uses AI every day and still reviews everything. He expects a human reviewer to stay in the loop for around 90% of what AI touches for at least the next three years.

A three-year race

The gap between leaders and the rest is widening. Hackett research shows 66% of top performers plan to deploy AI against master data within three years, compared with 35% of their peers. More striking still, 21% of the peer group have no plans at all.

Bertrand’s view is that this is not a gap but a divide. In an AI-driven environment, the advantage compounds, which is why the next 90 days are worth more than the next three years of planning.

Questions from the audience

How do we build a business case for supplier data? What number should we put in front of finance?

Bertrand’s advice: start with an audit, since every organization is different, and use real numbers rather than estimates. Payment term discrepancies are a good example: negotiated at 45 days, but held as 90 in the paying system. Jeremy added: pick the category of supplier data causing most problems, and start there.

Any tips on getting started if we have five ERPs?

Jeremy’s answer was to bring supplier data under one layer of governance and fix incoming data first. Otherwise, any cleanup will be out of date within three months. Bertrand agreed: a governed layer on top of the ERPs creates a single source of truth, which then drives all five.

Does procurement have to work with the MDM team to make this work?

In short, no. Jeremy pointed out that solutions focused on supplier data quality, covering onboarding and ongoing monitoring, can do this without a large MDM project. Bertrand agreed that many organizations have no MDM team. One can help with best practice, but it is not a prerequisite, so don’t wait for one.

Where to start in the next 90 days

Asked what a procurement leader should do in the next 90 days, both speakers began with the same step: audit your supplier data against AI readiness.

Bertrand’s three priorities:

• Audit what you have against AI readiness: duplicates, missing lineage and inconsistent classification.
• Price the problem. Quantify the cost of your data debt from a procurement standpoint, not an IT one.
• Name an owner. Define stewardship and accountability per supplier domain, not a committee that reviews everything.

Jeremy’s three priorities:

• Audit against AI readiness and make sure your data is up to date and correct.
• Run a data classification exercise to understand what you have.
• Map your processes step by step, to decide what AI should automate, what could be automated in other ways, and what should stay with human review.

Bertrand’s closing thought summed up the session. Supplier data quality is not an awareness problem. It is a methodology problem. AI can help unify some things, but it is not a silver bullet. The time to start is now: move in steps, and don’t wait for perfect.

Trusted supplier data is not a cleanup project you finish before AI. It is the foundation that decides whether AI works at all.

Next steps for AI-ready supplier data

Not sure whether you have a process problem or a data problem? The Supplier Data Reality Check sets out five questions that help separate the two, and build the case for fixing it.

HICX is a Supplier Management Platform for large global enterprises. The HICX Supplier Data Foundation creates one governed record per supplier across your ERPs and connected systems, working alongside them rather than replacing them.

---

About the speakers: