We have all had the same conversations about dirty data. Messy data attributes are preventing better personalization, better AI tooling, and expanded distribution networks. And that can all be true.
But practitioners are telling a different story. They are not waiting for clean data to start using AI. Instead, they are forging ahead using AI as an enabler, while putting guardrails in place to achieve a balance of speed and risk. As one executive told us, “While we see that AI does make mistakes, it operates at such a large scale that we see a ton of value.”
In this report, we look at four best practices we heard repeatedly from digital executives at manufacturers and distributors: scoping AI narrowly instead of waiting for a perfect foundation, letting AI do the merging and categorizing work that used to require a data scientist, focusing initiatives on channel partner needs, and putting governance in place as the initiative scales.
Scope AI to the Data It Actually Needs
When executives tell us they are waiting to invest in AI because the data is not ready, our first question is: ready for what?
The instinct to wait for a fully clean, centralized data layer before getting started on personalization, customer service, or product enrichment is understandable. But practitioners who have built successful AI-based initiatives consistently say the best way to make these projects work is to keep the scope small. Focus on the specific slice of data a given AI use case needs, and build from there.
A Vice President of Engineering at a large foodservice distributor dismissed the idea that companies should think about “cleaning their data.” Instead, he suggested, they should be collecting ideas from teams about projects where AI can help them build a foundation of clean data for one specific use case.
I am creating digital solutions for specific audiences, not trying to serve the needs of the broader enterprise to query all of our data.
VP of Engineering, large foodservice distributor
A senior eCommerce director at a life sciences manufacturer described how bringing in a Product Information Management (PIM) platform forced them to confront difficult questions not just about data readiness, but about how to use AI to enrich that data for future applications.
We are literally in the thick of preparing our data for the website, and we are using AI in two ways: to enrich the data now, and to get it ready for future AI applications. We are taking a multi-pronged approach, using AI to get the data ready while also thinking about how it can be leveraged in a future state.
Senior eCommerce Director, life sciences manufacturer
The pattern is consistent: define the business goal, select only the data that use case requires, then turn what works into reusable, higher-quality data for future projects.
Target AI Orchestration but Start Now on High-Value Use Cases
AI orchestration, where a central AI connects models, assigns tasks, and governs compute budgets across the enterprise, represents enormous long-term value. Eventually every business unit will be aligned on shared data and consolidated models. But that future is not today’s priority for most B2B firms.
The more common play among B2B practitioners right now is building AI use cases with short payback periods and high ROI. They have standardization, but just enough to move forward. They do not overstandardize, especially when AI projects have not yet demonstrated clear wins. Standardize incrementally after the business sees measurable value. It is fine for each business unit to use a different model for now. Orchestration can come after value is demonstrated.
A CIO at a global technology distributor described using AI to write the kind of data-analytic scripts that once required a trained data scientist to merge disparate sources, structured and unstructured alike. The same tools are now handling product categorization at scale.
AI is really good now at writing powerful data-analytic scripts, and you do not need many data scientists to do it. It also enriches and categorizes data, taking a full product set and organizing it by assortment, color, category. That used to take people offshore a lot of time. AI makes mistakes, but it operates at such a large scale that it delivers a ton of value anyway.
CIO, global technology distributor
The life sciences eCommerce director described a parallel challenge: unstructured and inconsistent data living in PDFs and web pages, with no shared definitions across five different product portfolios.
We have a lot of data that is completely unstructured, living in PDFs or on web pages. If you have five different product portfolios, you can end up with five different definitions for the same thing. We are using AI to help get everyone speaking the same language and consolidate that data into something structured.
Senior eCommerce Director, life sciences manufacturer
The foodservice executive added that AI is not only cleaning data that already exists. His analytics team is using it to generate net-new data sets. “As an example, they will figure out the menu for a restaurant, figure out what ingredients they would need to buy to make it, then figure out what we are selling them and, more importantly, what we are not selling them, so we can go after that share of wallet.”
Focus on the Needs of Your Channel Partners
Practitioners consistently reported that the growth of AI tools has re-energized their teams, with team members generating a high volume of ideas for how to apply the technology. One senior executive at an electronics manufacturer noted that the cost of distributing Copilot access had become prohibitive, raising the question of how to balance enthusiasm with the realities of building and running these products.
The answer practitioners kept returning to: focus on the customer. And in B2B, that often means starting with the channel partner.
One senior leader told us that because their company primarily sells through distributors, they began their AI data journey by asking how they could improve the data their distributors receive. They have been using AI tooling to enrich product data provided to distributors, which helps those distributors improve their GEO search performance. A downstream benefit: distributors can build chatbots on top of that enriched data to improve the customer search experience.
That initiative ultimately required hiring a full-time AI data scientist. And the foodservice executive was clear that human expertise remains irreplaceable in this work.
I would not just point an LLM at a set of Snowflake data and hope it figures out what we need.
VP of Engineering, large foodservice distributor
The pattern across organizations: teams closest to the customer or channel partner surface the most valuable AI use cases. The data science team’s job is to build the infrastructure to execute on those ideas repeatably.
Govern the Process as You Scale
None of the best practices in this report argue for moving fast without guardrails. The practitioners furthest along with AI-driven data work were also the most deliberate about controlling it.
A marketing executive at a global automotive manufacturer described an AI concierge experience that works well in large part because the underlying vehicle and specification data was already clean and centralized. Even so, every AI initiative at the company goes through a formal review process.
AI is something that is very controlled. All initiatives have to be submitted and analyzed. In the early days it was a bit of the wild west, with a lot of individual initiatives, and that became worrying. Now there is a lot more control: solution review boards, penetration testing, making sure the answers do not reference competitors, proper vocabulary. And because of the EU AI Act, we are being asked to be even more attentive to compliance.
Marketing Executive, global automotive manufacturer
The life sciences practitioner described a four-person governance group that convenes when questions arise about AI projects: the data governance committee, a solution architect, a commercial delivery IT lead, and an eCommerce lead. That structure ensures technical and business needs are met while maintaining consistency across initiatives.
Governance also means cost governance. The technology distribution CIO pointed out that AI spend based on token usage can escalate in ways that catch finance teams off guard.
Regardless of who you are with, there is a token cost. These providers would love for you to just consume, and the costs can be astronomical, hundreds of thousands of dollars. You need a governance framework, and the mindset behind approving an ROI has to account for the long-running cost of running these models, not just the cost per token at the outset.
CIO, global technology distributor
His recommendation: teams should meet bi-weekly or monthly to review project costs and reevaluate scope when spending exceeds budget. The discipline applied to the data itself, scoping, reviewing, and controlling cost, has to scale alongside the AI initiative, not get bolted on after the fact.