How to plan the rollout of a new AI tool in procurement

Buying a new AI tool is one decision. Making it work is many. This article covers what to plan before it goes live: who to involve, how to pilot it, and what to budget for.

How to plan the rollout of new AI tools in procurement

Key takeaways

Most AI projects in procurement don't stall because of the technology. They stall because of what surrounds it: a problem nobody defined, stakeholders who joined too late, a rollout that skipped the pilot, and costs that only appeared once the contract was signed. Start from a simple buyer's journey, bring everyone along the purchase-to-pay chain to the table, test with an experienced group, let trust decide what gets automated, and budget for the parts that don't show up in the demo. 

Where does a purchase really start? 

Purchasing doesn't start in procurement. It starts at a desk, a week before a new colleague arrives, when the office manager realises nothing has been ordered yet. 

They need a docking station, a headset, and a monitor, and they need them by this Friday. They look in the catalogue and can’t find the docking station. Or they find three that look similar and aren’t sure which one IT has approved. Or they don’t know a process exists for ordering single items at all. Sometimes the request reaches procurement late, and the team spends hours on something that should have taken a few minutes.  

The office manager might order from somewhere else, and the purchase never reaches procurement at all. It’s more common than it sounds. Before consolidating its indirect spend, the medical technology manufacturer BIOTRONIK measured a maverick buying rate of over 28%. And the cost goes beyond the price on the invoice. Every order that skips the usual route can bring indirect costs with it: a new supplier to set up, an invoice that doesn’t match a purchase order, a price nobody negotiated, and time spent chasing and correcting. 

That’s one reason teams start looking at AI tools: procurement is overloaded, and the promise of faster processes sounds very appealing. But the workload is a symptom. The problem often starts earlier, with someone who couldn’t find what they needed, or didn’t know where to look.

It's a useful place to start, because it changes the question. Not "Which AI tool should we buy?", but "Where does the need first occur, and what is getting in its way?" 

Which problem are you solving, and for whom? 

When AI comes up in a conversation with a procurement team, the answer to “what do you hope to gain?” is often less clear than the enthusiasm behind the question. Many teams know they want to do something with AI. Fewer can say which task it should take off their plate, or how they’d know it worked. 

It’s an easy place to end up. Leadership is asking about it, vendors are emailing about it, and colleagues are trying new tools every month. But a tool chosen before the problem is defined is hard to evaluate, because there’s nothing to measure it against.  

In Hackett’s 2026 Procurement Key Issues Study, the leading teams tie their AI investments to outcomes such as supply continuity, spend optimisation, and strategic business enablement. The starting point is the outcome, not the tool. 

A clear problem sounds different from a general ambition. “We want to use AI” gives a vendor nothing to work with. “New starters wait too long for standard equipment, because requests pass between three people, and we want to shorten that” does. You can test a tool against the second one. 

The simplest way to get there is to make the status quo visible first.

  • How does a manager order a docking station for a new starter today?

  • How many steps does it take, who touches it, and where does it wait?

Write it down, ideally with the time and cost of each step. It sharpens the problem, it shapes the list of things a tool would need to do, and it gives you the baseline you’ll want later for the business case. 

As you do, sort what you find into two lists: 

  1. Procurement’s effort: too many manual tasks, too many follow-ups.  

  2. User’s friction: they can’t find what they need, or don’t know the process.  

A tool aimed at the first can miss the second entirely. Once the problem is clear, the next question is who needs a say before anything changes. 

Who needs to be at the strategy table before you start? 

If you follow a docking station order from start to finish, you can map out the people who need a say in a rollout.  

  • The manager who requests it;  

  • purchasing, who source it;  

  • controlling, who watch the budget; 

  • the person who receives the parcel and books the goods receipt;

  • accounting, who pay the invoice;

  • and around them: IT, legal, data protection, HR, communications and,

  • where it exists, the employee representation team. 

Leaving any of them out can cause problems later: 

  • IT and data protection may raise concerns late, when changing course is more expensive. 

  • Communications may not hear about the tool, so nobody tells the wider company it exists. Procurement teams are often measured on usage – how much buying goes through the new channel – so a tool that nobody knows about can quietly miss that measure. 

  • Employee representatives, such as a works council may need to be involved. Requirements differ between countries and companies, and a tool that changes how people work is often a topic they care about. It’s worth finding out early what applies to you, and your own legal or HR team can tell you what’s required. 

The practical answer is simple: list every role between the person who needs the docking station and the person who pays for it, and ask each one what they’d need to see before they could say yes. Do that before you run a pilot, not after. 

How do you test a new tool before you scale it? 

One thing should be settled before anyone starts testing: leadership has already agreed. Once the right people are on board, the next decision is easy to rush:

Who uses it first? 

For most teams, the answer is a pilot: a small group tries the tool before everyone else does. Many organisations are doing it this way: the Hackett Study points out that 43% of organisations are actively pursuing AI deployment, but only 12% report large-scale implementation. Most are still running pilots or single-use-case deployments. That makes the pilot worth getting right, because it’s where you find out whether the tool is ready to scale.

What a pilot looks like depends on the company size. In a large company, it might be a single department or site; in a smaller one, a handful of colleagues. An experienced group is a good place to start: people who’ve used similar tools before, who are willing to give honest feedback, and who are given dedicated time for the pilot so it isn't extra work on top of an already full week.  

But experienced colleagues shouldn’t be the only voice. The tool will eventually reach team members who are less confident with technology, so include at least one of them, too. It also helps to include someone with a fresh perspective, who asks the basic questions on behalf of everyone else: Where do I click? What happens next? What if I get it wrong? Those questions are easy to skip in a group of experts, and they’re often the ones that show where the wider rollout will struggle. 

A few things help: 

  • Make it a privilege. Being chosen as a pioneer is motivating. A small incentive, even a small voucher, signals that the feedback matters. 

  • Build in feedback loops. Many short check-ins teach more than one long review at the end, and they mean the wider rollout benefits from what the pilot learned. 

  • Turn the pilot group into champions. They’re the first users, and they’re also the people who explain the tool to colleagues, which carries more weight than any announcement. 

  • Document what works and monitor it. Measure against the baseline you wrote down earlier. If the pilot shows a measurable time saving per person per day, that’s the evidence the next group needs to hear. If it doesn’t, you’ve learned that before it cost much. 

Agree at the start how the pilot ends, too: scale it, adjust it, or stop. 

What should run automatically, and what still needs human in the loop? 

It’s tempting to decide this once, on paper, at the start. In practice, it depends on trust, and trust builds in stages. 

Take the docking station again. Once it arrives, someone confirms in the system that the parcel was received, so the invoice can be paid. A tool that matches deliveries to orders and books that step automatically is appealing. But it comes with a reasonable worry: what if the system books something as received that never turned up, and the invoice gets paid anyway? 

So teams start with more manual work. A person stays in the loop, checks the output, and builds experience with how the tool behaves. Only when the results have held up do they automate the steps where it has proved reliable. This is where the pilot comes in. 

That’s a pattern worth borrowing for any AI tool. Sort the steps into three groups: 

  • Low-risk and easy to reverse: for example, suggesting a category for an incoming request. 

  • Needs a person to approve: for example, confirming a goods receipt or releasing a payment. 

  • Stays human: for example, negotiating with a supplier or choosing one. 

Revisit the list after the pilot, when you have results instead of assumptions. And however much runs automatically, a dedicated person stays accountable for the outcome. 

Will it fit what you already have? 

Most procurement teams are not starting from a blank page. They already work with several tools, and each new one has to fit in. The same Hackett Study found that 69% of organisations access AI through capabilities embedded in their existing procurement platforms. For many teams, the question is as much “What is already available to us?” as “What should we add?” 

That raises questions a demo rarely answers. Does this overlap with something you already pay for? How much effort is the integration, really? A standalone tool is simpler to introduce. One that has to connect to your existing systems makes people ask, reasonably, whether it’s necessary at all. 

Then there’s the learning curve. The easier a tool is to pick up, the more likely people are to use it, and usage is what decides whether the investment pays off. It’s a wider software problem, not only a procurement one: Zylo reports that, on average, 53% of SaaS licences go unused or aren’t used often enough to justify the spend. A tool that looks impressive and then sits unused is an expensive way to learn that lesson. 

Before signing, ask three things:

  1. Which existing tool does this replace or overlap with?

  2. Who would have to change how they work to use it?

  3. How will you tell in a few months whether they are? 

Which compliance questions should you ask early? 

For many European companies, where data is kept and who can access it is not a side question. At DPW Amsterdam 2026, data sovereignty came up repeatedly as a factor in which vendors procurement teams trust. Data protection and e-invoicing rules can shape which tools make the shortlist. 

For an AI tool, a few questions are worth asking before anything is signed: 

  • Where is the data stored and processed? 

  • Who can access it, including the vendor’s own staff and subcontractors? 

  • Is your data used to train the vendor’s models? 

  • What is logged? Is there an audit trail of who did what, and can you export it? 

  • If the tool touches invoices, how does it handle the invoice formats and retention rules you already follow? 

  • What happens to your data if you stop using the tool? 

Your own legal, data protection, and IT teams are best placed to answer these, which is one more reason they belong at the table from the start rather than at the sign-off. 

What will it really cost? 

The price on the quote is an easy number. The business case is harder, and it starts with the baseline you wrote down earlier: How long does the process take end to end today, and what does it cost? Without that, there’s nothing to compare against later. 

Then there are the costs that are easy to miss: 

  • Licences: Teams often start with a limited number and need many more once adoption grows. It’s worth asking what the price looks like at that point. 

  • Usage-based costs: Some AI tools charge by usage, often measured in tokens, the chunks of text the model reads and writes. So the bill can rise with not only how many people have a licence, but how much they use the tool. Ask how usage is measured, whether there’s a cap or an alert, and what a busy month would cost. 

  • Training and change management: Getting people to use a tool is work, and it takes time away from other things. In the Unite x Hackett’s Spotlight Report, procurement workloads are projected to rise 8% while headcount and budgets shrink, so that time competes with a growing backlog. 

  • IT effort: Integration can take longer than planned, especially where systems are older or user rights are complicated. 

  • Outside help: Mapping data, setting up user controls and connecting to existing systems often needs specialist support, and in our experience with procurement technology rollouts, teams sometimes only find that out mid-project. 

  • Limited vendor support; Consulting hours are often capped in the package, so extra help has to be bought. Ask what’s included and what’s billed by the hour. 

None of this means the investment isn’t worth it. It means the number you present to finance should include these lines, so there’s no surprise in month three. 

So what does a smooth first day look like for the new hire? 

The manager found the docking station in the catalogue, ordered it, and it arrived before the new colleague did. On Monday, they plugged in their laptop and got on with their day. Nobody thought about procurement – and that’s the point. 

Everything above exists to make that moment smooth: a problem worth solving, the right people involved early, a pilot that builds trust, automation that's earned, a tool that fits what's already there, and a cost picture that holds up. The teams that get value from AI in procurement are the ones who planned what happens when the tool goes live. 

 

This was the third post in our series on AI in procurement. In the previous two articles, we covered: 

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.

Why AI in procurement is only as good as the data behind it

Most teams know their catalogue could be better. Fewer can say what better means in practice. This article shows what good data looks like and how to keep it that way.

AI rollout questions procurement teams ask

Everyone along the purchase-to-pay chain: the people who request and receive items, purchasing, controlling and accounting, plus IT, legal, data protection, HR and communications. Employee representatives may also need to be involved, and requirements differ between countries and companies. List every role between the person who needs the item and the person who pays for it, and ask each what they’d need to see before they could say yes. Do that before the pilot, not after.

Start with an experienced group who are willing to give honest feedback and have dedicated time for the pilot. Include at least one less tech-confident colleague, and someone who asks the basic questions on everyone else’s behalf. Build in short feedback loops, turn the group into champions, and measure against the baseline you wrote down earlier. Agree at the start how it ends: scale, adjust or stop. And make sure leadership has already agreed, because a pilot isn’t the place to find out.

It depends on trust, and trust builds in stages. Sort the steps into three groups: low-risk and easy to reverse, needs a person to approve, and stays human. Start with more manual work, check the output, and automate only the steps where the results have held up. Revisit the list after the pilot, and make sure a named person stays accountable for the outcome.

The price on the quote is the easy number. Costs that are easy to miss include licences that grow with adoption, usage-based charges (often measured in tokens), training and change management, IT effort, outside help for mapping data and setting up user controls, and vendor support that’s capped by the package. Start from a baseline of what the process costs today, so there’s something to compare against later.

Ask where the data is stored and processed, who can access it (including the vendor’s staff and subcontractors), whether your data is used to train the vendor’s models, what is logged and whether you can export the audit trail, how the tool handles invoice formats and retention rules if it touches invoices, and what happens to your data if you stop using it. Your own legal, data protection and IT teams are best placed to answer these.

Is your catalogue ready for a new AI tool?