The CSCO's 12-Point Checklist for AI Pilot Approval

TL;DR: Most AI supply chain pilots fail for reasons that are visible before approval, not after. Run these 12 checks first. They cover the problem, the data, the owner, the money, the people, and the exit. If a pilot cannot pass all 12, send it back before it consumes a year of budget.

Every CSCO now has a queue of AI pilot proposals on the desk. Some come from vendors, some from the planning team, some from a board member who read about agentic supply chains on a flight. The hard part is no longer finding pilots to run. It is deciding which ones deserve a yes.

The checklist below is the test we hear senior leaders apply in the TFEST community. It sits one step earlier than building the business case for the pilot. The business case answers "is this worth doing?". This checklist answers "is this ready to start?". Run it in a 30-minute approval conversation before you sign anything.

Frame the problem before the tool

Three checks decide whether the pilot is even pointed at the right thing.

1. The pilot names a real problem, not a technology

Approve only pilots that start from a named, costed problem. "We want to try agentic AI" is not a problem. "Purchase-order exceptions take 3 days to clear and tie up two planners" is. A pilot built around a problem has a natural success metric. A pilot built around a tool spends its first 2 months looking for one.

2. There is a baseline number

Require a measured baseline before the pilot starts. If you cannot state today's cycle time, forecast error, or exception rate as a number, you will not be able to prove the pilot worked. The baseline takes a week to pull and it is the single most useful artefact the team will produce. Without it, every result is an opinion.

3. The data is good enough to trust

Confirm the data the pilot depends on lives in a system you already trust. Ard Verboon, Chief Procurement Officer at Schneider Electric, has been clear in public that autonomous tools only work on transactional data that is already clean. If the pilot needs six months of data cleaning first, that data project is the real pilot, and it should be approved on its own terms.

Set ownership and the rules of the game

Three checks on accountability. A pilot with no owner drifts; a pilot with no agreed finish line never ends.

4. One named owner holds the decision rights

Name a single business owner inside the supply chain function. Not the data science team, not the vendor. The owner holds the metric, the budget, and the call on whether the pilot scales. Mourad Tamoud, Chief Supply Chain Officer at Schneider Electric, frames agentic systems around exactly this question of who holds decision rights when a machine starts to act.

5. The success threshold is agreed in writing

Write down the number that counts as success before the pilot runs. "A meaningful improvement" is not a threshold. "Cut exception-handling time by 30% within 90 days" is. Agreeing it up front stops the goalposts moving once the team is attached to the work, and it gives you a clean basis for the scale-or-stop decision later.

6. There are explicit kill criteria

Define what failure looks like, in numbers, before you start. A pilot needs a way to fail honestly. If acceptance stays under a set level for 60 days, or integration slips past a fixed date, the pilot stops. Kill criteria are not pessimism; they are how you free budget for the next idea instead of funding a zombie project for a year.

Control the money and the scope

Three checks that keep a first pilot small enough to be wrong.

7. The cost ceiling includes integration

Set a total cost ceiling that counts integration, data engineering, and internal time, not just the licence. The software is rarely the expensive part. Integration and data work usually cost more. A ceiling that ignores them is the figure that turns a €100K pilot into a €600K one by the third quarter.

8. Scope is bounded to one workflow

Restrict the first pilot to a single workflow in a single business unit. One product family, one region, one process. Bounded scope means a wrong answer wastes days, not the network. It also produces a result you can read clearly, because nothing else changed around it. Breadth is what you earn after the pilot works, not what you start with.

9. Security and data governance are signed off

Get security and data governance sign-off as part of approval, not after. Any pilot that moves supplier, pricing, or customer data through a new tool is a governance question. Clearing it up front takes one meeting. Discovering the gap after go-live takes the pilot offline and costs you credibility with the board on every future request.

Plan for people and the exit

The last three checks are the ones most pilots skip, and the ones that decide whether anything survives contact with the organisation.

10. There is a human-in-the-loop metric

Track acceptance rate from day one. Acceptance rate, the share of recommendations a planner actually approves, is the earliest honest signal of whether the pilot is working. A figure that sits low for weeks rarely means a weak model. It usually means the data or the scope is wrong, and it tells you so before the 90-day review.

11. You know who loses work, and you have told them

Map who the pilot changes before it starts, and bring them in early. Reginaldo Ecclissato, Chief Business Operations and Supply Chain Officer at Unilever, has spoken about change as the binding constraint on operating-model shifts. A pilot that quietly automates part of two planners' jobs will fail quietly too, unless those planners helped design it.

12. There is a scale-or-stop gate

Set one decision gate where you scale, fix, or stop. A fixed date or milestone forces a real choice. At the gate you compare the result to the threshold from check 5, look at acceptance rate, and decide. Most pilots that lack this gate do not get killed; they simply fade, which is the most expensive outcome of all because nobody learns from it.

A pilot that clears all 12 is not guaranteed to work. It is guaranteed to teach you something useful inside a quarter, for a known cost, with a clean decision at the end. That is the bar worth holding. The agentic-AI roundtables on the TFEST26 agenda are built around the same question senior leaders keep returning to: not whether the technology works, but whether the organisation is ready to let it.

Save your seat for TFEST26 in Berlin, December 1 and 2, 2026

The CSCOs who get the most from AI in 2026 are not the ones who approve the most pilots. They are the ones who approve the right few, with a baseline, an owner, and an exit, then scale the ones that earn it.

— TFEST26 Editorial Team