A successful pilot puts an executive in an awkward position. There is finally evidence to discuss, but also pressure to make it stand for more than it proves.
An ordering assistant works with one product range. A presentation turns it into a promise about every commercial process. Customers adopt it in one market. The business case assumes the same response elsewhere. The distance between those statements is where the next investment can go wrong.
The useful question is how far the evidence travels. It helps turn enthusiasm into a scoped decision: what can we reuse, what changes, and what must we learn before committing further?
A public case with a useful boundary
At Holcim, my contribution was strategy and delivery leadership within the team developing an AI-assisted ordering experience. The published McKinsey interview describes a Spanish cement-ordering pilot using a channel customers already knew. It reports digital-ordering adoption moving from 25% to 93% and about 66% of initial order proposals being accepted. It also describes testing problems with unsuitable product suggestions and the option to reach a human agent.
Those are reported pilot outcomes from a published interview, not an independent audit, a global rollout result or a benchmark for another company. They concern customer adoption and initial proposals, not a quantified enterprise return.
I keep that scope visible because it is what makes the case useful. A concrete result gives the next team something to examine. “Autonomous enterprise” gives it much less to work with.
The case summary separates the published evidence from my contribution. It is an example of a commercial workflow, rather than proof that an industrial company can hand all its decisions to agents.
Ask what produced the result
Before reusing a solution, understand the conditions around it. The customer problem, channel, product information and available support all matter. A system can work because the task is well bounded and customers find the interaction convenient. Neither property automatically survives an expansion.
Adoption answers an important question: did customers use the offered route? It leaves others open. Did the complete order need less correction? Did support effort fall or move elsewhere? What did the service cost to operate? Were occasional failures acceptable for the intended scope?
Initial proposal acceptance is also useful, but it should not be renamed accuracy, retention or completed-order success. Each measure has its own definition. A funding decision needs the measures that connect to its business promise, including the work after the customer accepts.
A serious review can acknowledge a promising result and ask these questions at the same time. Doing so makes the case more credible, because the scope is clear enough to test.
Expand one boundary at a time
Here is a hypothetical next investment. An industrial company wants to take an ordering assistant from a straightforward product line into a business with many variants, delivery constraints and customer-specific agreements.
Some capabilities may transfer: conversation handling, identity integration, a customer interface and monitoring. The decision logic may change substantially. The new service needs current specifications, pricing authority, valid delivery options and a way to recognize requests that require specialist help.
Start by listing the differences, rather than describing the second rollout as a copy of the first. A new product range may create more risk than a new language. A familiar channel may still conceal a different buying process. A customer with several sites may need authorization rules that a single-site customer did not.
Choose the smallest expansion that tests a consequential difference. If the uncertainty concerns product suitability, adding an easy geography does not resolve it. If it concerns customer permissions, a demonstration using employee accounts proves little. The next scope should generate evidence the following investment actually needs.
Keep commitments authorized
An industrial order is a commitment involving product, quantity, price and delivery. Natural-language fluency does not confer authority to make that commitment.
Define which information the assistant can interpret, which proposals it can prepare and which actions require confirmation or review. Keep approved product and account constraints enforceable through the connected systems. When relevant information is missing, the service needs a useful handoff, not an invented answer.
The human receiving that handoff needs context: the customer's request, what was checked, what remains uncertain and what action is required. Escalation works only if someone can respond in time. An unattended exception queue turns a controlled design into customer delay.
As authority grows, test the failure paths again. A draft that can be corrected before sending has a different consequence from an order that has already reached dispatch. Recovery, customer communication and ownership belong in the expansion decision.
Fund the whole service
The economic comparison should include integration, data maintenance, review, support and the cost of corrections. It should also establish what the existing process costs and where the proposed service changes it.
Avoid assigning a cash value to every minute unless the organization can explain how the saved capacity will be used. Faster ordering may improve convenience, support growth or remove a bottleneck. Each is a legitimate possibility, but they are different business cases and need different evidence.
Make the next commitment reversible where possible. Set a bounded scope and explain what result would justify expanding, revising or stopping it. The decision should name the person who can act on that result. A pilot with no route to a funding choice can become permanent experimentation.
The Executive AI Mandate is a useful starting document for those responsibilities. It keeps the business claim, operational boundary and next decision together.
Give leadership the right level of confidence
The board does not need every technical detail. It does need to understand what was observed, what is inferred and what the new scope still assumes. A single blended success number obscures those distinctions.
I would bring a short comparison: the scope already tested, the proposed expansion, the important differences and the evidence planned to resolve them. Where the pilot does not answer a question, say so. Where a capability can be reused, make the reason clear.
That is a stronger basis for investment than a market forecast or a label suggesting inevitable autonomy. It gives leaders a way to support progress without treating uncertainty as disloyalty.
A good pilot earns the next question. The next investment should answer it.