When a manufacturing company contacts us, the first question we typically hear is: "Which AI agent should we implement?"
That's the wrong question.
Not because it has no answer, but because answering it without context leads directly to projects that get built, get demonstrated, and never reach production. The team doesn't adopt them. ROI isn't measured. The project gets shelved.
What we do before proposing any solution is a structured assessment. Thirty minutes with the COO or CFO. No presentations. No demos. Just questions.
This article documents that checklist: what we ask, why we ask it, and what we look for in each answer.
Why the diagnostic matters more than the technology
In manufacturing, processes have a characteristic that makes them especially sensitive to poor automation decisions: they are chained together. An error in reading a production order affects inventory, which affects logistics, which affects billing.
A poorly designed AI agent doesn't just fail to help. It can introduce errors into a chain where the tolerance for error is low.
That's why the diagnostic is not a commercial formality. It is the technical and business foundation on which any solution is designed.
Block 1: Data Flow — What is the team working with today?
The first questions aim to understand how information moves through the operation.
What systems are active? ERP, MES, WMS, spreadsheets, email. In companies with 50 to 500 employees, the typical picture is a combination of all of them. The ERP exists, but coexists with Excel because the ERP doesn't cover everything — or because the team doesn't fully trust it.
Where is critical data generated, and where is it consumed? If production data is generated on the plant floor but the report reaches the CFO 48 hours later after manual processing, that gap is a clear candidate for automation.
What percentage of the data is structured? An agent working with structured data (tables, defined fields, APIs) is faster to implement and more stable in production than one working with PDFs, emails, or unstructured documents. Both are achievable. But implementation time and risk differ.
Block 2: Integrations — What can be connected without friction?
The second block evaluates real technical feasibility.
Does the ERP have an API, or does it only export CSV? This is not a minor detail. An ERP with an API allows an agent to query data in real time. An ERP that only exports flat files requires designing a synchronization process that adds latency and failure points.
Is there someone on the team who can maintain a basic integration? We're not asking for an engineering team. We're asking that at least one person exists who can run a script, review an error log, or escalate a problem. If that person doesn't exist, the solution design changes accordingly.
What integrations are already active? If the company already connects its ERP to its logistics platform, the integration work for a new agent is reduced. If everything is siloed, the diagnostic must reflect that in the roadmap.
Block 3: Manual Operational Load — Where is time most frequently lost?
This is the block where real use cases emerge.
What tasks does the team perform repetitively each week? Consolidating production reports, cross-referencing inventory data with open orders, responding to internal queries about order status, preparing reports for management. In mid-size manufacturing companies, these tasks consume between 15 and 40 hours per week of qualified personnel.
How many people are involved in those tasks? If a task is handled by a single person, the operational risk is high and the case for automating it is strong. If five people handle it redundantly, the case is equally strong but for different reasons: data is inconsistent and the cost is higher.
What is the cost of an error in that task? In manufacturing, an error in a purchase order can mean stopping a production line. An error in an inventory report can trigger a stockout. Quantifying that cost — even in ranges — allows you to prioritize what to automate first.
A concrete example: at an industrial components company with 120 employees, the operations team spent approximately 25 hours per week consolidating production data from three plants into a management report. The process involved exporting data from two separate systems, cross-referencing them in Excel, and formatting the result. With an agent that automates that consolidation, the conservative estimate is between 80 and 100 hours recovered per month, with near-zero human error in the data cross-reference. At a personnel cost of 25–35 €/hour, the monthly savings range is between 2.000 and 3.500 €, not counting the value of having the data available in real time rather than with a 48-hour delay.
Block 4: Success Criteria — What has to happen for this to be worthwhile?
The final block is the most important — and the one most often skipped in superficial assessments.
What outcome do you expect in the first 90 days? If the expectation is to reduce the time required for monthly financial close, that's measurable. If the expectation is to "improve efficiency," it isn't. Part of our work during the diagnostic is helping to formulate success criteria that can actually be measured.
Who will operate the agent once it's deployed? An agent that no one operates is an agent that stops functioning within three months. Identifying the internal owner before building anything is a necessary condition for the project to survive.
Is there an assigned budget, or are we in an exploratory phase? This is not a commercial qualification question. It's a design question. If the budget is limited, the first agent must have clear, fast ROI. If there is room for a broader project, the design can be more ambitious from the outset.
What we do with the answers
At the end of the 30 minutes, we have enough information to do three things: identify the use case with the highest potential ROI, estimate implementation time within a realistic range, and determine whether the project has the conditions to reach production — or whether there are obstacles to resolve first.
If the conditions aren't in place, we say so. There's no point in starting a project that will end up in a drawer.
If the conditions are in place, we propose a first, well-scoped agent with defined success criteria and a timeline of 6 to 10 weeks to reach production.
The diagnostic is not a formality. It's the difference between a project that generates ROI and one that generates a slide deck.
If you'd like to apply this checklist to your operation, you can request a free diagnostic. The form takes less than two minutes and doesn't require scheduling a call right away.
