Manufacturing Intelligence: From Tribal Knowledge to Organisational Memory

Manufacturing businesses operate at the intersection of precision and uncertainty. Every inbound enquiry triggers a chain of decisions: feasibility, costing, capacity, margin, timeline. In most organisations, these decisions still depend on manual judgment, fragmented spreadsheets, and a handful of experienced people who carry the reasoning in their heads.

This has always been the case, and for a long time, it worked well enough. But as demand cycles shorten and product complexity increases, the approach stops scaling. Sales teams cannot respond quickly because engineering teams are occupied with routine estimation. Engineering teams cannot focus on complex work because they are pulled into repetitive costing tasks. And the knowledge that holds the whole operation together remains trapped in emails, drawings, and conversations that nobody will ever search through.

We have spent considerable time working on this problem at Mitochondria, and what we have learned has shaped a product called Cortex. This article describes how we think about manufacturing intelligence: not as an automation problem, but as a problem of visibility into reasoning.

The real bottleneck is not speed

The obvious framing is that manufacturing quoting is slow. Enquiries arrive with incomplete specifications, unclear drawings, and uncertain volumes. Teams go back and forth to gather missing inputs. Senior engineers get pulled into estimation work that junior staff could handle if they had the scaffolding. Turnaround times stretch. Conversion suffers.

All of this is true, but the speed problem is a symptom. The underlying issue is that the reasoning behind manufacturing decisions is invisible. An engineer evaluating a product makes dozens of judgment calls: process routing, tolerance classes, material selections, surface treatments, and exception handling. These decisions draw on years of experience and produce a cost sheet or an evaluation deck at the end. But the thinking that connects the inputs to the outputs is never captured. It happens, it produces a result, and it vanishes.

This matters because when that engineer is unavailable, or moves on, or retires, the reasoning leaves with them. The next person starts from scratch, makes different assumptions, and takes longer to reach a conclusion that may or may not be as good. Multiply this across an organisation, and you have a structural fragility that no amount of process documentation will fix, because the problem is not that people refuse to document. The problem is that documentation is a separate activity from the work, and it will always lose out to the next incoming enquiry.

What Cortex does differently

Cortex is Mitochondria's manufacturing intelligence product. It sits between the document dump an engineer receives and the finished output they need to produce.

When an enquiry arrives, Cortex takes the incoming documents (specs, drawings, BOMs, customer requirements) and organises them into a structured starting point. Not a finished product, but organised raw material that would otherwise take hours to assemble manually. Then it works with the engineer through conversation, asking specific questions about what remains: the process routing for this component, the tolerance class, the coating, and the exceptions that apply. The engineer answers. Cortex assembles the evaluation deck or cost sheet.

The part that matters is what happens underneath. Every answer the engineer gives is stored as a decision trace. Over weeks and months, the organisation accumulates a searchable, structured record of how products get evaluated and costed: which decisions recur, which exceptions keep appearing, which assumptions hold and which ones fail. The engineer never documents anything extra. Using Cortex is the documentation.

We call this an organisational brain, though internally it is known as a context graph. It is not a static repository. It is a living record that grows from the work getting done, and it becomes more useful with every enquiry processed.

Inbound sales as an intelligence problem

The bottleneck often starts before engineering gets involved. Inbound enquiries arrive through email, WhatsApp, web forms, and sometimes phone calls. Information is scattered, incomplete, and inconsistent. Someone has to read everything, figure out what is missing, go back to the prospect, wait, follow up, and eventually assemble enough clarity for engineering to begin.

Cortex can handle much of this upstream work. It interprets unstructured inputs (text, uploaded files, handwritten specification images, photographs of existing parts), identifies what is missing, and collects what it needs through conversation rather than rigid forms. This transforms inbound sales from a manual intake process into an intelligent qualification layer that hands engineering a structured brief rather than a pile of emails.

Costing without replacing judgment

The costing step is where most AI promises fall apart, because manufacturing costing depends heavily on experience that has never been codified. What does this process routing actually cost? What is the realistic yield for this material at this tolerance? What did we learn the last time we quoted something similar?

Cortex does not attempt to replace this judgment. It structures the process around it. It maps enquiry inputs to internal costing parameters, surfaces comparable historical jobs and their outcomes, highlights where assumptions need validation, and flags cases where confidence is low enough to warrant human review. The engineer still makes the call. Cortex ensures the call is made with the best available context, rather than relying on memory alone.

When enquiry handling and costing are structured this way, the quote becomes a natural output of the process rather than a separate assembly step. Pricing aligns with internal rules. Assumptions are documented because they were captured during the workflow, not after it. Timelines reflect actual capacity. The quote is reviewable, adjustable, and traceable back to every decision that produced it.

Knowledge that survives people

The most fragile aspect of any manufacturing operation is its dependency on specific individuals. Why did a past job succeed? Where did costs escalate unexpectedly? Why was a particular margin accepted for one customer but not another? These answers exist in someone's head. They are vulnerable to resignation, retirement, or simply forgetting.

This is where the organisational brain becomes a strategic asset rather than a technical feature. A Cortex deployment that has processed a thousand enquiries holds a structured record of how the organisation thinks about its products. New engineers can query it. Senior engineers can validate against it. Management can see patterns that were previously invisible: which product categories are most margin-sensitive, where estimation accuracy breaks down, and which customer types require additional validation.

The knowledge becomes accessible without anyone having to write it down deliberately. It accumulated as a byproduct of the work.

Governance in a commercial environment

Manufacturing intelligence systems handle sensitive data: pricing logic, supplier relationships, margin structures, and intellectual property embedded in process knowledge. Any system operating in this environment must be governed accordingly.

Cortex operates within defined boundaries. Role-based access determines who sees what. Every decision trace is auditable. There is a clear separation between what the system recommends and what it executes. The system can surface a comparable historical job, but it cannot commit to a delivery timeline without human approval. These boundaries are architectural, not policy-based, meaning they hold regardless of how the system is prompted or what questions are asked.

What this means in practice

The manufacturers we work with are not looking for AI because it is novel. They are looking for it because their current approach to evaluation and costing does not scale, their institutional knowledge is concentrated in too few people, and their response times are costing them business.

Cortex addresses these problems by making engineers faster at the routine work so they can focus on what is genuinely complex, and by capturing the reasoning that currently disappears after every completed job. The first effect is measurable in weeks. The second compounds over the lifetime of the deployment.

—

Mitochondria is an agentic AI product company based in Amsterdam, with operations in Pune. We work with organisations across Europe, India, Southeast Asia, and Australia. ISO 27001 certified. GDPR- and DPDP Act-compliant by architecture.