Summary

A contract chemical manufacturer competed on responsiveness and was losing on it. Customer inquiries arrived asking whether a product could be made to a given specification, at what cost, and with what risk. Answering meant establishing whether the company had made something comparable before, what the formulation was, what went wrong, and what the batch data showed at scale.

That knowledge existed. It sat in two estates that had never been connected. Formulation records, trial reports and deviation histories on one side, largely unstructured. Batch, process and quality data on the other, structured and queryable by a different team using different tools. The people who could bridge them were four senior formulators, two of them within a few years of retirement.

We deployed grounded retrieval across the formulation estate alongside natural language querying over historical batch data, with customer partitioning enforced at retrieval time. The system finds comparable prior work, cites the actual record, and shows what the process data did. It does not propose formulations. That boundary is deliberate, and it is what makes the system usable in a business where a confident invention is a safety event rather than an inconvenience.

Business Challenge

Summary: Answering a customer inquiry required searching two disconnected knowledge estates through four people, and the response time was costing contracts the company could have won.

Responsiveness is the product

Technical capability is broadly comparable across a shortlist. What separates suppliers at the inquiry stage is how quickly and how credibly they answer. A well-supported response in four days against a competitor's three weeks changes the outcome before any commercial discussion starts.

Turnaround was running into weeks, and the reason was never a lack of capability. It was the time taken to establish what the company already knew.

The knowledge sat in two estates that did not meet

The formulation estate. Development records, trial reports, scale-up notes, deviation investigations, technical data sheets. Largely documents, searched by chemists who knew where to look or whom to ask.

The process estate. Batch records, in-process parameters, yields, quality results. Structured and queryable, owned by process engineering, analysed with different tools by different people.

Every meaningful inquiry needed both. Has anything comparable been made, and what did the numbers do at scale. Nobody could ask that as one question, so it became two requests to two teams, sequenced, each with its own queue.

Four people were the index

Knowing that a cure profile had been attempted for another customer four years ago, that it failed on storage stability, and that a raw material substitution eventually resolved it, was knowledge held in people rather than systems.

That made the senior formulators the routing layer for the entire technical function, and it meant two upcoming retirements represented a loss no handover document was going to capture.

The confidentiality constraint made naive search impossible

Customer formulations are held under confidentiality. Work performed for one customer cannot inform an answer given to another. That is a contractual obligation rather than a policy preference, and a breach ends relationships and invites litigation.

Any retrieval system across this estate has to enforce partitioning at the point of retrieval and prove afterwards what was retrieved, by whom, and on whose behalf. Filtering results after the fact does not meet the requirement.

The cost of slowness ran both directions

Inquiries that ended in a decline still consumed senior technical hours, sometimes weeks after arrival. Inquiries that could have been won were lost to faster responses. Both traced to the same cause: the time taken to find out what the company already knew.

Solution Approach

Summary: Ground retrieval in the actual records, query batch history in plain language, connect the two, and enforce customer partitioning at the point of retrieval.

Step 1. Fix the boundary before building

The system retrieves and cites prior work. It does not propose novel formulations, suggest untested substitutions, or generate process parameters.

The reason is not conservatism. In this domain a plausible invention is physically dangerous, and a system permitted to generate formulations will eventually produce one that looks reasonable and is not. Outputs are constrained to retrieved records with their provenance, structured query results, and explanations of both. There is no output path that produces a novel formulation, so no prompt can elicit one.

Chemists retain the formulation judgment. The system removes the search.

Step 2. Structure the formulation estate

Records were indexed with the metadata that makes retrieval precise: product family, chemistry class, specification attributes, process route, scale reached, outcome, deviation history, and the customer partition each record belongs to.

Structured indexing was chosen over similarity-based search deliberately. When a chemist asks why a particular record surfaced, the answer needs to be an attribute match they can inspect. A similarity score is not an explanation, and in a confidentiality-bound estate it is not an audit record either.

Step 3. Grounded retrieval with citation

A chemist describes what the customer is asking for. The system returns comparable prior work with the source record attached, the attributes that matched, and a confidence value.

It returns the record, not a paraphrase. The chemist reads the original trial report, including the parts a summary would have dropped, which are frequently the parts that matter.

Step 4. Batch history in plain language

Process and quality history is queryable directly. Yield distribution across a product family, quality results filtered by process condition, deviation frequency by route or site, campaign performance over time.

Every answer returns with the fields, filters and aggregation behind it, so an engineer can see how the figure was constructed before relying on it. Follow-ups resolve in the same session rather than becoming another request.

Step 5. The bridge, which is the actual product

The two estates answer one question in sequence. Find comparable prior work, then show what the process data did for those specific campaigns.

This is what nobody could previously do without occupying two teams for a fortnight. A chemist working an inquiry sees the prior trial, the deviation that arose, the batches that followed, and the yield and quality distribution across them, in one working session.

Step 6. Partitioning and proof

Customer partitioning is enforced when the query is constructed, not applied to results afterwards. A user working an inquiry for one customer cannot retrieve another's records, and the boundary sits in the retrieval path rather than the interface.

Every retrieval is logged: who asked, what was returned, from which partition, on whose behalf. When a customer asks how their formulation data is protected, the answer is a demonstrable access record rather than a policy statement.

Cross-cutting controls

Sensitive commercial data masked before inference. Confidence scoring on every retrieval, with low-confidence results marked rather than presented as equivalent. Role-based access governing retrieval and query. Model tier selected against task complexity. Index and schema context served from cache, since both are needed constantly and change rarely.

Business Impact & Results

Summary: Inquiry turnaround compresses because searching stops being the bottleneck, and institutional knowledge stops depending on four people remaining employed.

Where the compression comes from

In descending order of size: establishing whether precedent exists, which previously required a person who remembered. Connecting formulation history to batch outcomes, which previously required two teams and two queues. Follow-up questions resolving in-session. Assembling quote evidence, which is largely retrieval rather than analysis.

The first is the largest, and it also shortens declines. Establishing quickly that there is no usable precedent releases senior technical hours from an inquiry that was never going to convert.

The value model, with client-supplied inputs

Input

Source

Technical inquiries per month

Commercial pipeline data

Share with usable precedent in existing records

Measurable in pilot

Senior technical hours per inquiry today

Time records or estimate

Loaded cost of senior technical time

Finance

Current inquiry turnaround, and competitor benchmark where known

Commercial

Win rate sensitivity to response time

Client's own historical win/loss data

Contribution margin per won contract

Finance

Two value streams sit on those inputs. Technical hours released, converted to cash at the usual discount rather than at face value. And contracts won that would previously have been lost on responsiveness, which is the larger number and the one only the client can evidence, from their own win/loss history.

We do not supply a win-rate elasticity figure. Anyone who does is inventing it.

What becomes measurable

The share of inquiries with usable precedent, which is a direct measure of how much of the estate is worth the indexing effort. Which product families have thin documentation, which is a records problem the technical function can prioritise. Retrieval confidence by query type, which shows where indexing needs work. Turnaround by inquiry type, separating those resolved from precedent from those needing genuine development. Retrieval access patterns, which serve confidentiality assurance directly.

Continuity

The most durable outcome is the least visible in a first-year business case. Knowledge that lived in four people becomes reachable by the technical function generally. When a senior formulator retires, the records they created stay findable by the attributes that make them relevant, which is not true of a handover document.

The honest limitation

Retrieval quality is bounded by record quality. Where trials were documented thinly, or outcomes were never written down, there is nothing to retrieve and the system will correctly return nothing.

This is worth saying early, because the reflex is to interpret an empty result as a system failure. It is usually an accurate report on the documentation. The system exposes where institutional memory was never written down, which is uncomfortable and useful in that order.

Key Takeaways

1. The bottleneck is establishing what you already know. Most inquiry turnaround is search, not analysis. Compressing the search compresses the response, without changing anything about technical capability.
2. Two knowledge estates, one question. Formulation history and batch data are held by different teams in different systems, and the question that determines the answer spans both. Connecting them is the product.
3. Retrieve and cite, never generate. In a domain where a plausible invention is a safety event, the system must be structurally incapable of producing a novel formulation. Enforce that in architecture, not in instructions.
4. Partition at retrieval, and be able to prove it. Confidentiality across customer work is contractual. Filtering results after retrieval does not meet the obligation. A logged access record does.
5. An empty result is usually a documentation finding. Where nothing comes back, the record was never written. That is information worth acting on rather than a fault to debug.

Conclusion

Contract manufacturers compete on responsiveness and lose on it for reasons that have nothing to do with technical capability. The capability is there, the precedent is usually there, and the delay sits in the gap between a question arriving and the organization establishing what it already knows.

That gap is structural. Formulation knowledge lives in documents and in people. Process knowledge lives in databases and in different people. The question that decides an inquiry needs both, and no one owns the join. So it gets answered by whoever has been there longest, when they have time, which is a system that works right up until it does not.

Any manufacturer whose inquiries route through a handful of long-serving experts will recognize the position. Fast answers when those people are available, slow answers when they are not, and an accumulating risk nobody wants to name because the assumed mitigation is documentation nobody has time to write.

The alternative is not better documentation. It is making the documentation you already have findable by the attributes that matter, connecting it to the process data showing what actually happened, and enforcing the confidentiality boundary in the retrieval path so the whole thing is usable in a business built on other people's intellectual property.

Worth measuring this quarter: take your last twenty technical inquiries and break down the time spent between establishing precedent and doing new work. If the first number dominates, your constraint is retrieval, not capacity and that's fixable a lot faster than hiring.

Looking to solve a similar business challenge? Connect with our experts to explore the right solution for your organization.