Cases
From manufacturing to finance and the public sector,actually used DARVIS
Finding the cause usually takes a few meetings and a few weeks of digging.
In the cost cases, a single analysis ran in 81 to 453 seconds. Open the full case to follow how each one went.
- 118Analyses closed, of 161 runs · 0 failed
- 4~6weeksTime to first rollout
- 5stagesSignal → evidence
Selected cases
What the companies that rolled it out got
Cost · every order recalculated
Three hours to find the cause —by then the decision was already made
Nobody could look at all 899 items, so only the top few were checked; now every one is scanned and the missed cases surface each month.
Where the money was leaking, unseen, in one business unit
One business unit of a large food manufacturer · 2026| Where money leaks | Before | Now | Preventable per year (scenario) |
|---|---|---|---|
| Money lost to finding out late | Known only after the loss was locked in | Flagged daily, before it locks in | 13M won |
| Money lost to not finding it | People checked only the top 5 items | All 899 items scanned · about 17 more found each month | 60M won |
| Money spent on the search itself | 3 hours per cause · 30 minutes to check inventory | 5 minutes | 38.5M won |
| Total | One business unit · about 9.3M won a month | about 110M won |
Hours and counts are measured; the amounts multiply them by conservative assumptions (labour cost, loss per case) and are scenarios. This is preventable money, not money recovered. The company is anonymised.
These numbers change before the next quote, before the next order.
See how a case like this gets reasoned through →
The totals were profitable; order by order, some were sold at a loss
Only after rebuilding the orders one by one did the loss-making stretch surface — and the same held on direct cost alone, before any overhead allocation. Counts, shares and amounts are the customer’s data and are not disclosed.
See the cost-loss screenPlant capacity and centre days-of-stock in one view, with people approving each transfer
Across N distribution centres and the biscuit and ice-cream categories, production allocation, inter-plant transfers and near-expiry ageing sit on one screen. It now covers an allocation and transfer recommendation screen, a settings page (per-centre plant priority and capacity) and an STO integration — under contract, in field testing, targeted to go live in October 2026.
See the inventory-loss screen- Electronic components
Three AI agents taken through a full PoC
An internal chatbot, DB sync automation and equipment-name validation — three problems solved the same way.
Read the full Company I case - System semiconductor
Validated in two weeks against 109 real questions
Manuals that never stop changing in a 24-hour plant, tested with the words the floor actually uses.
Read the full Semiconductor manufacturer case One question across the documents of a $700M-revenue maker
Documents scattered across internal systems, verified in three stages and pulled into one answer.
Read the full Global parts manufacturer caseFrom instrument data to an audit-ready report
We tested whether qualification reports could be produced without passing through human hands.
Read the full Pharmaceutical manufacturer case- Automotive parts
Material planning verified in real time
Scattered systems joined behind one chatbot, with the data context unified.
Read the full Company W case From interpreting a forecast model to company-wide data
It started with making the model’s reasoning readable to people.
Read the full Shipping company case- Financial data
150 SQL users find metrics in plain language
A 1.5-month PoC confirmed Text-to-SQL, plain-language profile search, report automation and access control. Quantified gains are still being measured.
Read the full Company K case - Public administration
Shorter response times in administration
Repetitive administrative work absorbed by an AI chatbot, shortening response times.
Read the full Public agency case
The common procedure
Every case goes through the same five stages
Signal detection
Items that drift outside their usual range are flagged first, so we know where to look.
Read-only connection
ERP, MES and cost sheets connected read-only. No changes to your systems and no redeployment.
Recalculation per order
Cost rebuilt per order and per item — starting from individual records, not from totals.
Breaking down the contribution
Against last month, which factor pushed how much — broken out in money.
Evidence assembled
Every answer is delivered with the table and column it came from.
Observed in our own execution log as of August 2026 — a single analysis runs in 81 to 453 seconds, and a question yields 2.38 candidate causes on average. These are observations of runtime and count; they do not mean money saved or recovered.
How this differs from other tools
Different from a BI tool
BI shows you the metric. Here, when the metric drifts outside its range we flag it first, then leave the contribution and evidence together.
Different from a systems integrator
An SI fixes your systems. Here nothing is touched — we only read, and analyse.
Different from doing it in Excel
What someone recalculated by hand every close now runs every month on the same basis.
How to read these numbers — the figures in each case are observations from recalculating that site’s own data. They are not money we recovered, so please don’t read them as savings or ROI. Performance figures are published once attribution has been checked.