Inward / production systems

Red Bumerang/Verticals/Inward

Industrial
process AI.

A process nobody has written down cannot be automated, only guessed at. We map how the work actually runs, put AI where it earns its place, and connect the systems that were never meant to connect.

Book a 30-minute meeting See what a build looks like
~2 minTo draft a customer proposal, down from as much as a month
~30,000Industrial plants scored for suitability, with zero site visits
78 → 1Undocumented sheets rebuilt as one production application
16,005Calls reviewed overnight, with full coverage rather than a sample
Order first

Nobody has written this process down. That is the project.

Ask three people how a process runs and you get a diagram from 2019, a colleague who simply knows, and a workbook with seventy-eight tabs: three conflicting accounts, none of which reflects how the work actually runs. This is why industrial AI fails: the model is pointed at a description rather than the operation. So we begin by reading the work itself: what moves, who really decides and which system nobody can export from. That map is a deliverable in its own right, before any model is chosen. Only then does automation have a process it can reliably follow.

We do not automate the process you describe. We automate the one that runs.

Begins withReading the work

Never begins withThe diagram on the wall

The shape of a build

One feed. Two processes. Two outputs.

Every system we put into production has the same skeleton. Your data enters once — from the ERP, or from whatever has been standing in for one — and drives the processes that used to be done by hand. Both write into a single source of truth, so the record and the reasoning behind it live in one place and can be audited afterwards. Two things come out: an internal application your team works in daily, and documents that are generated rather than typed. And every run feeds the next one.

Process schematic

  • Source data
  • Automated step
  • Step with AI
  • Source of truth
PROCESS 1 PROCESS 2 ERP data one feed in no re-entry Step 1 automated AI Step 2 model makes the judgment Step 3 automated Step 1 automated AI Step 2 model makes the judgment Step 3 automated Single source of truth records + context AUDITABLE RESULT 1 Internal web app your team works in it daily RESULT 2 Automated documents generated, not typed FEEDBACK · EVERY RUN INFORMS THE NEXT ONE
1 feed2 processes2 AI steps2 outputs Schematic · step labels stand in for your own

Why one feed

Re-entry is where the truth splits. The moment the same figure is typed twice, two versions of it exist and somebody spends their week reconciling them.

Why one source of truth

An answer nobody can trace is not an answer. Records and the context behind them sit together, so a decision can be explained months later.

Why two outputs

People need a place to work; the business needs paper. One application your team uses every day, and the documents that used to be typed one at a time.

Where AI earns its place

The question is never whether to use a model. It is which step deserves one.

Most of a working system is ordinary code. Put a model where a rule belongs and the project becomes unpredictable, expensive and difficult to audit. Put a rule where judgment belongs and the automation may be switched off by month three. Four tests guide the decision.

Test 01

Repeated judgment → the model

Reading a document, scoring a plant for suitability, deciding whether a call went the way it should have. It is the kind of work someone repeats a thousand times in much the same way but has never been able to document.

Test 02

Exact rules → code

A discount table, an approval threshold, a tax rule. If it can be written down precisely, a model adds only another way for it to go wrong.

Test 03

No export button → reverse engineering

Two of the sources in our production systems had no API at all. In industry that is the normal case, not the exception, and it is usually the reason the process was still manual.

Test 04

Policy says no cloud → your environment

Frontier models where they are permitted, a private deployment inside your own infrastructure where they are not. The decision is made per system, not once per company.

In production

Two systems, described without naming anyone.

Client work is covered by non-disclosure, so the companies remain anonymous and the numbers provide the evidence. Both systems are running today; neither was a pilot.

Case 01 · Order handling

Seventy-eight sheets, no owner, no specification

The workbook ran the order process. Nobody had written down how. Every rule lived inside a formula, every exception in somebody’s memory, and the file could no longer be edited safely by more than one person at a time. We read it end to end, recovered the rules rather than guessing them, and rebuilt it as a production application with roles, permissions and an audit trail in days rather than quarters.

Spreadsheet to system

  • Legacy workbook
  • Rebuilt in production
ORDER INTAKE UNDOCUMENTED WORKBOOK PRODUCTION APPLICATION 78 sheets · no owner · no specification Delivered in days · roles, rules, audit trail
78 sheets readRules recovered, not guessedProduction in days

What was read

The whole file, not a sample. Seventy-eight sheets, every formula and every cross-reference, including the tabs nobody had opened in two years.

What was recovered

The rules, stated explicitly. This included several the business did not know it was applying, and two it no longer agreed with.

What was delivered

An application, not a spreadsheet with a nicer front end. Roles, permissions, an audit trail, and more than one person able to work at once.

Case 02 · Call quality

Every call reviewed, not a sample

Quality assurance in a contact centre normally means a supervisor listening to a handful of calls a week and extrapolating from them. We processed 16,005 calls and 434 hours of recordings overnight without supervision, flagging them by category: where the script broke and where no next step was agreed. Coverage increased from a sample to every call, leaving supervisors more time to coach the people identified by the flags.

Call quality · 100% coverage

  • Reviewed audio
  • Flagged segment
SCRIPT MISS NO NEXT STEP AGREED CALL QUALITY · 100% COVERAGE 16,005 CALLS · 434 H OF AUDIO REVIEWED OVERNIGHT
100% of calls, not a sampleFlags by categoryNo supervisor review time

What changed in coverage

From a handful a week to all of them. A sample can show overall patterns; it cannot identify which call lost an account.

What changed in speed

434 hours of audio, reviewed overnight. The flags are on the supervisor’s desk before the shift starts, not three weeks later.

What changed for the team

The supervisor spent less time listening and more time coaching. The scarce resource was never the audio. It was the attention.

Three more, in short

A national telecom operator

Up to 1 month→

~2 minutes

Drafting a customer proposal. The whole sales team now uses it daily.

An EU energy optimisation group

A site visit each→

~30,000 scored

Every industrial plant in a country scored for suitability from desk data, without a site visit.

Red Bumerang · our own operation

1 month→

4–8 hours

Campaign preparation. Client onboarding went from four a month to two a day. We ran it on ourselves first.

Where it runs

Nothing gets replaced.

The system attaches to what you already run: the ERP, the CRM, the spreadsheet nobody documented, the national data service with no export button. Deployment is decided system by system, according to your policy.

Models
Frontier models where they are permitted, or a private deployment inside your own environment where they are not
Inference
Run on our infrastructure or yours; where your policy requires it, nothing is routed through a third-party AI cloud
Languages
Speech and text systems work directly in Slovenian, Croatian, Latvian, Lithuanian, Estonian and Hungarian; they are trained for each language rather than translated
Sources with no API
Reverse engineered — two of the sources in our production systems had none at all
Data residency
Stays wherever your policy requires it to stay — including entirely inside your own network
Handover
Your team uses the application daily; we continue to support and improve it after go-live

Processes we have rebuilt run in

Manufacturing·Telecom·Energy·Construction·Contact centres·Food & beverage·Logistics·Applied research

Next step

Bring us one process.

Thirty minutes, no deck. Bring the one that eats a week every month, or the one only a single person knows how to run. We map where the order is missing, which step, if any, deserves a model and what the first version would cover. You leave with a clear outline and a first step, even if we do not work together.

Book a 30-minute meeting or write to info@redbumerang.com