Guidance on public programmes such as Manufacturing AI Innovation and the AI Voucher — eligibility confirmed in a conversation

Details
Skip to content
DFINITE
1-minute fit check
DARVIS productDARVIO consulting
KO 한국어EN English

DARVIS · agentic workflow · week 3

Every question that comes up at month-endruns through a step-by-step analysis flow.

Ask in one line and it passes through five stages to an answer that carries its evidence.
That flow is designed by our consultants, and the AI assistant works through it in order.

How it works

From one manual to an answered question

  1. 01STEP

    Designing the process manual

    A step-by-step flow a consultant assembled from interviews with your veterans. The veteran’s instinct is not erased; we ask when it forms and what would make it wrong, then write that down as rules.
    Written in business language, so it can be reviewed without any IT knowledge.

    DARVIS Agentic WorkflowFictional Manufacturing Co. · process manual · v1.2
    Read-OnlyExample — fictional data
    Process manual v1.2Business language · 5 stagesReviewed by production · 2026-06-14
    01Anomaly detectionDETECTAutoML

    First, whether a close metric has left its usual range.

    Data it reads
    • MES.daily_yield
    • ERP.cost_actual
    Pass conditions — fail one and it doesn’t move on
    • At least 30 days of observation
    • Outside the usual range (2σ)

    What it emitsList of flagged metrics→ into QUERY input

    • Written in business language
    • Reviewable on the floor
    • Kept as an asset after review
  2. 02STEP

    Testing a user question

    A plain-language question comes in, the data map (the ontology) is read as context,
    and the right one of four tools is called.

    DARVIS Agentic WorkflowFictional Manufacturing Co. · question test · staging
    Read-OnlyExample — fictional data
  3. 03STEP

    Answer rendered, with citations

    Tables, charts and a summary — and every answer cites its evidence automatically.
    If the evidence is thin, it doesn’t answer.

    DARVIS Agentic WorkflowFictional Manufacturing Co. · answer with citations
    Read-OnlyExample — fictional data

How does it run?

Built in five stages, run in two modes

01

Building it

Five workflow stages

Framing the problem

First we agree which metric moving out of range counts as a real problem.

with Consultant
Definition of a process anomaly

Writing the process manual

You receive the step-by-step manual and simply review it. No IT knowledge required.

with Consultant
Process flow diagram

Building the analysis line

The manual you reviewed becomes the analysis steps as written. Nothing for you to build.

with AI engineer
Analysis line assembled

Bringing in the AI assistant

An AI that reads your data map starts taking questions in plain language.

with AI engineer
A veteran operator

Fitting the analysis tools

Four tools are attached — query, search, forecast, connect. Which one to use is the AI’s choice.

with Built into DARVIS
Instruments and analysis equipment
02

Running it

Two modes
A

Standard workflow

Most routine questions

Questions that repeat every month follow the order written in the manual.

Example questions“What was line A’s yield yesterday?” / “What was last week’s average hourly rate?”

B

Autonomous workflow

The harder cases

For unknown problems with no manual, it explores the data map on its own — and a person reviews what comes back.

Example questions“Why is the margin wobbling this month?” / “I can’t tell what’s causing it”

The two questions that come first

How is it different from general AI, and does it overlap with what we already have?

Q.

How is this different from a general AI like ChatGPT?

A general AI doesn’t know your company. DARVIS works on your data map, and every answer arrives with its evidence.

DARVISGeneral AI
Your company’s contextTakes your data map — equipment, materials, process relationships — as contextDoesn’t have it
Citing evidenceEvery answer shows the table, row count and datesAbsent or unreliable
Tool useCalls only the tools it is given (Text2SQL · RAG · AutoML · MCP)The LLM answers directly
Data leavingNothing sent outside. Runs on your own serversSent to an external server
Q.

We already have ERP, Databricks and Snowflake — do we need this too?

It sits at a different layer. DARVIS is a layer on top, so it runs over whatever you have underneath. Leave your ERP and MES exactly as they are.

The people doing the work

They ask in plain language and get an answer

  • Department head
  • Executive
  • Non-specialist

Data infrastructure

The foundation that stores and processes data at scale

  • Databricks
  • Snowflake
  • Self-hosted

Operational systems

Where your operating data comes from

  • ERP
  • MES
  • QMS
  • File server

Whatever you choose for infrastructure, this goes on top of it.

Frequently asked questions

What holds people back from putting AI on the floor

The ontology this screen reads

A question like “which products are heading for a stock-out, and what do we do” doesn’t end at one table. The answer comes from following relationships — product to inventory, inventory to orders and suppliers. That chain of relationships is what the workflow takes as context.

Entities in play — Product · Inventory · Supplier · Order

What an ontology is →
ProductEntitySKU, Name, CategoryInventoryEntityOn-hand, AvailableSupplierEntityName, LeadTimeOrderEntityOrderID, Date, QtyFacilityEntityLocation, Capacity

Choose a stage

Back to the DARVIS overview

Next step

Pick where to start

Month-end cost explained

  • Read-only connection to your systems
  • One plant · one product family · one close
  • Comparison · contribution · evidence
  • Starts in 4–6 weeks

AX kickoff guide

  • The three most common false starts
  • Four conditions for picking the first use case
  • A 90-day roadmap
  • A reporting frame for executives, plus a 10-point self-check

Technical review pack

  • Read-only integration and data flow
  • On-premise and air-gapped options
  • TLS 1.3 · AES-256 · SSO/RBAC
  • An engineer review call if you want one