Cortex · Operational Mode

A mode where the AI never sees your data

Smart Forge works in two modes, and you choose which one to use. In Reasoning Mode the model reads your results. In Operational Mode it does not: it works from names and descriptions only, and your records stay in your systems.

01

The data that needs AI most is often the data you cannot send to it

Most AI assistants answer a question by reading. They fetch the rows, send them to a model, and the model reads them and writes a reply.

For a lot of business data that is fine. For some of it, it is not. Patient records, product formulas and designs, payment card details, the personal data of your customers: this is data that policy, contract or plain caution says should not be handed to an outside model. The teams that hold it are often told the same thing, which is that AI is not for them.

Operational Mode exists for those teams. It turns the usual arrangement around. The model never reads your data. It only decides what should be done, and Smart Forge does it.

The same question, two ways
Which patients missed a follow-up this month?
A typical AI assistantSends the patient records to the model.the data leaves your systems
Operational ModeThe model sees column names. You see the list.
Figure 1. The same question, answered two ways.

02

A model can pick the right tool without seeing a single row

Think of a dispatcher in a large office. She knows which department handles which request, and what each one needs to be told. She does not open the files. When a request comes in she routes it, and the department does the work and answers the person who asked.

That is the model's role here. It is given a catalogue of your tools: the name of each one, a sentence about what it is for, and the names and types of its columns. From your question it chooses a tool and fills in what to pass. Then it steps aside. Smart Forge makes the call on your system, and the result comes back to Smart Forge and onto your screen. The model is not shown it.

Step through it below.

How it works

Follow-up questionsThe same rule holds when you ask for more

Ask for a chart of those items by supplier, and the model still does not get the rows. It writes a few lines of code from the column names alone: group by supplier, count, draw a chart. Smart Forge runs that code on the data it already holds. The table and the chart you see were built from your data, inside Smart Forge.

The model writes the steps. It never sees the numbers.

Because the code is written before it runs, it can be read. Cortex can explain in plain language what the generated code is going to do.

03

What stays, and what travels

It is worth being exact here, because "the AI never sees your data" is a claim that deserves checking.

Three things travel to the model: the question as you typed it, the names and descriptions of your tools, and the names and types of their columns. Everything else stays where it is: every row, every value in it, and the credentials used to fetch it.

One consequence follows. If you type a patient's name into a question, that name is part of the question, and the question is sent. Ask about the data, not with it.

Pick a kind of data below to see what stays in your systems and what the model receives.

This is not a compliance certification. The widget shows the kind of data each rule set covers, and how Operational Mode treats it. Your compliance team decides what applies to your data.

04

Scheduled agents: designed once by AI, run without it

The same idea goes further for work that repeats. Describe a job in a sentence: every hour, raise a purchase request for any item below its reorder level. The model designs the agent from the names of your tools and columns, in the same way it answers a question. The columns it relies on are checked before the agent is saved.

From then on the model is out of the picture. Each run is ordinary code. It reads, matches and acts with no model call, so no rows are sent anywhere, there is no AI cost per run, and the same inputs give the same result every time.

Every hourrun at 09:00
  1. low_stock()214 items read
  2. match supplier214 matched
  3. raise request12 raised
model calls in this run: 0
Figure 2. One run of a scheduled agent. The model designed these steps once. It takes no part in running them.

Set against the usual way of running an AI agent, the difference is in every row.

The usual way, and the Smart Forge way

The usual wayWith Smart Forge
Every run sends rows to a modelScheduled runs call no model at all
Each run costs AI tokensRuns cost no AI tokens
A wrong column fails halfway through a runColumns are checked before the agent is saved or run
Each follow-up question sends the data to the model againFurther analysis runs as generated code on the data already retrieved

05

What it costs you

Operational Mode has a trade-off, and it is better to hear it here than to find it later.

A model that cannot read your results cannot think about them. It cannot notice that most of the late orders came from one site, or explain why a number moved, or follow a hunch into a second system. Those are jobs for Reasoning Mode. Operational Mode is built for operational work: looking things up, listing, filtering, charting and carrying out routine tasks.

So the choice depends on the question as much as on the data. If the question starts with what, which or how many, Operational Mode will answer it and nothing will leave your systems. If it starts with why, you need a model that reads.

There is a way to have both. Run the model inside your own network, and Reasoning Mode can read your data without that data leaving your walls. Bring your own model.

06

The two modes, side by side

You choose the mode for each chat and for each agent, and you can use both: Operational Mode for look-ups and for data that must stay where it is, Reasoning Mode for the questions that need thinking.

Compared onOperational ModeThe model never sees your dataReasoning ModeThe model reads the results
What the model seesNames and descriptions only. No records.Your question and the results the tools return
Best forLook-ups, lists, charts and routine tasksWhy, what changed, what should we do
How it answersPicks the tool. Smart Forge runs it and shows you the rows.Plans, investigates and explains
Depth of analysisLimited: it works without reading the dataFull: it can follow a pattern across several calls
Sensitive dataNo data reaches the model at allGuardrails can hide chosen values from the model
AI costTokens only for choosing the toolTokens for every result the model reads