Cortex - Chat with your systems

Ask your systems anything, in plain language

Cortex answers questions from live data across every system you connect. It picks the tools, runs them as you, works the numbers exactly, and shows every step. When one question needs several systems, a team of agents works it out together.

01

Ask the way you would ask a colleague

Most questions about the business have an answer sitting in your systems. Getting it usually means a report request, a query, or someone who knows where to look.

Cortex removes that step. You type the question as you would say it, and Cortex finds the answer in the systems behind it: the ERP, the warehouse, the CRM, a database, a ticketing tool. If the answer lives in two systems, it reads both and puts them together.

Sales and financeWhich customers with open deals over $50k have overdue invoices?
WarehouseWhich waves will not finish before the 16:00 trucks?
ITWhich service accounts have not signed in for 90 days?
ServiceWhat is still open for this customer, in every system?

02

How an answer is made

Behind each answer is the same short sequence. The model reads your question and the descriptions of the tools this agent may use, and chooses the ones that fit. The tools run as you, on your connections, so you only ever see what you are allowed to see. Large results are kept as tables and worked with ForgeQL, so totals and comparisons are computed, not estimated.

How it works

03

Two ways to think

Every chat runs in one of two modes, and you can switch between them. Operational Mode is for questions with a direct answer: the model decides what to run, and your data goes straight to your screen without the model ever seeing it. Reasoning Mode is for questions that need thought: the model reads the results, with sensitive values masked, and explains what they mean.

04

When one question needs a team

Some questions need more than one system. What is putting this week's deliveries at risk? Part of the answer is in the warehouse, part is with transport, and what to ask next can depend on what the first system says.

Multi-agent orchestration hands a question like this to a team. A coordinator plans the work and gives each part to a specialist agent that works on one system. Parts that do not depend on each other run at the same time. A part that needs another's result waits for it. And a specialist with nothing to do is skipped: if the warehouse finds no short items, nobody asks purchasing about them.

  1. Each part gets a briefA short task, the facts known so far and the saved results it may use. No data travels in the brief.
  2. Each specialist stays in its laneOne system, its own tools, its own guardrails and approvals. The coordinator has no tools on your systems at all.
  3. Results are saved, not retoldEach specialist saves what it found and hands back a short receipt with a summary and the results it saved.
  4. The answer is computedThe coordinator runs queries on the saved results to build the answer and its charts, instead of stitching the agents' text together. Every figure shows the agent and system it came from.
  5. Every step stays on screenEach decision with its reason and evidence, each agent's tool calls and receipt, and any agent that was skipped, with why.

A question can use up to eight specialist steps, three at a time. If it needs more, the answer shows what was found and what is still missing. When the same question applies to every system and you already know what you want back, an Enterprise query does it in one call instead, with far fewer tokens. Read about Enterprise queries.

05

Keep the answer

A good answer is often one you want again tomorrow. In Operational Mode, pin it to a dashboard and the tile reruns the saved code on its own schedule. Because the code is already written, the refresh needs no AI and uses no tokens.

In Reasoning Mode, turn it into a report instead. Pick a template and colours, approve the plan, and the Report agent builds the PDF.

For work that should happen without anyone asking, the same question can become a deployed agent on a schedule. Read about deployed agents.

06

What sits behind a chat

You always talk to an agent, and the agent decides what is possible. Your team sets each one up with the connections and tools it may use, the skills it follows, and the guardrails that apply. The model is chosen per chat. The agent also has a memory: it keeps what it learns about you, such as your name, your email and how you like answers written, so you do not have to repeat them. People only see the agents they are allowed to use.

Order analystAn agent your team set up for operations questions

Model

Chosen per chatAny provider, your cloud or your own servers

Connections

SalesforceJD EdwardsBlue YonderManhattan AssociatesDev, Test or Prod, chosen when you ask

Tools

24 allowedOnly these can be called

Skills

Late order triageWeekly summaryKnow-how the agent follows

Guardrails

Read-onlyMask emailsSet per agent

Memory

Name: AlexEmail: alex@company.comShort answers, detail on requestRemembers your preferences from one chat to the next
Figure 1. What sits behind a chat: the agent you talk to, and what it is allowed to use.

07

What to know

  1. It answers from the tools it hasIf no tool reaches the data, Cortex says so instead of guessing. A new tool closes the gap.
  2. It sees what you can seeEvery call runs as you. Cortex cannot reach further than your own access.
  3. The steps are always shownEach tool call, its arguments and its result stay on screen above the answer.
  4. The model is your choiceHosted, in your own cloud or on your own servers, chosen per chat.