01
One order lives in four systems. No screen shows all of it
An order is born in the ERP, picked in one warehouse, topped up from another and handed to a carrier. Each system knows its own part of the story, and only that part.
So a simple question, where is this order right now, has no simple place to be asked. Someone logs into the ERP, then the warehouse system, then the other warehouse, then transport, and pieces the answer together by hand. By the time it is written down, part of it is out of date.
There are three usual ways out, and each has a cost. You can build integrations between the systems, which takes months and has to be redone when one of them changes. You can copy everything into a data warehouse, which gives you yesterday's answer. Or you can put an AI agent on each system and have them work it out together, which works, but spends a great deal of AI on what is really a look-up.
Enterprise queries were made for the case those three miss. You already know what you want from each system. You just want it from all of them, in one go, in one table.
02
Decide once what to ask. Then ask it everywhere
The idea rests on a small piece of preparation. For each system, your team writes a tool that answers the question in that system's own terms: a query for the ERP, a command for the warehouse, an API call for transport. The only rule is that every version returns the same columns.
On top of those sits the Enterprise tool. It has no query of its own. It is a short definition that says which systems to ask, how to combine what comes back, and what the columns of the result are.
When someone asks where an order is, the model sees one tool and picks it. Smart Forge sends the question to every system the tool covers, in a single request, and combines the answers with a query. Each row of the result says which system and which instance it came from, so nothing is lost in the merge.
Step through it below.
03
Why it uses so few tokens
Every time a model reads, plans or writes, it uses tokens, and tokens are what AI costs. The way to spend fewer is not a cheaper model. It is to ask the model to do less.
Compare two ways of answering the order question. With a team of agents, a lead agent reads the question and plans. An agent for each system then reads its brief, chooses its tools, reads the results and writes a hand-off. The lead agent reads the hand-offs and writes the answer. The model is working at every one of those steps, for every system, each time the question is asked.
With an Enterprise query the model makes one decision: which tool, and which order number. Everything after that is code. The systems are asked by code, the answers are merged by code, and the table is drawn by code.
Where the model does the work
- Done by the model, uses tokens
- Done by code, uses none
When you already know what to ask, the model only needs to decide once.
The saving grows with the number of systems. A fifth system adds a row to the table. It does not add another agent that has to think.
In Operational Mode the saving is complete. The combined table goes to your screen and not to the model, so the rows themselves cost no tokens at all. Read about Operational Mode.
The usual way, and the Smart Forge way
04
Where it fits
An Enterprise query suits any question where you know what you want from each system before you ask. These are examples of the kind of question it is built for.
The ERP, each warehouse and transport report what they know about one order, in four rows.
Where is order SO-48290 right now?Every warehouse reports what it holds of one item, so you see the whole position before you promise a date.
How many units of A-204 do we have, across all sites?When someone leaves, each system reports whether it still has an account for them.
Which systems still have an account for this leaver?Each ticketing tool reports its open items, in one list you can sort your own way.
What is still open for the warehouse team?The same tool serves more than chat. A deployed agent can call it, and a ForgeView screen can be built on it, so the combined table becomes something people open every morning and not something they have to ask for.
And if you want to work on the result, total it, group it or find what stands out, Reasoning Mode treats the combined table like any other result and can query it with ForgeQL. One request gathers the data from every system. The thinking then happens on a single table. Read about Reasoning Mode.
05
Enterprise query, or a team of agents?
The two are easy to confuse, because both span several systems. They answer different kinds of question, and you can use both.
Use an Enterprise query when the same question applies to every system and you know what you want back. Use multi-agent orchestration when each system needs its own investigation, or when what you ask the second system depends on what the first one said.
| Compared on | Enterprise queryOne call, every system | Multi-agent orchestrationOne agent per system |
|---|---|---|
| The question | The same for every system | Different for each system |
| What the model does | Picks one tool | Plans, investigates and writes up |
| What comes back | One table, each row labelled by its source | A written answer with its evidence |
| AI cost | Tokens only for choosing the tool | Tokens for every agent and every hand-off |
| Best for | Status, positions and lists across systems | Working out why something happened |
06
What to know before you rely on it
Four things are worth knowing, so the table is never read for more than it says.
- Each system needs its own version of the toolYour team writes it once for each system. A system without one is not asked.
- It puts rows side by sideThe result stacks what each system returned and labels the source. It does not decide which rows belong together. That reading is yours, or Reasoning Mode's.
- It answers when the slowest system hasThe request waits for every system it covers, so the answer arrives with the slowest reply.
- A system you are not connected to is left outThe source column shows which systems answered, so check it before you conclude that a system had nothing to say.