Context for agents
An agent knows only what it is handed.
A group’s answers used to pass through a person who knew what no tool recorded. More of that work now goes to AI agents such as Claude, and an agent knows only what it is handed.
This paper sets out what that person carried and what a good group answer needs. It ends with how a group can prepare its definitions before it hands that work to agents.
What a person carried
A group’s knowledge sits in the tools each company runs: a ledger, a CRM, an online store, a shared drive. Each tool holds its own records, in its own words.
The knowledge that made those records usable sat somewhere else. A person knew which spreadsheet was current and which figure the accountant had restated. That knowledge was never in a tool.
The same person knew which company counts freight in cost of goods and which in overhead. They knew whose month closes late, so an empty cell meant late, and that two companies sell to the same customer.
Group answers were good because that person corrected them on the way. Each correction lived in a habit or a note on a spreadsheet, where no system could read it.
What an agent is handed
An agent reads what its connection returns, the tables and files of the systems it can reach. It cannot ask the controller which file is current.
So an agent answers from the records as they stand. Freight sits wherever each company put it, a late month reads as an empty one, and one customer under two names counts twice.
The answer arrives fluent and complete in form, whatever it missed. An agent repeats a wrong definition in every answer, for everyone who asks.
An agent also knows nothing of who is asking. Handed every company, it answers anyone about every company.
The risk grows with trust. The more of the reading a group hands to agents, the further a wrong definition travels before anyone checks it.
Groups and scale
A single company has one ledger and one vocabulary. A group has many companies, each with its own tools and its own words for an order or a channel.
The work of reading a group grows with every company and every tool. Connecting an agent to one system of one company keeps getting cheaper.
Two things stay scarce: one definition that holds across companies, and a boundary on who sees each one. Neither lives in any single company’s tools, because both belong to the group.
Groups also hold their companies in different ways. Some own every company outright, and some hold part of each, sometimes several levels deep. A group figure therefore needs a stated scope and a stated share.
A good group answer
A good group answer can be described with no product in it. Four needs decide whether people and agents can rely on it.
- Each company keeps the tools it runs, and the group reads them where they are
- Each figure means the same thing at every company, written down with its owner
- Each person and each agent sees only the companies granted to them
- People and agents read the same definitions, and every figure traces to its sources
An answer that meets all four can be checked. It names the dataset it read, the defined table behind its figures. It says when that data was last refreshed, and a company with no data shows as missing.
Kinds of tool
A group can turn to several kinds of tool, and each was built well for its own job. Each answers part of a group question, and none holds a group’s answer in one place.
| Category | Built for | What each group question costs |
|---|---|---|
| A controller and spreadsheets | Any question, answered by hand | A round of emails and exports for each question, and knowledge held by one person |
| Portfolio monitoring | The figures each company chooses to report | A request to every company each period, and a new template for each new question |
| Consolidation | The group’s books, closed from each company’s ledger | A finished close before each answer |
| One ERP for the group | Every company on one system | A migration for each company, and for each new one |
| Business intelligence and a data team | Data a team has modeled first | A model per company, built and kept by a data team |
| AI agents linked to each tool | One system of one company at a time | A connection per tool and per company, each read without the group’s definitions |
Each kind holds some of the four needs. A group answer depends on holding all of them together.
Preparing a group
A group can prepare before it chooses any tool. Most of the work is writing down what a person already knows.
- Pick the questions the group chases most, and the figures each one needs
- Define each figure once, with its owner and how each ledger maps to it
- Name each shared customer, supplier and product once, and each figure’s system of record
- Describe each dataset for outsiders, with what it means and notes on its columns
- Before any agent connects, decide which people and agents may see each company
- Agree that a missing value shows as missing in every total
- Record what each answer read and when, and keep each board pack as sent
This work makes the group’s own reporting more reliable first. It is also what an agent needs before a group can trust its answers. Our companion paper, Defining a figure once, works through the second step in detail.
Gatlas
We built Gatlas for this layer, starting with our own companies, whose month ends gave us the first datasets and checks.
Gatlas reads each company’s tools and files, defines each figure once for the whole group, and hands it all to your people and your agents. Each reader sees only the companies you grant.
Every figure traces back to its sources. A missing value shows as missing. Each dataset says when its data was last refreshed.
Thank you for reading.