Structuring Intelligence

Léa leaves Thursday's committee with another question: which renovation plans across the portfolio are actually ready to proceed?

By Monday, she has gathered the available reports, budgets and meeting minutes. The information is there, but comparing it means reading hundreds of documents and bringing their different descriptions into a consistent form. Until recently, the preparation alone would have been enough to leave the question aside.

The cost of preparing information helps decide which questions get asked. Léa now has a team of human and AI agents that divides assignments and checks its work, and continues while she is elsewhere. She gives that organisation a further responsibility: building the information base it needs to answer.

The question left aside

Léa's database already contains addresses, floor areas and energy figures. Someone chose those fields, collected the information and arranged for it to be updated. The approval of a renovation budget may only appear in meeting minutes; if it was never collected, she has to open the documents again.

Making that detail available across a portfolio has traditionally meant bringing analysts, data specialists and engineers together. They locate sources, agree on definitions, write extraction code, connect records to buildings and check the results. Then they maintain the whole arrangement as documents and requirements change.

There are two costs here: turning documents into usable information, and building a reliable way to keep doing it. Agents take on parts of both, reading the reports and writing the code that collects, checks and updates the results. Together, these steps form what is often called a data pipeline.

For Léa, that brings the renovation question within reach of this week's review. She gives her team an assignment:

“Compare the renovation plans across the portfolio. Separate intentions, approved projects and completed works; show the evidence, and flag anything that needs clarification.”

Words become something we can compare

The documents describe progress in different ways. “The owner is considering replacing the heating system” and “a replacement is under discussion” both describe an intention. “The replacement was approved at the March meeting” records a different stage.

The agent reading each passage classifies the project using the agreed stages: considered, approved or completed. It extracts dates and budget amounts, attaching each entry to the relevant building and source. Different wording no longer requires someone to rewrite every description by hand.

Once those descriptions have a consistent place and meaning, Léa can count how many projects have approval, how many have a documented budget and how many lack a start date. Words written about individual buildings become a comparison across the portfolio.

This is what structuring information achieves. The underlying documents remain, with their detail and nuance, while selected facts become easier to find, combine and use.

The team builds the process too

The coordinator already divides the portfolio review between collection, analysis and verification. For this assignment, it also assigns responsibility for building the extraction process. The agent doing that work uses the analyst's definitions, while the reviewer tests the results against the original documents.

Those responsibilities depend on one another. Before the comparison runs across the portfolio, the team needs to agree on what counts as an approved project and check that the extraction preserves that distinction.

One agent builds and tests the process. Code handles repeated steps (collecting files, linking them to buildings and saving results), while an agent reads the passages that need interpretation. The reviewer checks a first set of results against the documents before the work expands across the portfolio.

On Monday evening, Léa leaves while the team continues processing the reports. During the night, the reviewer finds a budget marked “approved” where the source says “submitted for approval”.

The reviewer sends the entry back for correction. The agent responsible for the extraction tightens its instructions and adds a test for that wording; the reviewer then checks the other approval entries for the same mistake. The correction improves both the current table and the process used to produce it.

Another building has no budget mentioned in the available files. Its entry stays “not documented”, because absence from the documents does not establish that no budget exists.

By Tuesday morning, the overnight correction is recorded in the shared work history. Léa's attention is needed elsewhere: two documents disagree about a project's status, and the coordinator brings her the passages and their dates. She asks the property manager for confirmation; that entry remains unresolved while the remaining comparisons continue.

These checks belong inside the process. They give the team a way to catch errors, revisit affected results and bring genuine uncertainty back to the person responsible.

A process that survives the next document

When a new report presents its renovation schedule differently, the agent responsible adapts the process. The reviewer checks it against both the new report and earlier examples, so a change that solves one problem does not quietly introduce another.

The savings build on both sides: less manual reading for each comparison, and less technical work to create and maintain it. The team reuses what already works, concentrating on new material, exceptions and checks.

The coordinator keeps track of those changes alongside the ongoing analysis. Léa sees what is complete, what is being checked and where a decision is needed, without having to supervise each document.

Thursday's next question

By Thursday, Léa has a comparison of renovation plans, with evidence and unresolved entries clearly marked. She also has something to reuse: organised facts, a working process and the definitions and corrections accumulated along the way.

At the committee, someone asks which approved works scheduled for next year could affect rental income. Her team starts with the plans and dates already checked, then brings in the lease information needed for that next comparison.

Her organisation carried the assignment while she was elsewhere; it now retains a foundation for the next one. Checked information, reusable processes and corrections accumulate together, so less of each investigation begins from zero.

The question Léa brought back from the previous committee has become a comparison she can use, and the work behind it has made the next question easier to pursue.