Companies now publish long sustainability reports, and investors, lenders, and analysts want the numbers inside them. Reading hundreds of reports by hand is slow, so teams turn to AI to pull out emissions, energy use, and targets automatically.
The difficulty is that a sustainability number is rarely just a number. "12,400" could be tonnes, thousands of tonnes, or megawatt-hours. It could cover one plant or the whole group, and one year or three. Emissions report data extraction depends on labels that capture all of that context. This post explains what those labels need and why subject knowledge matters.
A useful extraction record holds more than a value. Annotators typically label:
Leave out any one of these and the number becomes hard to compare. Good sustainability report annotation treats the value as one field among several, not the whole answer.
Imagine two reports that both state the same emissions figure. One covers the whole group, and the other covers a single division. One is for the calendar year, and the other for a fiscal year ending in March. Stored as plain numbers, they look identical and compare as equals.
An extraction model learns from its labels. If annotators capture only the value, the model will return values and ignore everything around them. Analysts then get a table that looks tidy and cannot be trusted.
Units are a common source of quiet errors. Reports mix tonnes, kilograms, and thousands of tonnes. Energy appears in kilowatt-hours, megawatt-hours, gigajoules, or fuel volumes. Some tables state the scale once in the heading, such as "in thousands," and never repeat it in the cells.
Annotators need rules for these cases. Should the label record the figure as printed, or convert it? A written approach, applied the same way every time, prevents a dataset in which the same unit is read three different ways.
Emissions are usually grouped into three scopes. Scope 1 covers direct emissions from a company's own operations. Scope 2 covers emissions from purchased energy. Scope 3 covers the wider value chain, including suppliers and the use of products.
Scope 2 adds a wrinkle, because reports often show it in two ways, one based on location and one based on market purchases. Both may appear in the same table, and they are not interchangeable.
Boundaries matter just as much. A report may cover facilities the company controls financially, or only those it operates, or may leave out recent acquisitions. The note describing the boundary is often a sentence in a methodology section, far from the table it applies to. Labels need to connect the two.
Sustainability reports mix facts with intentions. A page may state last year's measured emissions, a goal for ten years from now, and a description of offsets purchased. They sit side by side and read alike.
For this reason, labeling must separate them:
A model trained on labels that blur these categories may report a target as if it were an achievement. In a field where claims draw scrutiny, that is a serious error.
Report tables are rarely simple. Headers span several columns, multiple years sit side by side, and subtotals sit between rows. A figure may carry a footnote marker that changes its meaning, such as a note explaining that the number excludes one site.
Companies also restate earlier figures when methods change or data improves. The same year can appear in two reports with two different values. Annotators need a rule for marking a figure as restated, and for linking it to the footnote that explains the change.
Footnotes should be treated as part of the data, not as decoration. A label that ignores them will quietly return the wrong number.
Not every figure sits in a table. Some appear only in charts, in captions, or in narrative paragraphs. Annotators need guidelines on which of these to label and which to skip, so the model does not learn to guess numbers from bar heights.
Narrative text raises the same question. A sentence saying "emissions fell significantly" has no value to extract, while "emissions fell by a stated amount against a base year" does. Annotators should know the difference, and the base year, when given, should be captured as its own field.
People who work in emissions accounting or sustainability reporting read these documents differently. They know how scopes fit together and why a boundary note matters. They can tell a target from a result at a glance, and they know when two figures look alike but cannot be compared.
They also help write the label guide. Specialists define each field, decide what counts as a restatement, and explain how to treat unusual layouts. This matches Globik's approach of using verified domain experts for work that depends on judgment. For climate tech AI training data, that expertise is what separates data analysts can use from data they have to recheck.
Before a sustainability extraction project begins, a team can settle five things:
Sustainability reports hold valuable data, but only when each number is captured with its unit, scope, period, and boundary. For another example of labels shaped by field knowledge, see our guide on weed versus crop segmentation for precision spraying: why the mask edge decides the spray. Strong ESG data extraction begins with people who understand what the numbers mean.
It is the labeling of sustainability reports so AI can pull out figures along with their unit, scope, period, and boundary.
The same figure can mean very different things depending on its unit, scope, reporting period, and the sites it covers.
Targets are future goals, not achievements, and mixing them into measured data produces misleading outputs.
Footnotes can change what a figure includes, and restated figures replace earlier values, so both need explicit labels.
They understand scopes, boundaries, and reporting conventions, so they can write clear rules and review the work accurately.
