ESG data extraction that moves your holdings from estimated to reported, every figure cited to its page
Holdings move from estimated to reported ESG data, each figure cited to its page.
See one case, screen by screen ↓ESG data extraction, also called sustainability report data extraction, is reading the figures an issuer reports — Scope 1, 2 and 3 emissions, energy, water, waste, board gender diversity, the gender pay gap — out of its sustainability and annual reports and into a standard ESG template, for the SFDR PAI statement, PCAF financed emissions and client reporting. ESG Data Extraction does it with seven agents that read each report in its own language, twice, cite every figure to its page and table, and propose its data-quality score; an analyst confirms and a different steward approves each issuer.
AI in asset management: a buyer's guide to agentic AI for investment managers →
Updated
The SFDR principal adverse impact statement and PCAF financed-emissions reporting both reward a figure the issuer reported over one a vendor estimated. Vendor ESG data covers the large listed names, so the small-cap, emerging-market and frontier holdings of a global book sit on revenue-based estimates at PCAF score 4 or 5. Their own sustainability and annual reports exist — often only in Korean, Japanese, Chinese, Portuguese or Bahasa, with tables in thousands of tonnes or terajoules, fiscal years that end in March, and parent-only figures next to consolidated ones.
So ESG data collection for those holdings is manual: analysts copy the figures into a spreadsheet template by hand, thousands of reports a season, between April and October. A figure in thousands read as units moves a fund's carbon footprint by a factor of 1,000. A March year-end lands in the wrong reference year. The second check is a sample, not every figure, and when an auditor asks which page a published Scope 1 number came from, nobody can show it.
Estimated split for one report of about 150 pages and 38 template fields, by hand (about 2 to 3 analyst hours) and with the solution, where the analyst confirms what the agents read.
Seven agents find, read twice, convert, check and score every figure; an analyst confirms, a steward approves.
Quillfield Capital, a fictional global asset manager, holds Daeyang Chemical Co., Ltd., a Korean chemicals issuer, across six funds: $412.6 million. Its Scope 1 sits on a vendor estimate of 1.18 Mt at score 4. Its 156-page Korean sustainability report came out on June 30, 2026; the English version is not expected before December. Here is what happens next, in the working solution.
Helena's home screen shows 214 issuers waiting for a steward (23 due today, 4 overdue), 1,118 of 1,686 gap issuers on reported data, 287 open exceptions and 2,913 of 3,040 reports read. Overnight the Intake agent matched 212 new reports, the Second reader agent read 4,812 figures blind and sent 41 disagreements to analysts, and the Plausibility agent held 17 new outliers and found the issuer's own explanation for 11 of them.
Morning brief: “41 issuers are ready for a steward since yesterday, Daeyang Chemical first: its Korean report moves Scope 1 from an estimate of 1.18 Mt to a reported 1.42 Mt. Client reporting read back 2 figures differently in this morning's load; I have not retried on my own.”
In the ESG template grid, every issuer is a row and every field a column, each value with its data-quality score. Daeyang's Scope 1 cell opens page 87, Table 4-2, 온실가스 배출량 (천 tCO₂eq): the cropped Korean table beside the English in template units, with the conversion written out. The holding moves from 4, estimated from revenue, to 1, reported and verified, on approval.
Units agent: “1,421.6 thousand tCO₂eq × 1,000 = 1,421,604 tCO₂e” · fiscal year Jan 1 – Dec 31, 2025 → FY2025.
The Extraction agent found Table 4-2 on page 87 and read 1,421.6 thousand tonnes at 96% confidence. The Second reader agent read the same cell blind and agrees. The Plausibility agent compared it with last year's 1,380,220 tCO₂e (+3.0%), chemicals peers at −4% to +7% and the vendor estimate of 1.18 Mt, which was 17% low: plausible. The Assurance agent found that limited assurance covers the figure on page 148.
Helena asks the cell in plain words. The Units agent answers that the parent-company figures are in the exchange business report on page 212 — Scope 2 of 598.0 kt, with no Scope 1 for the parent alone — and that the template's boundary rule B-01 says consolidated. It will not switch on its own; an analyst can edit with a reason code, and the steward will see it.
“Rule B-01 says consolidated, so I will not switch on my own.”
Hazardous waste reads 38,412 t against 21,904 t last year, outside the ±50% band, so the Plausibility agent held it. Then it searched the report and found footnote 3 on page 89: a one-time disposal of spent catalyst after aged catalyst was replaced in 2025. Seo-yeon Kang, who reads Korean, accepted the hold at 9:10 a.m. with the reason code “One-time event explained by issuer”.
The Plausibility agent never releases a hold itself.
The issuer review walks the run: report found, tables read twice (38 fields: 31 reported, 7 not disclosed), units and fiscal year, plausibility, the data-quality score, the analyst's confirmation at 9:12 a.m., and the steward's approval, by a different person. Beside it, what approval changes in Quillfield's financed emissions data: EVIC $14.9 billion, and financed Scope 1 from 32,675 t estimated to 39,366 t reported.
“One boundary choice. Scope 2 uses the consolidated 608.3 kt from the report, not the parent-only 598.0 kt in the exchange filing (rule B-01).”
The approval shows the Scope 1 score moving from 4 to 1, the 31 figures each with page and table, and Seo-yeon Kang's confirmation. Helena picks the reason “Figures match the report; assurance scope checked” and ticks the attestation. Had she edited a figure on this issuer, the solution would have offered the approval to another steward.
“I checked the cited pages, the holds and the assurance scope, and I did not edit any figure on this issuer.”
Daeyang Chemical is approved and queued for batch PB-2026-1019-B at 6:00 p.m. London. The Publishing agent will load it to the ESG data warehouse, the portfolio management system, the PAI engine, the financed-emissions engine and client reporting, then read every value back.
The Publishing agent never publishes an issuer a steward has not approved.
The next batch holds 42 issuers and 1,428 figures for 5 systems, Daeyang among them. In the morning's batch, client reporting read back 698 of 700 values: it stores water in thousands of m³ and rounded two values on the way in. The agent flagged the difference and did not retry on its own; out-of-cycle loads and rollbacks need the Chief Data Officer.
The data-quality screen recalculates after every approval: the 1,686 gap issuers have moved from an average score of 4.3 at the start of the season to 2.4, and 88% of the listed book's financed emissions now rest on reported data, from 71%. Coverage of all 14 PAI indicators is shown by fund, ahead of the PAI statement due June 30.
The same issuer, seen by five of the people who carry the season — what their work looked like, and what it looks like now.
Seven agents find, read twice, convert, check and score every figure; an analyst confirms, a steward approves.
Watches filing calendars, exchange feeds, issuer sites and the document store; matches each report to its issuer by ISIN, LEI or name in any script, and detects its type, language, fiscal year, restatements and duplicates.
Finds the tables that hold each template field and reads the value, unit, year and boundary, with page, table and cell.
Reads the same tables blind, with a different method, and compares its reading with the first.
Converts units, fiscal years, currencies and boundaries by the template's rules, and writes the arithmetic out.
Compares each figure with last year, restated years, sector peers and vendor estimates; holds outliers and looks for the issuer's own explanation in the report.
Reads the assurance statement and its scope, page by page, and proposes the PCAF data-quality score for each scope.
Prepares deltas for approved issuers, loads them to the warehouse and downstream systems, reads every value back and keeps versions for rollback.
Confirm, approve. The agents propose; a named person decides.
Ask a question or say what to do. Every answer cites the page, and the agents act only with a person's say-so.
Why did Daeyang Chemical's score move from 4 to 1?
Because its own report replaces the vendor estimate. Page 87 of the Korean sustainability report gives Scope 1 of 1,421.6 thousand tCO₂eq, which is 1,421,604 tCO₂e, up 3.0% on last year and inside the peer range. Hanbit Assurance's limited assurance covers Scope 1 and 2 on page 148, so the score becomes 1 once a steward approves.
Which Asia Pacific UCITS holdings are still on estimates?
In the Asia Pacific Equity UCITS Fund (Art. 8), four holdings still carry estimates (score 4 or 5): Daeyang Chemical (awaiting your approval), Hoshimura Electric Works (2 exceptions due today), Tung Hsin Semiconductor (being read, 62%) and Siam Riverside Foods (report expected Nov 14).
Hold every Scope 3 rise above 40% for a second look
That changes threshold T-03 (Scope 3 band, today ±35%), which I can't change myself. I have drafted proposal RP-122 for the Head of Sustainable Investing: on last season's data it would have held 23 more issuers and missed no known error.
Prepare the evidence pack for Daeyang Chemical
The pack holds all 38 fields with page, table and the cropped Korean source beside the English, the report's fingerprint, nine agent steps with their versions, the accepted hold on hazardous waste, Seo-yeon Kang's confirmation and the steward's approval.
13 screens from the working solution, on its sample data. Pick one to see it large.
The thresholds and template rules the agents work to are settings. Each change is proposed by one person, approved by another, versioned and logged.
| Setting | Default | Choose from |
|---|---|---|
| Pre-confirm for the analyst | Both readers agree at 97% or more | Below it, the analyst decides with both readings |
| Hold an emissions figure | Year on year beyond ±35%, or outside the sector peer band | Proposal RP-121 to widen the Scope 3 band to ±45% waits for the Head of Sustainable Investing |
| Hold water or waste | Year on year beyond ±50% | The hazardous waste band was widened from ±40% by change T-07: too many one-off disposals were held |
| Match a report automatically | Intake confidence 95% or more | Below it, a person confirms or changes the issuer |
| Score 1 (verified) | Signed limited or reasonable assurance that covers the figure | Reported without that assurance scores 2 |
| Boundary (rule B-01) | Consolidated figures | The parent-only figure is recorded when both exist |
| Fiscal year (rule F-01) | The calendar year holding most of it | FY ending Mar 31, 2026 → FY2025 |
| Steward approval | 2 business days | Escalation to the Head of Sustainable Investing after 3 business days |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample data · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published figures (a leading ESG data vendor — carbon data on about 16,000 listed issuers; a carbon-data provider — vendor data reaches up to about 70% of a portfolio's financed emissions; SFDR Delegated Regulation (EU) 2022/1288 — 14 mandatory PAI indicators, statement each June 30; PCAF — data-quality score 1, reported and verified, to 5, estimated). Quillfield Capital and the people, issuers, funds, assurance providers and figures named on this page are fictional — characters and sample data in the working solution.
ESG data extraction, also called sustainability report data extraction, is reading the figures an issuer reports — Scope 1, 2 and 3 emissions, energy, water, waste, board gender diversity, the gender pay gap — out of its sustainability and annual reports and into a standard ESG template, for the SFDR PAI statement, PCAF financed emissions and client reporting. ESG Data Extraction does it with seven agents that read each report in its own language, twice, cite every figure to its page and table, and propose its data-quality score; an analyst confirms and a different steward approves each issuer.
Yes. The standard template has 38 fields mapped to the 14 mandatory PAI indicators, the PCAF inputs and European ESG Template (EET) fields, or the agents fill your own template, imported from Excel or CSV. The data-quality screen shows reported, estimated and not-disclosed shares for each of the 14 indicators, by fund.
The Assurance agent reads the assurance statement and its scope. A figure covered by signed limited or reasonable assurance is proposed at score 1; a reported figure outside that scope at score 2. In the demo, limited assurance over Scope 1, Scope 2 and energy gives score 1, and its Scope 3 emissions data, 9.86 Mt across five reported categories and not covered, gives 2. The score on the holding moves only when a steward approves.
Yes. The demo season covers reports in Korean, Japanese, Chinese, Portuguese, Spanish, Indonesian, Turkish, Polish, German, French and other languages. Every cell shows the cropped table in its original language beside the English, so an analyst can check it without translating, and the agents translate a footnote when asked.
The Units agent applies your template rules and writes the arithmetic out: thousands of tonnes to tonnes, terajoules to MWh, a fiscal year ending March 31, 2026 to FY2025, and consolidated over parent-only figures, with the parent-only figure recorded. It never changes a rule and flags every boundary choice for the analyst.
The Plausibility agent compares it with last year, restated years, sector peers and the vendor estimate, and holds it when it falls outside the band. It then looks for the issuer's own explanation in the report and shows it to the analyst, who accepts with a reason code or edits. It never releases a hold itself.
Yes. The agents read, convert, check and propose; they never publish, release a hold, move a score or change a rule. An analyst confirms every issuer, a different steward approves it with a reason code and attestation, and every load to the warehouse is read back, with rollbacks approved by the Chief Data Officer.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your ESG template, unit and boundary rules and a sample of past issuer reports — and it goes live once every quality gate has passed. Setup includes a pilot back-test on issuers your analysts already did by hand, compared figure by figure.
Issuer questions answered in minutes, every claim pinned to its source page.
1-min demoPortfolio Company ESG Data CollectionEach company confirms prefilled values, and no outlier reaches an LP unexplained.
1-min demoESG Risk MonitoringControversies in holdings scored within hours, and no fund's screen breached unseen.
1-min demoWe'll run ESG Data Extraction on a sample of your own issuers' reports, against the figures your analysts already keyed.