Same Baht, Opposite Sign: How AI Summary Turns a Whole-Market Search Into Who Gains, Who Loses and Who Hedged — With Every Row Cited

Two Thai listed companies, writing about the same currency move in the same year. PTT, in its management discussion for FY2025 (SEC filing, page 2):

"Thai Baht appreciated against US$, with the average foreign exchange rate strengthening from 35.5 Baht/US$ in 2024 to 33.0 Baht/US$ in 2025, resulting in lower revenues denominated in US$."

And TEAM, in FY2026 Q2 (SEC filing, page 1):

"The improvement was supported by a decrease in the cost of sales when translated into Thai Baht, resulting from favorable foreign exchange movements, as the Company has a high proportion of raw material purchases denominated in foreign currencies."

Same baht. Opposite sign. One company sells in dollars and books fewer baht; the other buys in dollars and pays fewer baht.

That is the whole reason a market-wide currency question is hard. Finding the passages is the easy part now — a theme search across the market returns them in a second. The hard part is the next hour: reading twenty-five passages from twenty companies and sorting them into who was helped, who was hurt, who had already covered it, and who merely mentions it. That sorting is what AI Summary on Boom Leverage Terminal does, and this article is about exactly what it does, what it refuses to do, and how to check it.

What AI Summary is — and what it is not

AI Summary went live on the Terminal's scan page on 11 October 2026. It is a switch on the page you already use. When it is on, the summary is written over the result set you just searched — the same passages, the same companies, the same fiscal years you paid credits to see. It does not go back to the internet, it does not read anything you did not search, and it does not bring its own opinion.

The format is fixed, and it is fixed on purpose:

• Grouped by impact angle, not by company. The headings are the angles of the question — typically who gains, who loses, who hedged, who only names it as a risk factor — named for the topic you searched. Headings that sort by data type ("rows with numbers", "rows without numbers") are forbidden by the rules the model writes under.
• One lead sentence per group, saying who is in it and how they were affected.
• A four-column table per group: Company · Period · Main impact · Cause / context. One row is one company. If a company spoke about the topic in several periods, it stays one row citing every source.
• Every row carries a citation number [n] that points at a numbered source card under the summary — ticker, fiscal year, page, and the verbatim passage. A row without a citation is not allowed.
• ▲ and ▼ only when the source states the direction itself. If the filing does not say better or worse, no arrow.
• Years exactly as the source labels them. A source labelled FY2026 Q1 stays FY2026 Q1. The model is forbidden from converting a Western year into a Thai Buddhist-era year or back, and from abbreviating to two digits — more on why below.
• Companies the sources do not discuss are simply absent. An angle with nobody in it gets no heading. A company the sources do not clearly place goes under "named as a risk factor", not into a guess.

Why the format is locked: a summary that can choose its own shape can hide its weakest claim in a paragraph. A table with one company per row and a citation in every row cannot. Every cell is either traceable to a passage you can open, or it is not there.

The same engine also writes a summary in two other places: over a single company's Health Check (every cluster for one company — a separate walkthrough is in the Health Check article), and over one cluster in one fiscal year for that company. The rules differ slightly per surface; the citation contract does not.

What gets checked before you see a word of it

The summary is not shown the moment the model finishes. It goes through a validator first — the full mechanics are in how the citation validator checks every line, but the short version for this surface:

1. Every line that states a fact must cite a source, and the cited number must exist in the list you can see.
2. Every number in a line must appear in the text of a source that line cites. Numbers are what models invent, so a number that is in no source at all is cut, not repaired.
3. A line whose words plainly came from a different source than the one it cites is flagged as mis-cited, and gets one repair pass to move the footnote — the repair may fix a citation, never invent a figure.
4. A Buddhist-era year that no cited source can account for is cut. This rule exists because of a real production run on 11 October 2026: sources labelled FY2026 Q1 came back in the draft as "1/2566" and "1Q66" — the model had "translated" 2026 into the wrong Thai year (the right one would be 2569). The prompt now forbids the conversion and the validator catches it if the prompt is ignored.
5. The "(transcript)" label on an Opp Day source comes from the source's own metadata, not from the model. On the first run of the per-angle table, the model wrote "(transcript)" on 18 of 24 rows when none of the cited sources was an Opp Day transcript. Anything the model writes in that slot is now stripped before validation.

If some lines fail, the lines that pass are shown with a note saying how many were cut. If no cited line survives, the whole summary is withheld and you see the sources instead.

Rule from model validation: a model output you cannot reconcile line by line is not a weaker output — it is an unvalidated one. The validator is not there to make the summary look better. It is there so that what reaches you has already survived the first question I would ask of it.

What it costs, exactly

Credits on the Terminal are spent per action, and AI Summary has its own line:

Three conditions sit on that 2:

• It is charged only after the validator. If the AI fails, times out, or the answer is withheld, you are not charged for the summary.
• It is charged only when at most half of the lines were cut. A mostly-wrong answer is free — the model is not paid for guessing.
• The replay is matched on what you actually asked: the same surface, market, title and scope. Change the years, the axes or the company list and it is a new summary.

Separately, AI usage counts against a weekly AI allowance that resets every Monday (UTC) and is shown in the account menu under today's credits. AI Summary also requires you to be signed in; searching without signing in works exactly as before.

Worked example: the stronger baht, FY2025–FY2026

To make the format concrete, I ran the theme search the way a reader would — the Thai phrase for "the baht strengthening", financials and risk axes, FY2025–FY2026 — plus a second search on hedging language. Below are real passages from that result set, quoted exactly as the engine served them, each with the filing link and page.

Companies describing a cost or gain benefit:

• TEAM, FY2026 Q2, page 1: "The improvement was supported by a decrease in the cost of sales when translated into Thai Baht, resulting from favorable foreign exchange movements, as the Company has a high proportion of raw material purchases denominated in foreign currencies."
• PATO, FY2025 Q2, page 3: "However, due to appreciation of the baht within the quarter, the exchange rate was in between 34.03-32.60 baht to a dollar resulting in an increase in forex gain."
• BGRIM, FY2026 Q2, page 2: "Net Profit – owner of the parent increased significantly to THB 676 million in Q2'2026, compared to THB 7 million in Q2'2025, reflecting the above-mentioned factors plus a non-cash transaction of unrealised FX gain from the net position of outstanding USD debts and foreign currency transactions."

Companies describing a revenue headwind:

• PTT, FY2025 Q3, page 2: "Thai Baht appreciated against US$, with the average foreign exchange rate strengthening from 35.9 Baht/US$ in 9M2024 to 33.3 Baht/US$ in 9M2025, resulting in lower revenues denominated in US$."
• MEGA, FY2025, page 8: "Excluding the currency impact from the appreciation of THB against USD YoY (and the decline in Mega We Care Myanmar business, largely attributable to 1H25), Mega We Care revenue in other markets recorded low double-digit growth."

Companies describing a hedge:

• TVO, FY2025 Q3, page 3: "The Company managed to hedge foreign exchange risk and commodity price risks with a forward contract. This resulted in a foreign exchange and derivatives gain of Baht 48 million."
• WINMED, FY2026 Q2, page 2: "The Company managed foreign exchange risk through FX Forward Contracts to control import costs and reduce the impact of exchange-rate fluctuations."
• SSF, FY2026 Q1, page 2: "The Company and its subsidiaries will continue to enter into forward foreign exchange contracts to hedge against the risk of currency fluctuations"

Companies that name the move but whose quoted line does not state the impact:

• CM, FY2025 Q3, page 1: "appreciation of the baht, which strengthened from the average exchange rate of foreign sales of 34.87 baht per dollar to 32.17 baht per dollar, or an average appreciation of more than 7.7%"
• YUASA, FY2026 Q1, page 2: "the appreciation of the Thai Baht against the US Dollar."

How those rows are arranged

Here is the same set laid out in the AI Summary format. I composed this table by hand from the quotes above for this article — it is an illustration of the format, not a captured AI output. Citation numbers refer to the order of the quotes in this section.

Who gains — companies whose own text ties the stronger baht to lower costs or a currency gain.

Who loses — companies whose own text ties it to lower reported revenue.

Who hedged

Named as a factor — the quoted line does not say which way it cut.

Look at what the rules force here. MEGA gets no arrow, because its sentence describes growth excluding the currency effect rather than stating a direction. CM sits under "named as a factor" even though its numbers are the most precise in the set, because the quoted line does not say whether the move helped or hurt — you have to open page 1 to find out. That is the correct behaviour. The format is allowed to be less satisfying than a confident paragraph. It is not allowed to be wrong in a way you cannot see.

A workflow you can copy

This is the way I use it on a theme I do not yet understand.

1. Type the theme, not the conclusion. "The baht strengthening", not "who benefits from a strong baht in 2025". Years go in the year filter, not in the question — the reasoning is in why you type a theme, not a year.
2. Set the scope deliberately. Whole market first. Pick the fiscal years and the axes — up to five axes is one credit.
3. Read the hit list before you turn the summary on. Thirty seconds of scanning tells you whether the search found the theme or something adjacent. A summary over the wrong rows is a well-formatted wrong answer.
4. Turn on AI Summary. Two credits, only if the checked answer is shown.
5. Read the table for structure, not for conclusions. Which angles have many companies? Which have one? Is the "who hedged" group empty — and is that because nobody hedged, or because the search did not reach hedging language? (In my example, the hedging rows came from a second search.)
6. Open two citations in every group. Click [n], read the passage, then open the filing page. If the row's "main impact" is not what the passage says, you have found the limit of the summary for this topic.
7. Export it. "Copy Markdown" or "Download .md" gives you the title, date, scope, the summary with its [n] markers, and every source — ticker, fiscal year, page and quote — with the disclaimer attached. That file is an audit trail, not just notes.
8. Ask the follow-up in chat. "Continue in chat from this source set" opens Boom Leverage AI with the same sources already loaded, so a question like "which of these gave a sensitivity figure?" re-uses the rows instead of searching again.

Practice I would hold anyone to: never forward a summary you have not clicked into at least twice. Two random citations per group is cheap, and it is the difference between a research note and a rumour with footnotes.

The traps — read these before you trust any row

1. The sign of an impact is not its size. "▲ higher forex gain" tells you the direction the company reported. It does not tell you whether that gain was material to the year. PATO's sentence and BGRIM's sentence are both "gains"; one is a line in a quarterly discussion, the other is attached to a profit figure. The table puts them side by side because they share an angle, not because they matter equally.

2. A currency gain on cash or debt is not an operating benefit. BGRIM's own text calls its gain "a non-cash transaction of unrealised FX gain from the net position of outstanding USD debts". TTA, in the same result set, reported "realized/unrealized gains from foreign exchange of Baht 243.5 million, mainly from unrealized gains on US Dollar-denominated cash and deposits" (FY2026 Q2, page 2). Those are balance-sheet translation effects. Grouping them with TEAM's cheaper raw materials under "who gains" is technically correct and economically misleading if you stop reading at the heading.

3. Periods disagree, and the summary does not reconcile them for you. The baht did not move one way for two years. TTA's FY2026 Q2 passage describes "the appreciation of US Dollar and Canadian Dollar against Thai Baht from year-end 2025" — the opposite direction from PTT's FY2025. ALUCON in FY2026 Q2 writes that "The U.S. dollar remained strong against major currencies" (page 1). A table spanning FY2025–FY2026 can legitimately hold rows from both regimes. The Period column is there so you can see it; the summary will not merge them into one story, and you should not either.

4. The summary only sees the rows you bought. It summarises the result set of your search — your years, your axes, your company scope, your depth of history. If a company discussed the theme in a section your axes did not cover, or in a year outside your filter, it is not missing from the market; it is missing from your search. An empty angle is a statement about the result set, never about the world.

5. The search can return neighbours of the theme. A theme search finds passages close in meaning, and close is not identical. In the raw result set for this example, rows like "Cash and cash equivalents 330.87 250.46 80.41 32.11%" came back alongside the currency passages — a balance-sheet line with no currency content. The summary rules tell the model to skip companies whose sources do not discuss the topic, but you should still read the hit list (workflow step 3).

6. "Not stated" is an answer. When a row says the impact is not stated, the summary is telling you the truth about the passage. The temptation is to fill it in from what you know about the company. If you do, write it as your judgement, not as the filing's.

Where I draw the line: the summary is a sorting tool over evidence. It is not a forecast, not a ranking of who will outperform, and not a substitute for reading the page. Use it to decide which five passages deserve your next hour.

Where this fits

AI Summary is the bridge between two things the Terminal already did: the cross-sectional search that finds every company talking about a theme, and the chat that lets you interrogate a set of sources. The search answers who mentioned it. The summary answers how do they split. The chat answers what exactly did this one say, and compared to which other.

All three run on the same rule: no line reaches you without a passage you can open behind it.

Try it, or put it in front of a desk

Sign in at terminal.boomleverage.com, run a theme search across the whole market, and switch on AI Summary. A free account gets daily search credits and a weekly AI allowance with no card on file — enough to run the stronger-baht example above and check every row against the filing yourself, which is what I would want you to do first.

If your team sorts market-wide themes into exposure groups for a living — a research desk, an asset manager, a credit or treasury function — the scoping conversation for seats, export volume and history depth is at boomleverage.com/th/enterprise or contact@boomleverage.com.

Disclaimer: This article is written for education and to explain how to read disclosures that listed companies make publicly. It is not investment advice, it recommends no security, and it guarantees no outcome. · The example table was composed by hand from the quoted passages for illustration; it is not a captured output of the product · AI Summary is generated by AI and can be wrong — the product shows "summarised by AI, may contain errors; check the sources before use" under every summary, and every row should be checked against its cited passage and the original filing · Grouping a company under "gains", "loses" or "hedged" reflects only what the quoted passage states for that period; it is not an assessment of the company's overall exposure, and none of these groupings has been backtested against market outcomes · Filing passages are quoted from documents published by the listed companies via the Securities and Exchange Commission, Thailand; credit prices and features described are as of 11 October 2026 and may change · Investing carries risk; past results do not guarantee future ones