如何用 Claude 追蹤公司在業績電話會議反覆談論的內容，並避開 Microsoft 的拼寫陷阱

All outputs are real tool calls to the Equibles MCP server from 21 October 2026, trimmed.

What management keeps bringing up, call after call, is often the earliest sign of where a company is putting its weight. Counting a word across a dozen transcripts is easy for an AI agent. Counting it correctly takes one extra step, and Microsoft is a good example of why.

Connect it first

In Claude, open Customize → Connectors → Discover, search for Equibles, click Connect to Claude and Authorize, then switch it on in a chat from + → Connectors. In ChatGPT: Plugins, search Equibles, Install plugin. In Claude Code, claude mcp add --transport http equibles https://mcp.equibles.com/mcp and then /mcp to sign in. The free plan covers 100 requests a day; other clients are in the setup docs.

1. Count the phrase across calls

The prompt: "Using Equibles, count how often Microsoft said 'agents' on its last 12 earnings calls." The agent calls GetTranscriptPhraseTrend, which counts whole-word, case-insensitive matches in each transcript, gives a rate per 10,000 words so long and short calls compare fairly, and splits prepared remarks from Q&A.

The switch is sharp. Almost nothing for a year, then 10 to 21 mentions on every call from October 2024 on. And it's management driving it: in the latest call, 16 of the 17 mentions are in the prepared remarks. The quotes come back too, so the context is checkable: "You can think of agents as the new apps" (January 2026).

2. Count the word it might be replacing

The obvious follow-up: "Now do the same for 'Copilot'."

Call        Copilot   Per 10k
FY2024 Q1      32       41.8
FY2025 Q1      63       73.5   ← peak
FY2025 Q4      40       53.9
FY2026 Q2      28       32.4
FY2026 Q3      27       29.0
FY2026 Q4      13       13.3

That looks like a clean story: Copilot fades from 73.5 per 10,000 words to 13.3, agents take over, and in the latest call "agents" (17) finally beats "Copilot" (13).

It's wrong.

3. Check spelling variants before believing a trend

Read the quotes the tool returns and one of them says "M365 co-pilot", with a hyphen. A whole-word match for "Copilot" doesn't count that. So: "Count 'co-pilot' too."

Call        Copilot   co-pilot   Combined   Per 10k
FY2025 Q4      40         0         40        53.9
FY2026 Q1      35         0         35        44.9
FY2026 Q2      28        11         39        45.1
FY2026 Q3      27        19         46        49.4
FY2026 Q4      13        16         29        29.6

From January 2026 the transcripts spell it both ways. Put back together, Copilot held steady through April 2026 and only dropped in the July call, and it's still mentioned more than agents: 29 to 17.

The general rule: before reading a trend into one word, ask the agent to count the variants. Hyphens, spaces and joins ("co-pilot", "GenAI" or "gen AI"), singular and plural ("agent" isn't matched by "agents"), and related forms. "Agentic", for instance, went from zero to 5 a call over the same period, all but two of those in prepared remarks.

4. Use the split between prepared remarks and Q&A

Prepared remarks are what management chose to say. Q&A is what analysts asked about, plus the answers. Copilot's Q&A count went from 29 in October 2024 to 4 in July 2026 (9 with "co-pilot"). "Agents" in Q&A has been 0 to 3 in most calls. Management talks about agents a lot more than analysts ask about them.

The prompt I reuse

Using Equibles, track "<PHRASE>" on <TICKER>'s last 12 earnings calls:
1. Mentions per call, rate per 10k words, prepared remarks vs Q&A.
2. The same for every spelling variant and related form; combine them.
3. The same for the term it may be replacing.
4. Two quotes per call so I can check the context.
5. Flag calls with no transcript text; never treat them as zero.

What a word count can't tell you

• Meaning. Twenty mentions can be enthusiasm or twenty answers to worried analysts. Read the quotes.
• Revenue. Talking about agents isn't selling them. Pair the count with the revenue breakdown before drawing conclusions.
• Missing calls. When a transcript isn't on file, the tool says "no text" rather than zero. A gap is a gap, not a drop.

Every company's calls are also on its earnings-calls page, like Microsoft's.

Nothing here is investment advice.