Docs

How Amdahl processes your data

Amdahl processes your CRM, calls and email into one searchable shape before any query runs: speaker-attributed, joined to the deal, tagged across 10 dimensions, and about five minutes fresh. Your agent reads computed facts instead of raw text.

Your agent reads computed facts, not raw text. Amdahl processes your CRM, calls and email before any query runs. A question then costs a few thousand tokens instead of a few hundred thousand, and the answer cites where it came from.

What it reads

Connections sync your sources: CRM accounts and deals, call and meeting transcripts, and email. list_connectors (or the Connections reference) lists every source you can connect today. Each sync runs on the server. You never send this data with a request.

What it does to the data

  • One shape across every source. A call, a meeting note and an email land in the same structure, so one query reads them all.
  • Speaker-attributed. Each thing said is attributed to its speaker, so a search can read only what your customers said, not your own team.
  • Joined to the account and the deal. A quote knows which company and which deal it came from, so you can scope a question to closed-won accounts or to one account.
  • Tagged across 10 dimensions. Classifiers tag every message before any query runs.
  • About five minutes fresh. New records are ready to query about five minutes after they reach the source system.

The API names behind these words (corpus, warehouse, interaction, grain) are in the Glossary.

What your agent gets back

Your agent asksIt getsCall
A number, a list or a slice of dealsRows, counts, win rates and cycle time by segment, with the SQL that produced themSearch
What customers said about a topicQuotes in their own words, with their sourceSearch
Whether a message holds up against your customersAn evaluation, and the claims nothing in their words backsOptimize a message with evidence: "workspace"
One question that needs several readsOne cited answerResearch

Why it matters

Raw transcripts are expensive and unreliable context. On amdahl.ai's context engineering page, 200 calls come to about 800K tokens of raw transcript, while the same calls, processed, answer a query in about 2K tokens: about 400x fewer. The same processing makes answers repeat. See Proof.

What does not change

  • Your data stays scoped to your workspace. Every read carries your workspace filter and any data-scope rule your admin has set.
  • Search never changes your data, and nothing Amdahl does sends a message or writes to your CRM.
  • A record in your CRM is not a customer. CRMs create company records from inbound mail, so your data also holds vendors, recruiters and press. For ICP, voice-of-customer and win/loss questions, scope to accounts with at least one call or deal.

The full list is under Data handling.