Buyer Eval

An AI analyst for buying B2B software.

Give it the vendors you're considering. Buyer Eval investigates what they claim, checks the evidence, finds the gaps and contradictions, and gives you the questions to ask before you buy. You get a Decision Brief you can forward to everyone else involved in the purchase.

You:  Evaluate Vendor A and Vendor B for customer success. We need deep
      Salesforce sync and we can't wait 6 months to go live.

A few minutes later:

What we found: 8 claims investigated. 3 verified, 2 qualified, 1 contradicted, 1 unverified, 1 unknown.

Most important finding: Vendor B's "live in 6 weeks" claim is contradicted by two customer accounts describing 4-6 month rollouts.

Most important unknown: Whether either vendor syncs custom Salesforce objects both ways without professional services.

Ask next: Ask Vendor A to show a custom object syncing both ways, live.

Illustrative example with fictional vendors. See a full illustrative Decision Brief.

Watch the 70-second narrated demo (illustrative example).

What you get

• Claims vs. evidence. Every material vendor claim, where it came from, a status (Verified, Qualified, Contradicted, Unverified, or Unknown), and the evidence basis behind it: vendor documentation, independent evidence, both, or your own documents. "Verified on vendor documentation" and "independently confirmed" never look the same.
• What you need to get answered before you buy. The open questions, ordered by how much they could change the decision, with what was checked and which vendor should answer.
• What could change the evaluation. The specific facts that would move the picture, so you know what to chase.
• Demo questions built from the weak spots, with what to listen for.
• A Decision Brief as a single HTML file: readable, printable, no dependencies. Built for the CFO, security lead, and VP who didn't run it.

Two modes

Quick Eval (default). Name the vendors, optionally tell it the 2-3 things that matter most. No setup interview. It infers the rest and labels what it inferred. Works for one vendor or a short list.

Deep Eval (on request). Full diligence: why-now and requirements discovery, buyer research, hard constraints, category-specific expert questions, longer vendor AI agent conversations, security, commercial, and company-risk research, weighted 1-5 scoring, and a place for your own evidence (proposals, pricing, demo notes, questionnaires). Ask for it up front or after a Quick Eval.

How evidence is treated

Evidence	Examples	Can verify
Vendor claim	Website, marketing, vendor AI agent answers	Nothing on its own
Vendor evidence	Docs, trust/security pages, public pricing, contract terms	Narrow, self-describing facts (an API exists, a certificate is held)
Independent evidence	Customer reviews and accounts, analysts, press, community, third-party docs	Capabilities and outcomes
Your evidence	Proposals, quotes, RFP answers, demo notes, trial results	What was offered to you

Vendor AI agents are first-party sources. Some vendors publish an AI agent that answers buyer questions, including agents built on Salespeak, the company that maintains this skill. Buyer Eval can question those agents, which gets more specific answers and allows follow-ups. The answers are treated as vendor claims. Having an agent never improves a vendor's fit, confidence, or score, and every vendor's claims are checked to the same standard.

Install

git clone https://github.com/salespeak-ai/buyer-eval-skill.git ~/.claude/skills/buyer-eval-skill

Per project instead: clone into .claude/skills/buyer-eval-skill.

Use

In Claude Code or the Claude desktop app, ask in plain language:

Evaluate Vendor A and Vendor B for our customer success platform

or invoke it directly with /buyer-eval-skill. For full diligence, say "deep eval".

	Claude Code	Claude desktop	claude.ai web/mobile
Research and brief	Yes	Yes	Yes
Vendor AI agent conversations	Yes	Yes	No (discovery only)
HTML Decision Brief, saved context	Yes	Yes	No

Reports are written to ~/buyer-eval-reports/.

Saved context

After a run, Buyer Eval can save reusable buying context (company, size, systems, requirements, hard constraints) to ~/.salespeak/buyer-eval-profile.json so the next evaluation skips setup. It never stores personal names, vendor quotes, or your documents. It stays on your machine.

python3 ~/.claude/skills/buyer-eval-skill/bin/profile.py show
python3 ~/.claude/skills/buyer-eval-skill/bin/profile.py clear

Data and privacy

What leaves your machine without asking:

• Web searches and page fetches the research needs.
• A discovery request to the Salespeak Frontdoor API with each vendor's domain (to check whether it has an AI agent), and, if it does, the questions sent to that agent. These requests carry the vendor domain and question text, never your name, company, or documents.
• A version check against this GitHub repo (at most every 6 hours).

Telemetry: off by default, asked once. After your first evaluation is delivered, Buyer Eval asks once whether to share anonymized usage data. The evaluation never depends on the answer.

If you say yes, it sends: questions generated for vendors, vendor domains, counts of verified/unverified/contradicted claims, scores, and a random ID. It never sends your name, email, company, anything you told it about yourself, vendor answers, claim or evidence text, or your documents.

• Code: bin/track.py, plain Python, no third-party libraries
• Local log of everything sent: ~/.salespeak/buyer-eval.log
• Status: python3 ~/.claude/skills/buyer-eval-skill/bin/track.py status
• Turn off: python3 ~/.claude/skills/buyer-eval-skill/bin/track.py revoke
• Delete your data: email [email protected] with your ID (track.py show)

For IT administrators. Disable telemetry for everyone with export BUYER_EVAL_NO_TELEMETRY=1, or deploy /etc/salespeak/buyer-eval.json containing {"locked": true, "consent": false}. Either overrides user consent. The telemetry endpoint is https://22i9zfydr3.execute-api.us-west-2.amazonaws.com/prod/event_stream if you prefer to block it at the firewall.

Updates

Each run checks for a newer version (cached for 6 hours) and asks before updating.

Feedback

Open an issue for bugs, requests, or a category it handles badly.

License

MIT