A customer harness is the software around AI agents that act on customer relationships: what they read, what they can do, who approves it, and how you know it worked. The explainer covers the term. This is the question list I would take into a vendor evaluation.
Before you talk to a vendor
Answer these about your own stack first. A harness runs on them.
- Do we have one written ICP, specific enough to build an account list from?
- Do sales, marketing and RevOps agree on stage definitions, with entry and exit criteria?
- Is lead source stamped once at creation and never overwritten?
- Which of our messaging sources is current: website, deck, sequences, enablement? Do they agree?
- Who approves anything that reaches a customer today, and is that written down?
If more than one answer is “no,” fix those first. An agent will act on whatever version of the definition it reads.
Data and context
- Which objects and fields can the agents read, and can we restrict them by team or segment?
- How fresh is the data the agent sees: live, hourly, nightly?
- Can the agents read our own documents (positioning, ICP, playbooks), and which version wins when two disagree?
- Can we see exactly what context an agent had when it took an action?
Rules and approvals
- Which actions can an agent take alone, and which wait for a person? Who sets that, and where?
- Can approval rules differ by segment, deal size or region?
- What stops an agent from contacting the same person from two motions at once?
- How are rule changes recorded, and can we roll one back?
Evidence and logs
- Is every agent action logged with its inputs, output and the rule that allowed it?
- When an agent writes a claim about our product or a customer, can it show the source?
- Can we export the logs?
Pipeline measurement
- Can the platform report opportunities created, meetings held and win rate for agent-touched accounts, against a holdout?
- How does it attribute a deal that a person and an agent both touched?
Leaving
- If we leave, do we keep our rules, prompts, definitions and logs in a format we can use elsewhere?
- What does it cost as agent volume grows: per seat, per action, per credit?
How to score the answers
Weight questions 1 to 5 and 17 most. The first five decide whether agents will act correctly. Question 17 decides whether you will ever know if they did. If a vendor answers 6 to 16 well and cannot answer 17, you will be able to count what its agents did, and you will not be able to tell whether they created pipeline.
For a working example of these rules in a small operation, see How I run SandsDX on a GTM harness.