# A proof of concept proves something only when it runs your exports, against criteria you wrote down before the first demo

PROOF OF CONCEPT, read 10 Nov 2026

**The short answer.** Run the proof of concept on your own exports for four weeks, with pass criteria written first: connectors working, a known-overlap merge, your scoring rule reproduced, tickets routed to real owners, and a report an auditor would accept. Score each criterion on the same sheet for every candidate.

## Set the terms before the first call

Vendor demos use clean data. A proof of concept earns its name only on yours. Agree in writing: the duration (four weeks is workable), the data (your exports, not sample data), who at the vendor supports it, whether the data is deleted at the end, and that the same script applies to every candidate. Expect to run your own security review before production findings leave your network, so start that early.

Sandbox and trial availability varies. Phoenix Security, for instance, publishes a free tier of up to 1,000 assets on its pricing page, and Vulcan Cyber's site navigation listed a free tool; treat these as vendor statements about offers that may change. Nucleus publishes no pricing (quote only), so a proof of concept is also where you learn the price.

## Write the pass criteria first

Pick six to eight criteria and write a threshold for each. The list below is a starting set; adjust it to your scope.

- Connectors: the day-one sources from the connector plan load without custom engineering from the vendor, in under a set number of days.
- Merge: on a 50-host known-overlap set, the unique count matches your hand count within a stated tolerance, and every merged record lists its sources.
- Rule: the scoring ladder from the risk scoring guide can be expressed, with KEV as a hard override, and a replay of 90 days of findings lands in tiers within 5 percent of your own calculation.
- Routing: 25 findings across five asset classes are assigned to the right team by rule, with the number of manual corrections counted.
- Ticketing: a ticket opens in your real system, updates on a status change, and closes the finding on resolution, and the reverse.
- Evidence: a report shows open items past due by tier and age, SLA attainment, and the history of one item's tier changes.
- Operations: connector failure is visible within a day, and a user without admin rights cannot change the rule.

## Use data that includes the ugly cases

Include a scanner export with duplicate hosts, an asset with no owner in the inventory, a finding with no CVE, a CVE with no EPSS score, a KEV entry on a low-criticality asset, and a CVSS 5.4 vulnerability that is in KEV. The KEV Lag Ledger has real examples: CVE-2021-26829 was listed on 2025-11-28 with CVSS 5.4 and an EPSS of 0.00252 that day. The platform should place it in your top tier because KEV listing says so.

If the vendor asks for cleaned data, ask what they cleaned, and then compare. The cleaning is the work you will do after buying.

Use real volume. A platform that handles 5,000 findings smoothly may behave differently at 5 million. Ask the vendor what volume it has tested at, and load a full week of your exports to see.

## Test the claims you cared about

Go back to the statements that moved the candidate onto your shortlist, and test each. If a vendor states that AI agents merge duplicates or assign owners, ask what the agent did on your data, whether a human can review it, and where the record is kept. Brinqa describes thresholds for when human validation is required for its attribution agent; ServiceNow's documentation says AI agent records are read-only by default and must be duplicated to modify. Whether those behaviors match what you need is something only a trial shows.

Capture timings: first load, daily sync, rule changes applied to existing findings, report generation. Measure on your volume.

## Score and decide

Use one sheet with the same criteria and weights for every candidate, filled in by two people who watched the same sessions. Keep a column for what the vendor did for you that a customer could not. If the vendor's engineer fixed a problem by hand, record it as a cost.

At the end, estimate year-one effort on your side: connectors to build, rules to write, owner data to clean. The platform that scored best with the vendor's engineer present may not be the one that runs best with yours. Record the decision and the evidence, so that the next audit and the next platform change start from a record.

If a candidate publishes no public technical documentation, say so in the sheet. A proof of concept can cover what the vendor shows you, but not what you will need to read in month eight.

Keep the rejected candidates' sheets as well. When the chosen vendor's renewal comes up, the record of what the alternatives did on the same data is your negotiating position and your audit trail for the selection.

## Sources

- Phoenix Security, pricing (vendor page): https://phoenix.security/pricing/ (read 10 Nov 2026)
- ServiceNow documentation, Vulnerability Response: https://www.servicenow.com/docs/api/khub/maps/x8~ohSTeWLioGLVI5x3Vgg/topics/16Miv0EjmvVF7qlqVbs6Aw/content (read 10 Nov 2026)
- Brinqa, AI layer (vendor page): https://www.brinqa.com/platform/ai-layer (read 10 Nov 2026)
- CISA, Known Exploited Vulnerabilities catalog (JSON feed): https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json (read 10 Nov 2026)
- ServiceNow documentation, AI skills and agentic workflows: https://www.servicenow.com/docs/api/khub/maps/x8~ohSTeWLioGLVI5x3Vgg/topics/RO3F5ODkc6PeIizmnEuGDg/content (read 10 Nov 2026)

Last reviewed 10 Nov 2026. Canonical: https://riskbasedvulnerabilitymanagement.com/guides/running-a-platform-proof-of-concept/
