The test bench for AI agents
Measure, inspect and govern your AI agents
Pivotal Bench scores every answer of your agents against a reference, maps their instructions, tools and dependencies, and rates the quality of their prompts. First connector: Dust.
Free during the beta-testing phase
Three jobs, one platform
Benchmark
Every answer is judged Correct, Partial or Incorrect. Compare test sets across platforms, annotate differences and generate ranked reports.
Inspection
See inside the platform: which agents exist, what they are told to do, which tools and skills they use, who calls whom.
Governance
A Craft.ecQ score from 1 to 5 per agent, company-wide shared spaces, reader/editor roles and a history of reports.
What you measure
- 3
- verdicts: Correct, Partial, Incorrect
- 2
- scores: strict and weighted
- 8
- Craft.ecQ criteria per agent
- 1–5
- instruction quality score
How it works
- STEP_01
Connect
Add your Dust API key and workspace to a space.
- STEP_02
Import
Upload your questions and expected answers as CSV. Columns are detected automatically.
- STEP_03
Run
Agents answer in the background, in parallel. AI judges every answer and a second look re-checks doubtful verdicts.
- STEP_04
Decide
Compare datasets side by side and generate an executive report with charts and recommendations.
Security & confidentiality
Your data belongs to your company. Never sold, never read without your consent.
Isolated per company
Each company only sees its own data. Even the platform administrator cannot read it.
Support on authorisation
Access granted by your administrator: 1 h, 24 h or 7 days, read or edit, revocable.
Encrypted per company
Questions, answers, agent instructions and reports stored unreadable with a key unique to your company.
Tamper-proof audit log
Every opening, export, public link and support access is recorded, exportable as CSV.
AI without retention
No training on your data, plus optional masking of e-mails, phones, IBANs and cards.
Purge with certificate
Your administrator deletes everything at any time and gets a dated certificate.
The publisher
Published by Pivotal Skills AI, the enterprise generative AI specialist
Pivotal Bench is designed and published by Pivotal Skills AI, which helps companies deploy generative AI. The platform was born in the field: measuring, understanding and governing the AI agents our clients put into production.
Discover Pivotal Skills AISister platform
Digital Skills Analyzer: measure adoption, skills and ROI of your agents
Pivotal Bench evaluates the quality of your agents. Digital Skills Analyzer, Pivotal Skills AI's other platform, measures what your teams actually do with them — to steer change management at scale.
- 01
Usage and skills
Real user usage and generative AI skill levels.
- 02
Agent adoption
By population, group, department, subsidiary or business unit.
- 03
Costs and ROI per agent
Costs in credits, tokens or euros, and the return on investment of each agent.
- 04
Recommendations
Concrete recommendations to drive change at scale.
Frequently asked questions
What is Pivotal Bench?
+
Pivotal Bench is a platform that measures the quality, speed and cost of AI agents, inspects their configuration (instructions, tools, skills, sub-agents) and helps govern them across a company.
Which platforms are supported?
+
Dust is the first connector. ChatGPT, Claude and MCP are planned. Results from any other system (Glean, Copilot…) can already be imported as CSV and compared.
How are answers judged?
+
An evaluation model compares the meaning of each answer with the expected answer and returns a justified verdict. Partial or Incorrect verdicts get a second look against the question, and you can override any verdict by hand.
What is the Craft.ecQ score?
+
Craft.ecQ is an 8-criteria prompt-writing framework: Context, Role, Actions, Format, Tone, Examples, Constraints and Quality control. Pivotal Bench rates each agent's instructions from 1 to 5 and explains every criterion.
Who can see the results?
+
Everything belongs to your company, recognised by its e-mail domain. Admins manage members; each space is shared as read-only or editable.
Your AI agents deserve serious measurement
Connect your Dust workspace and get a first inspection and a first benchmark within minutes.
