Buying a company
The seller says AI now does a large share of the work, which affects your offer price.
Model providers, AI vendors and data providers report their own handpicked numbers. We test your systems and data on your own questions, records and tasks, whether you bought them or built them. Then we work with you to improve them and find opportunities to increase your revenue.
A support chatbot solved 61% of 400 conversations, not the 78% its vendor reported
Reportedcounted as resolved by the vendor
Measuredproblem solved, by our evaluation
Differencecounted as resolved but not solved
Why 156 of 400 chatbot conversations failed
We check your AI system's answers and actions on real customer questions and tasks. Beyond the overall score, you receive a breakdown of where the system fails, the root cause of each failure and the unusual cases it gets wrong. Then we help you correct the main causes and run the evaluation again.
Guide to containment rate vs resolution rateCustomer records before and after cleanup
| Before | After | ||
|---|---|---|---|
| Acme Limited | 3 Jun 2026 | Acme Ltd | 3 Jun 2026 |
| Acme ltd. | 2026-06-03 | ||
| acme | June 3 | ||
| Bolt Co | blank | Bolt Co | missing |
We check customer records, knowledge base documents and data you buy, collect or pay others to label. We compare them with their sources and, for bought or labeled data, with your contract. We find the duplicates, gaps and errors that lead to wrong answers, then clean them up with your team and add what your knowledge base is missing.
Can you clean our data or build a knowledge base?How a grader scored 100 tasks, checked by experts
| Grader passed | Grader failed | |
|---|---|---|
| Experts: solved | 62 | 6 |
| Experts: not solved | 14 | 18 |
For teams that build or sell AI models, we check benchmarks and training environments (the practice tasks a model learns from) and count how often their graders pass a wrong answer or fail a correct one. Then we help you rewrite the weak tasks and rerun your scores.
How to check if a benchmark score can be trustedThe seller says AI now does a large share of the work, which affects your offer price.
A support chatbot, phone agent or other AI product did well in a trial, and you have to decide whether to sign the contract or roll it out to more teams.
Your support chatbot's dashboard says most conversations were resolved, and you want to know how many customers had their problem solved.
Your AI reviews insurance claims, helps with loan decisions or writes patient notes, and you need to show a regulator or auditor how you tested it.
You need to compare models on your own tasks, including what each correct answer costs, or to show that a new version of your model is better than the old one.
Your system works with audio or video, sometimes while a customer waits, so how fast it responds matters as much as whether it is right.
Evaluation
Improvement
When an acquirer, investor or auditor hires us to evaluate another company's AI, our work ends after the evaluation.
We recommend a change only when the evaluation results support it, so every recommendation comes with the evidence behind it. A few clients have asked us to build tools for them, and we keep that work separate from our evaluations. When we compare vendors for you, we never include our own products.
You receive the graded examples, the scoring rules and the code that runs the evaluation, so you can rerun it after each vendor update or on any model you consider. We can also set up a dashboard or connect the evaluation to your systems so you can monitor the results over time.
How a rerun catches a drop after a vendor update
Accuracy and cost of five AI models on the same tasks
If you need it, we can run every analysis on your own servers or in your cloud account, so no data leaves your systems. We follow GDPR and your rules for data retention and deletion.
We believe AI will soon do a large share of the world's work, and that it should be tested as carefully as the bridges, medicines and aircraft we already trust.
Our founding team has built and evaluated real-time AI systems used by millions of people. We've worked with Fortune 500 companies and published benchmarks. We've also built training and evaluation environments for large language models such as Meta's Llama and deployed fast private models that run on the device itself instead of in the cloud.
We've trained and deployed models across North America, Europe and Asia under several sets of rules for handling, retaining, cleaning and using data.
We've seen models with strong benchmark scores and vendor claims still fail real users. Now we bring that experience to companies that depend on AI systems and data.
We're a small team, still in stealth, and we're taking on a few clients. You can hire us directly or through your diligence or audit firm. Tell us what you're working on, and we'll reply within 24 hours.
Contact us