Run the same audit tasks through each AI model, score the results from 1 to 5 on five criteria, set the weights that matter to your firm and see the ranking. The scores come from your own testing.
Weights are relative and add to 100. They are used as shares of the total, so they do not need to sum to 100.
A weighted score is a summary, not a decision. Treat confidentiality as a gate rather than a weight: a model that cannot meet your client-data obligations should be out whatever it scores elsewhere. Use real, anonymised audit tasks, and run each model on identical inputs. Models change often, so re-test before relying on a result.
Public benchmarks say little about how a model handles a depreciation schedule, a TDS reconciliation or a bank confirmation. Build a small set of real tasks with known answers, remove client identifiers, and run every model on the same set. Score what you can verify, such as whether the figures tie and whether sources are named, and keep the written scores so the choice can be explained later.
You set a weight for each of five criteria and score each model from 1 to 5 on each. The weighted score is the sum of score times weight, divided by the total weight, so it stays on the 1 to 5 scale.
Models are ranked from highest to lowest. You can add or remove models and change the weights to see how sensitive the ranking is.
The scores should come from tests you ran on the same tasks for every model. The tool only does the arithmetic.
Model A scores 4, 3, 4, 3, 3. Model B scores 3, 4, 3, 5, 4. Total weight is 100.
