AI Activity Map
Map your legal work to the AI that should run it
A map of the activities your work consists of, each one coloured by the AI that fits: a frontier model in the cloud, a cheap fast one, your own servers, or a small model on the device. Four tiers, one rule, and the rule is printed under every map so you can disagree with it.
The three maps
Law firmLawyers and staff at a firm21 activities: 1 cloud frontier, 6 cloud lite, 6 on-prem, 8 localIn-house legalCounsel inside a company28 activities: 1 cloud frontier, 9 cloud lite, 8 on-prem, 9 local, 1 not yetGovernmentLegal, policy and CIO staff in a department21 activities: 0 cloud frontier, 5 cloud lite, 11 on-prem, 5 local
The four tiers
- Cloud frontier
- A frontier model over an API. Usage-based, the highest price per token, unpredictable at volume. Leaves your domain. The most capable: reasoning, long context.
- Cloud lite
- A fast, cheap SOTA-class model over an API. Usage-based, cheap per token, still unpredictable at volume. Leaves your domain. Skilled, not frontier.
- On-prem
- Open models on your own GPUs. Fixed: expensive to stand up, flat to run. Stays in your domain. Open models, which lag the frontier.
- Local
- A small open model on a laptop, workstation or appliance. No marginal cost. Never leaves the device. Routine work only.
The maps are drafted from the tokenomics of legal work and from how the three practices actually run. Every verdict comes from one four-line rule, and the table under each map shows the facts it was applied to. Where a row reads wrong to you, that is a fact to correct, not a colour to argue with.