An agent reads the records your team creates. If those are thin, it will be confidently wrong.

LitifyAI is the vendor's own set of purpose-built legal capabilities — named agents for intake qualification, conflict checking, matter summaries, total policy, case value and strengths and weaknesses, plus a Damages Assistant, Instant Demands, a Document Assistant, AI time capture and AI invoice review. Litify ACE, its agentic case expert, extends this toward driving work forward rather than only reporting. Agentforce is Salesforce's general agentic layer, available to any org: you define topics, actions and grounding data yourself. Packaged agents are faster where one fits your job; Agentforce is the answer where nothing fits. Most firms end up using both.

Usually not in the same phase, and this is one of the few things we are firm about. During an implementation the data is at its thinnest and the team is at its busiest, so agents configured then read records that do not yet reflect how the firm will actually work. We assess readiness during discovery and normally sequence agents into the phase after go-live, once the case types are settled, stages are being updated reliably and the documents are where they should be. The data work that makes agents viable improves reporting and operations regardless.

That is what a readiness assessment measures, and it is measured per capability rather than as a single score — intake qualification needs very different data from matter summarisation. We look at field completeness on exactly what each agent reads, case type and stage consistency across the record set, document coverage and classification quality, and activity capture rates by role. The output is a viable-now list, a not-yet list, and a specific description of what would have to change to move an agent from the second list to the first.

The org's sharing model, if it is designed to. Agents operate within the access their configuration grants, so this is a confidentiality question before it is an AI question — an agent that can read every matter in the firm is a problem regardless of how well it performs. We design agent access explicitly against the sharing model, including ethical walls and matter isolation, and we enumerate the actions an agent may never take. This is part of the guardrail phase, before any agent is enabled, rather than something checked afterwards.

A human review step catches it, which is why we place one before any consequence and name an accountable reviewer rather than leaving it to whoever is nearest. During a pilot every output is reviewed; at scale that becomes sampling, with monitoring on override rates and output quality over time. Every agent action is logged with what it saw, what it produced and who approved it — which is what you need when a client or a regulator asks how a conclusion was reached. Every agent also has a documented switch-off procedure and a named owner.

Only if a baseline is measured before switch-on, which is why we insist on it. "It feels faster" is not a result and will not survive a partners' meeting. We agree a measurable claim and an exit criterion in advance — time to qualify an intake, time to draft a demand, override rate on suggested time entries — and measure against the baseline at the end of the pilot. If the criterion is not met we switch the agent off and tell you why. That is a successful pilot: it cost weeks rather than a firm-wide rollout.

Sometimes, and the honest answer depends more on your data than on your size. The capabilities with the best value-to-risk ratio for most firms are narrow and checkable: intake qualification, conflict checking, time capture assistance and invoice review. They work because the remit is tight, the output is reviewable in seconds, and the data they need is usually already being captured. Matter summarisation and damages work need much richer records and are where firms most often enable too early. We will tell you which group you are in.