Lemma and AI
Agents retrieve and structure facts. Lemma decides with published rules. Explanations travel back so people hear the truth, not a guess.
Trustworthy AI is not one system doing everything. It is a clear split: agents gather, Lemma judges, agents speak.
01
Agents retrieve and structure facts from documents, systems, and users. They turn messy context into the data a Spec expects: quantities, statuses, dates, eligibility signals. That work is flexible. The decision that follows is not.
02
Lemma evaluates published rules deterministically. Same Spec and data always yield the same outcome. Prices, eligibility, and actions come from rules leadership can read and sign, not from a prompt that changes tomorrow.
03
Every Lemma result can carry an explanation tree: which rules applied, which exceptions fired, why a value or veto landed. Agents do not invent a rationale. They narrate the trace Lemma already produced.
04
Agents read those explanations and serve end users with truthful, actionable information: what was decided, what was missing, what to do next. The conversation is natural. The mandate behind it is fixed.
Chapter 01
Agents move at machine speed. They draft workflows, fill forms, and call tools before a human has finished reading the brief. That speed is the opportunity. It is also the risk: when the outcome lives only in opaque code or a shifting prompt, nobody can challenge it.
Leadership needs a mandate it can own. Eligibility, pricing, and care protocols are not improvisation. They are commitments. An agent that invents the rule while answering the question leaves the organisation without a signature, an audit trail, or a way to change policy without retraining the model.
The scarce resource is control. Opportunity only holds if systems and agents run on provable rules that people can read, version, and certify.
Chapter 02
Lemma is an open language for business rules. Teams write Specs that stakeholders can read and machines can evaluate with certainty. The same Spec and the same data always produce the same outcome. Evaluation is deterministic and stateless.
Rules stay in a form people wrote on purpose: principles, exceptions, vetoes. They are not buried in application code or diluted across prompts. When policy changes, you change the Spec. Agents keep gathering facts; they do not become the policy.
That separation is what makes AI usable in regulated and high-stakes work. Agents stay flexible where context is messy. Lemma stays strict where the answer must be the same every time.
Chapter 03
A bare number is not enough for a customer, a case worker, or a regulator. Lemma can return an explanation with every result: which rules applied, which exceptions won, what data was missing. That tree is the handshake between the decision engine and the agent.
The agent does not invent why a claim was denied or a rate was applied. It reads the explanation and translates it into language the end user can act on. If something is still missing, the agent asks for it. If a veto fired, the agent says so plainly.
Truthful service depends on that chain. Retrieve, decide, explain, serve. Break any link and trust collapses into plausible fiction.
Chapter 04
Lemma ships an MCP server so assistants can check, show, and evaluate Specs safely from
a local prefix. Run lemma mcp
beside your policies when you want process isolation and agent access without handing
over the decision itself.
On LemmaBase.com, hosted MCP lets assistants work with published rules: list, search, show, run, and open motions when you grant write access. See LemmaBase MCP for connection details and the tool catalog.
In both cases the pattern is the same. The agent supplies structured data. Lemma returns values and explanations. The agent serves the person on the other side.
If your agents already gather facts and talk to users, put the mandate in Lemma. The conversation can stay flexible. The decision stays accountable.