AI and delegated authority glossary

Software that pursues a goal across multiple steps, deciding what to do next and using tools along the way, rather than answering a single question. Commonly described as a model plus memory, planning and tool use. The test: what does it do when nobody is asking it anything?

Lilian Weng, LLM Powered Autonomous Agents, Lil'Log, 23 June 2023

See also Agentic AI, Harness, Tool use, or function calling

Marketing an assistant, chatbot or scripted automation as an autonomous agent. Gartner estimated only around 130 of the thousands of vendors claiming agentic AI were the real thing.

Gartner, Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027, 25 June 2025

See also Agent, Agentic AI

AI built to act toward goals across several steps rather than respond to single prompts. The label alone tells you nothing about capability.

See also Agent, Agent washing

A system that runs the operational work of a business, reading submissions, checking appetite, drafting responses and building bordereaux, and delivers finished work with an evidence trail, rather than helping a person with one task.

See also Decision trail, Extraction

Under the EU AI Act, a machine based system operating with some degree of autonomy that infers from its inputs how to generate outputs such as predictions, content, recommendations or decisions. Rule based software generally falls outside this definition.

Regulation (EU) 2024/1689 (the AI Act), Article 3(1); European Commission, Guidelines on the definition of an artificial intelligence system, 6 February 2025

See also Machine learning, Generative AI

The share of bound risks that sit inside the underwriting appetite agreed with your capacity provider, measured per underwriter and per program.

See also Binder alignment, Live book

The record of every action taken in a system, human and automated, retained so a single decision can be reconstructed later. The practical test: how long does it take to answer “why did we write this risk?”

See also Decision trail, Least privilege

How closely bound business conforms to the binding authority agreement, scored at the point of bind rather than sampled afterwards.

See also Appetite adherence, Delegated authority

A periodic report from a coverholder or MGA to its carrier, listing risks, premiums or claims written under a binding authority. Traditionally monthly or quarterly. Traditionally late.

InsTech and distriBind, Making delegated authority data exchange simple

See also Live book, Coverholder

The processing power consumed when a model runs, and the reason nothing here is free. Relevant whenever a vendor offers unlimited usage of something that has a real unit cost.

See also GPU, Token, Inference

A system's own estimate of how reliable a particular output is. Useful only when it drives behavior, meaning low confidence produces a flag rather than a guess.

See also Hallucination, Grounding

How much text a model can hold in mind at once. Large now, but finite, and exceeding it drops earlier material, usually silently. A 400 page schedule plus a binder plus your appetite statement may not fit at once; how a system handles that has real consequences for accuracy.

See also Token, Retrieval augmented generation (RAG)

A company granted delegated underwriting authority by a Lloyd's managing agent under a binding authority agreement, able to bind risks on behalf of the syndicate within agreed limits.

See also Delegated authority, Bordereau

Where your data physically sits and is processed. A contract question with a physical answer: which country, which provider, and whether you get to choose.

See also Training data, Least privilege

The full record of inputs, rule checks, human actions and machine actions behind a single underwriting outcome, retrievable for one risk on request.

See also Audit trail, Grounding

An arrangement where an insurer or managing agent delegates underwriting, and sometimes claims handling, to a third party such as an MGA or coverholder.

See also Coverholder, Binder alignment

Turning unstructured content into structured fields. A scanned ACORD form becomes named fields in your policy administration system.

See also Extraction completeness, Structured output, Grounding

How much of a submission file was captured, structured and made searchable. Different from accuracy, which is about whether what was captured is correct. Ask about both.

See also Extraction, Submission capture rate

Additional training on specific examples to change how a model behaves. Changes style and format; does not teach it facts about your book. Fine tuning could teach your house style for quote letters. It will not teach it what is in this month's submissions. For facts, use grounding.

See also Grounding, Training data, Retrieval augmented generation (RAG)

The largest and most capable models available at any given moment, currently from a handful of providers. Training one costs hundreds of millions of dollars, which is why nobody in insurance is building one.

See also Model, Parameters, Large language model (LLM)

AI that produces new content, including text, structured data, code and images, rather than only classifying or scoring what already exists.

See also Machine learning, Large language model (LLM), Agentic AI

The specialized chip most AI processing runs on. You will never buy one. It explains why every question has a cost rather than being free once the software is installed.

See also Compute, Inference

Tying output to a specific source document so the model works from your material rather than its training. In practice: every extracted field links back to the page it came from. Grounded output is checkable in seconds; ungrounded output has to be checked from scratch.

See also Hallucination, Retrieval augmented generation (RAG), Decision trail

Constraints on what a system may do or say, whether technical, policy or contractual. Never bind without approval. Never quote outside appetite without a referral.

See also Least privilege, Human in the loop

Fluent, confident, false output. A property of prediction rather than a defect. A summary giving an attachment point of $5 million when the slip says $2 million, formatted so correctly that nobody checks.

See also Grounding, Confidence score, Retrieval augmented generation (RAG)

The software around a model that assembles context, calls tools, enforces limits and maintains state. Most of the difference between two products lives here rather than in the model.

See also Agent, Model, Tool use, or function calling

A person approves before an action takes effect. Appropriate for anything that binds, prices or commits the firm.

See also Human on the loop, Guardrails, Least privilege

A person supervises and can intervene, but the action does not wait for approval. Appropriate for chasing a missing loss run. KPMG treats escalation between the two as non negotiable.

KPMG, Agentic AI untangled: navigating the build, buy, or borrow decision, January 2026

See also Human in the loop, Guardrails

A single run of a model, taking input and producing output. Usually the unit you are billed for.

See also Token, Compute

A model trained on very large quantities of text to predict what text comes next. The technology underneath ChatGPT, Claude, Gemini and most current AI products.

See also Model, Frontier model, Parameters

The principle that any actor, human or automated, holds only the permissions its task requires. In practice it means an automated action never exceeds the authority of the person it runs for.

See also Guardrails, Audit trail, Human in the loop

A continuously current view of portfolio composition, appetite adherence and binder alignment, as opposed to a bordereau describing a position that has already changed.

See also Bordereau, Appetite adherence, Binder alignment

Systems that learn patterns from data rather than following written rules. Predictive models used in pricing and reserving are machine learning, and predate generative AI by decades.

See also Generative AI, Model, AI system

The trained system that does the reasoning. Most insurance AI products are software built around a third party model. When a vendor says “our AI”, ask whose model is underneath and what happens if that provider has an outage.

See also Harness, Frontier model, Large language model (LLM)

A model that handles more than text, typically images too. Relevant because so much of your inbound arrives as photographs and scans.

See also Extraction, Model

The numbers inside a model that hold what it learned during training. Frontier models have hundreds of billions of them. The “large” in large language model.

See also Large language model (LLM), Frontier model, Training data

The instruction given to a model. In a purchased product, usually written by the vendor and invisible to you.

See also Model, Harness

Searching a document set for relevant passages and handing them to the model with the question, so the answer comes from your documents rather than training data. “What does our binder with this carrier say about vacant property?” With RAG the system finds the clause and answers from it. Without it, the model invents a reasonable sounding clause.

See also Grounding, Hallucination, Context window

Output produced in a fixed format, such as named fields, rather than free prose. What allows AI output to flow into a policy administration system.

See also Extraction, Tool use, or function calling

The share of inbound submissions that get logged and triaged. The gap between that and 100% is business you never knew you turned away.

See also Extraction completeness, Time to measured outcome

How long between deployment starting and a baselined metric actually moving. The honest version of the go live date most vendors quote.

See also Submission capture rate

The unit of text a model processes, roughly a word fragment. The basis of most usage pricing. If you are billed per token, a heavy schedule season costs more than a light one.

See also Context window, Compute, Inference

A model's ability to act beyond producing text: query a database, send an email, run a calculation. The difference between a system that drafts a quote for you to copy out and one that puts the quote in your system.

See also Agent, Harness, Least privilege

The material a model learned from. Distinct from the data you send it in use, and worth keeping distinct in any contract you sign.

See also Fine tuning, Data residency, Parameters