The questions, with strong and weak answers
What business problem does your AI solve for a legal department, and how would I measure it?
Strong answer
A specific operational outcome: faster turnaround on requests, lower outside counsel spend, more questions answered without a lawyer, better visibility for leadership. They can show the report.
Weak answer
"It makes lawyers more productive," with no number attached.
LegalOperator's answer
AI assigns a UTBMS code to each invoice line when the firm's bill has none, groups the lines into workstreams and checks every line against your guidelines. A person on your team approves, holds or returns each invoice. How E-billing works.
Where does my team use this, and which of the systems I already own does it connect to?
Strong answer
It works inside email, chat, and the office tools people already use, and they name specific systems (document stores, contract repositories, ticketing, finance) that connect through configuration rather than a services project.
Weak answer
A tour of their web app and a promise about integrations.
LegalOperator's answer
Iris, the AI agent for the legal team, works in the web app, Slack, Teams, Zoom Chat and email, and inside Claude, ChatGPT and Microsoft Copilot through the MCP Connector. LegalOperator integrates two-way with Jira, Ironclad, Agiloft, TeamConnect, CounselLink and Onit Unity. Every channel, source and export.
Which AI models run in your product, what does each one do, and who decides when to change them?
Strong answer
More than one model, each matched to a job (a fast one to sort and route, a strong one to reason over a document), and a named person who tests new models against real tasks before they go live.
Weak answer
One model name, or "we use whatever the provider recommends."
LegalOperator's answer
More than one kind of model, each matched to its job: frontier large language models to reason over documents and draft, decision models that sort and route each message and report how confident they are, embedding models for search, and open-source models alongside commercial ones.
The last time you upgraded a model, what changed for your users?
Strong answer
Nothing to relearn. Answers got better, screens and workflows stayed the same, and they can name the month and what improved.
Weak answer
A release with new buttons, or they cannot remember.
LegalOperator's answer
You can test a model change yourself. An admin on your team pins real questions to an eval set and reruns it at any time, including after a model change.
Where do the answers come from, and what happens when the AI does not know?
Strong answer
Answers come from your approved content and your own records, each one cites its source, and when the answer is not there the AI says so and hands the question to a person with the context attached.
Weak answer
It uses the model's general knowledge, or it always answers.
LegalOperator's answer
Iris answers from your company's own approved content, with citations, and not from the open internet. Matter research links every answer to its source documents and says "nothing found" when your matters do not cover the question. How matter research cites sources.
Can it complete work, and can I see everything it did?
Strong answer
It can take actions inside the system, shows a preview, waits for a person to confirm, and every AI call is logged with what it read, what it answered, which model ran, and what it cost.
Weak answer
It drafts text and people do the rest by hand, or logging is a roadmap item.
LegalOperator's answer
Every change Iris proposes is a preview that a person on your team confirms. Nothing changes until they do. What Iris does and how your team confirms.
Is my data used to train any model, and what security proof do you hold today?
Strong answer
No, it is in the contract and backed by their terms with each model provider. SOC 2 Type II or ISO 27001 in hand, a penetration test within the year, or a dated plan if they are early.
Weak answer
"We take security very seriously."
LegalOperator's answer
Customer data is not used to train AI models, by LegalOperator or by the AI providers it uses; those providers are bound by contract not to train on it. SOC 2 Type II audit in observation, report expected November 2026. A third-party penetration test was completed in July 2026 and all findings were remediated; the summary is available under NDA. Security and data handling in full.
Ask two more questions
The seven are about the AI. These two are about the deal, and they apply to any vendor.
What are the contract terms, how do I leave, and can the price change without my say?
A strong answer is short terms the vendor can point to in writing, a stated way to leave, a published price, and a clear no to automatic price increases. A weak answer is terms you only learn after a sales call. LegalOperator's answer: no lock-in, and no automatic price increases. Prices are on the pricing page. Read the terms.
Can I try it on my own data before I buy?
A strong answer is a trial or proof of concept on your own documents or invoices, with a stated deliverable, a stated time to results and a stated date when your data is deleted. A weak answer is only a scripted demo on their sample data. LegalOperator's answer: the 3-Invoice Pilot. Send three recent invoices, LEDES or PDF. Within one business day they come back coded, grouped into workstreams and checked against your guidelines. No charge, nothing to install, nothing changes in your current tool. Pilot data is deleted from the app after 30 days. Backups expire on their normal schedule. Start the pilot.
If you want the thinking behind the questions, start with why legal teams stall on AI, and what to ask instead.