A government lawyer doesn’t want a summary of the statute. They want the statute. The exact text, the section number, the proviso clause, the word “shall” instead of “may.” Hand them a fluent paraphrase and you haven’t helped them — you’ve handed them a liability with a confident tone. This is the part most Copilot Studio projects get backwards, and it’s why so many AI agents in legal and policy contexts get piloted, distrusted, and quietly shelved.
A Lawyer Doesn’t Want a Summary. They Want the Source.
In most office contexts, summarization is the feature. You have forty pages, you want the gist, the model gives you the gist. Fine. But move that same behavior into legal, policy, records, or personnel work and the gist is exactly the wrong output. In law, the paraphrase is not the rule. The text is the rule. The difference between “within 30 days” and “within a reasonable time,” between a mandatory “shall” and a permissive “may,” between a definition that includes contractors and one that doesn’t — those distinctions are the entire job. A summarizer smooths them out by design.
So the right question for a citation-bound government agent isn’t “how good is the summary.” It’s “can the lawyer see the controlling text, verbatim, with a pinpoint citation, in one click — and does the agent refuse to answer when it can’t find it.” Get that right and you have something a general counsel will actually rely on. Get it wrong and you’ve built a very articulate way to give bad legal guidance at scale.
Copilot Studio’s Default Behavior Is the Problem
Here’s the uncomfortable part, and it’s straight out of Microsoft’s own documentation. When you add knowledge sources to a Copilot Studio agent and turn on generative answers, the agent retrieves relevant content from those sources, summarizes it, and generates citations where possible. Read those two qualifiers again. The default output is a summary. The citation is best-effort.
Both defaults are fine for an IT helpdesk bot. Both are disqualifying for a policy or personnel agent. “Summarizes” means the controlling language gets rewritten before it reaches the person who needs the exact wording. “Where possible” means that on some answers there’s no citation at all — and the ones that look cited aren’t always verifiable. There’s a well-documented trap where citations resolve perfectly in testing and then go non-clickable in production because the underlying content was ingested from a structured file the platform can’t turn into a resolvable reference. Your demo looks airtight. Your production agent points at nothing.
None of this is a knock on the platform. Copilot Studio is the right tool. It’s a knock on shipping the defaults into a context where the defaults are a legal exposure.
Summarization Is a Liability, Not a Feature
The danger isn’t that the model is usually wrong. It’s that it’s usually right, fluently, and then occasionally drops a qualifier or a “not” while sounding exactly as confident as it did on the correct answers. A summary that omits the exception clause reads as authoritative as one that includes it. The reader has no signal that this particular answer is the one that’s going to get them in trouble.
In law, the paraphrase isn’t the rule. The text is the rule.
Then there’s auditability, which is where legal work lives or dies. When an answer gets challenged — and in government it will — “the agent told me roughly this” is indefensible. “The agent returned Section 4.2(b) verbatim, here’s the link, here’s the document version” is a record you can stand behind. A summarizing agent destroys the audit trail at the exact moment you need it most. And reliance compounds the risk: the more useful the agent is on the easy questions, the more a busy lawyer trusts it on the hard one where the paraphrase quietly went sideways.
The Pattern: Citation-Bound, Zero-Summarization
The fix is an architectural decision, not a prompt tweak. You design the agent so the language model’s job is to find the right passage, not to rewrite it. The model is a router and a retriever. The controlling text passes through to the user unaltered, wrapped in a pinpoint citation and a document version. The model finds the passage; it does not get to rewrite it. That single constraint — retrieve, don’t summarize — is what separates an agent a general counsel will rely on from one they’ll politely kill after the pilot.
How You Actually Build It
The work is unglamorous, which is exactly why it gets skipped. You ingest sources in a form the platform can resolve to a clickable, verifiable reference — not a structured file that demos clean and dies in production. You constrain the agent to return the controlling passage verbatim, with section number and version, and you instruct it to refuse rather than improvise when retrieval comes back empty. “I can’t find controlling text for that” is a feature. A confident paraphrase of nothing is the failure mode you’re engineering against.
Then you test the refusal path harder than the happy path. Anyone can demo the question that retrieves cleanly. The agent earns trust on the question where the answer doesn’t exist in the corpus — and it says so instead of filling the silence. That’s the test a lawyer runs in their head before they rely on it, so it’s the test you run first.
Why GCC Makes This Non-Optional
In a commercial tenant, a hallucinated paraphrase is an embarrassment. In a regulated government environment, it’s a record — discoverable, auditable, and attached to a decision someone made in reliance on it. The audit trail isn’t a nice-to-have you bolt on later; it’s the deliverable. An agent built to operate within Microsoft’s FedRAMP-authorized GCC (Government Community Cloud) boundary, aligned to NIST 800-171 control objectives, with citation-bound output and full audit logging, is the only version that survives contact with a real agency. The commercial “good enough” playbook doesn’t transfer. It just fails more expensively.
Who Builds This
I do — directly. Puget Sound AI is a solo, veteran-owned small business (VOSB). When you hire me, you get the engineer who scopes the retrieval design, builds the agent, and hands your staff the documentation to run it. No account manager translating your requirements into a slide deck, then invoicing you for the slide deck.
If you’ve got a policy, personnel, or records use case where the paraphrase is a liability and the text is the rule, that’s the work I do. Let’s talk.