Every team that embeds an assistant inside its own product eventually hits the same review. Someone asks a question the system should not answer, and it answers anyway — fluently, confidently, and from the wrong scope.
That failure is rarely a model problem. It is a design decision, and it comes from shipping one outcome where the business actually needs three.
The three outcomes an answer can have
This section is analysis.
When a user asks an application's assistant a question, exactly one of three things is true. The system found permitted content that answers it. The system found the content, but this user is not cleared to see it. Or the system found nothing that supports an answer at all.
Those are three different facts with three different owners. The first is a success. The second is an access-control event, and someone in security or compliance needs to know it happened. The third is a content gap, and someone in operations needs to know what was asked.
Most implementations collapse all three into a single output: plausible text. That is the design decision worth reopening, because a system that cannot distinguish "I don't know" from "you may not know" cannot be audited, cannot be tuned, and cannot be defended in a review.
There is a second-order cost that rarely appears in the business case. A user told "no results" when the content exists but is restricted will file a ticket. That ticket will be resolved by widening access, because the fastest fix for an unanswerable question is permission. The interface design quietly becomes an entitlement design.
What the published guidance actually covers
NIST's AI Risk Management Framework is organized around four functions: Govern, Map, Measure and Manage. They are not sequential phases. Govern runs across the other three, and it covers exactly the questions above — who is accountable for a system's outputs, what limits apply, and how those limits are evidenced rather than asserted.
The framework is voluntary guidance rather than a certification. It will not tell you what to build. It will tell you which questions a reviewer will eventually ask about what you built.
OWASP's Top 10 for Large Language Model Applications is more concrete about failure modes. Prompt injection sits at the top of that list, and excessive agency appears as an entry in its own right — the risk of granting a component permissions beyond what its task requires. Both describe the same underlying property. An answer is not a control, and the boundary has to be enforced somewhere other than the prompt.
CISA's Secure by Design initiative pushes the obligation upstream, stating that "Every technology provider must take ownership at the executive level to ensure their products are secure by design." Its premise is that the burden has been "placed disproportionately on the shoulders of consumers and small organizations and away from the producers of the technology."
For a team embedding AI in its own product, that is a useful mirror rather than a criticism of anyone else. You are now a producer, and the customer downstream of your assistant is downstream of you.
Three places embedded intelligence can live
This section is analysis. None of the three is a vendor recommendation.
A vendor's API inside your application. Your product calls an external assistant service, passes the user's question and the permitted context, and renders what returns. Fastest to ship, and the integration surface is small. The trade-off is that scope, logging and identity have to be re-established on your side, because the vendor cannot know who your user is or what they may see.
Your own retrieval layer, with a model at the end. You control what enters the context window. This is where the three outcomes become possible, because you write the code that decides whether a chunk was permitted, absent, or refused. The cost is that you have now built a retrieval system, and retrieval systems drift as content, permissions and schemas change.
A third-party assistant bolted onto a portal. Common in enterprise software, and the one that ages worst. It sees what the signed-in user sees, which is correct, and it records what the platform chose to record, which is usually not enough to reconstruct why an answer was given.
The choosing question is not which is smarter. It is which one makes the three outcomes visible to the person who has to defend the system in eighteen months.
Where the review queue actually grows
Code quality is the part of this story that gets underestimated, and it is a capacity problem before it is a quality problem. When generation gets cheap, the volume of generated change rises faster than reviewer headcount. The scarce resource stops being the ability to produce a change and becomes the ability to prove one is correct.
NIST's Secure Software Development Framework, SP 800-218, is guidance rather than a certification, and its premise is that security practices belong inside the development lifecycle instead of in a review at the end. An embedded assistant is now part of that lifecycle, which means its prompts, retrieval rules and permission checks are artifacts that need version control and review like any other code.
The NIST Cybersecurity Framework, in its 2.0 form, organizes outcomes around six functions: Govern, Identify, Protect, Detect, Respond and Recover. The observation that the last two receive the least rehearsal is this publication's analysis, not a finding in the framework.
The practical version: if nobody can explain, from logs alone, why the assistant returned a specific answer to a specific person last Tuesday, the system is not governed. It is merely running.
What JWA's own product description covers
Bordy is described by its publisher, JWilliams & Associates, as offering chat, optional in-depth research, selective personal memory, conversation transfer, shared rooms and channels, and image analysis at heybordy.com.
For application work, bordycode.com distinguishes two modes: Embedded Intelligence for application research, context and memory, and Developer Intelligence for primary coding, QA companionship and hybrid orchestration. Developer capabilities, compatible tools, repository access, build execution and entitlements are confirmed per integration rather than assumed from a feature list.
One detail from this publication's own operation is worth naming, because it is a working configuration rather than a market claim. NewsForge, the system that produces this edition, uses Bordy's scoped structured-output research API with durable queued requests, and NewsForge retains approval and publication authority.
That last clause carries the design point. The model returns structured results. A separate system decides what gets published, and it is named in the record when it does.
Four questions for the next integration review
- Which of the three outcomes does the interface show? Not in the logs — on the screen. A user told "no results" when the content is merely restricted will open a ticket that gets closed by widening access.
- Where is scope enforced? If the answer is "in the prompt," it is not enforced. Prompts are inputs, and inputs can be argued with.
- What does the record contain? The question, the permitted sources consulted, the sources excluded and why, the model and settings used, and the person who received the answer. Reconstructable without interviewing anyone present.
- Who owns an unanswerable question? Somebody has to read the list of questions the system could not answer. That list is the most useful content-gap report the organization will produce, and most teams discard it.
A 30-day check
- Days 1 to 5. Ask the embedded assistant twenty questions whose answers you already know, including five the current user is not cleared to see. Record exactly what it says for each.
- Days 6 to 10. For any answer drawn from a restricted source, trace how that content reached the model. That path is your real scope boundary, whether or not anyone designed it.
- Days 11 to 15. Pick three answers and reconstruct them from logs alone. Where the reconstruction fails, note which artifact was missing and who should have produced it.
- Days 16 to 22. List every place a permission decision is made: application code, retrieval filter, prompt, vendor platform. Count how many of those places have a test.
- Days 23 to 30. Write the escalation rule. What the assistant does when evidence is insufficient, who is notified when it is not, and what the user is told in each case.
What remains unestablished
- Measured accuracy of any grounded-answer system. No benchmark result was reviewed for this edition, and none is asserted.
- Whether exposing three outcomes reduces support volume or increases it. No measurement was reviewed in either direction.
- What any specific vendor's retrieval or scope enforcement actually does in production. Only published descriptions were reviewed, and those describe intent.
- Any customer outcome, cost saving, conversion change or headcount effect from embedding AI in a business application. None is asserted.
- Whether the patterns described here transfer to a specific industry's regulatory obligations. That requires legal review this edition did not perform.
- Market data. No stock, index or valuation figures were reviewed for this edition.
Evidence notes and limits
NIST's AI Risk Management Framework is cited for its organizing functions only — Govern, Map, Measure and Manage — and for the proposition that it is voluntary guidance rather than a certification. It is not quoted, no specific control is asserted, and no conformity determination is drawn from it.
OWASP's Top 10 for Large Language Model Applications is cited for two entries only: that prompt injection is listed first, and that excessive agency appears as a separate risk concerning permissions granted beyond a task's requirement. No quotation is reproduced and no scoring or assessment is asserted.
CISA's Secure by Design initiative is cited for two quoted statements: that "Every technology provider must take ownership at the executive level to ensure their products are secure by design," and that the cybersecurity burden has been "placed disproportionately on the shoulders of consumers and small organizations and away from the producers of the technology." It is voluntary guidance, and no certification or measured outcome is asserted from it.
NIST Special Publication 800-218, the Secure Software Development Framework, is cited in paraphrase for the proposition that security practices belong inside the development lifecycle rather than in a terminal review. No practice identifier is asserted and no conformity claim is made.
The NIST Cybersecurity Framework, in its 2.0 form, is cited for its six named functions and nothing else.
Bordy's capabilities and limits are described only per JWilliams & Associates' published product descriptions at heybordy.com and bordycode.com, and are labeled as the publisher's own description rather than independently verified performance. The description of how NewsForge uses Bordy's scoped structured-output research API with durable queued requests, and retains approval and publication authority, is JWA's own account of its own configuration and is not offered as a claim about what any other organization should build or will achieve.
No benchmark, audit, contract, deployment record or third-party test was reviewed. The three-outcome framing, the three placements, the four questions, the 30-day check and the reading of the review queue as a capacity constraint are this publication's analysis. Nothing here establishes that any organization reduced risk, improved accuracy or prevented an incident.
The advertisement above is separate from this analysis and is not a source for any claim in it.