Twenty-four places searched. Twelve systems in the boundary. Every call yours.
The agent finds where CUI lives and where it flows, in files, email and AI tools, and draws the smallest boundary that holds it. You decide what goes inside.
The real world
- Award
A clause with a date
The solicitation says CMMC Level 2 by award. Someone is asked which systems hold CUI, and the answer is a guess.
- Drift
CUI goes where work goes
Drawings are exported to the file share, attached to email, sent home for the weekend, and pasted into a chatbot to check a tolerance. Most copies lose their markings on the way.
- Sprawl
The boundary becomes everything
Nobody can draw it with confidence, so every system goes in, and every one of the 110 requirements applies to every one of them.
- Spend
Controls bought before the map
Budget goes to tools for systems that never held CUI, while the shop-floor USB drives and the unencrypted backup tapes wait.
- Assessment
The assessor finds it first
A shared link, a personal inbox, a prompt log. The finding lands weeks before award, when there is no time left to fix it.
Watch it work
Loading…
What it’s worth
On the fictional company in the demonstration, an early trace keeps eight systems out of the assessment and puts a number on what waiting would cost: the contract that requires Level 2.
- $4.2M
- A year of contract value at risk if Level 2 is not in place by award
- $160K
- Of assessment work avoided: 8 of 20 systems kept out of the boundary
- $42K
- Of tracing hours saved: 280 hours of searching by hand
Compared
Data-loss prevention tools already search for markings, and compliance platforms track the 110 requirements. Here is what each does, and where an agent adds something new. In every column, people make the final call.
| What it does | By hand | Data-loss prevention tools | Compliance platforms | CUI Shield Agent |
|---|---|---|---|---|
| Finds CUI banners, portion markings, distribution statements and export-control phrases | Partlyinterviews and samples | Yespatterns you write | No | Yes |
| Recognizes CUI that has lost its marking, such as a drawing pasted into a chatbot | Partlyif someone knows | Nono marking, no match | No | Yesreads the context, gives a reason |
| Tells a leftover banner on a public brochure from real CUI | Yes | Noflags it | No | Yes |
| Reads AI-tool logs for CUI that reached a public service | Partly | Partlyif it sees that traffic | No | Yes |
| Draws the smallest boundary along the system inventory’s connections | Partlyusually everything goes in | No | Partlyyou enter it | Yesyour security lead approves |
| Maps each gap to NIST SP 800-171 with the fix | Yesslowly | No | Yesa checklist | Yesfrom the facts it found |
| Prices the contract value at risk | No | No | No | Yes |
| Drafts the boundary memo and the enclave design | Yesslowly | No | Partlytemplates | Yesa draft to approve |
| Blocks or encrypts CUI in transit | No | Yesits core job | No | Noproposes; your tools enforce |
| Signs the CMMC assessment | Yespeople | Yespeople | Yespeople | Yespeople |
● Yes◐ Partly— No
People’s hours, from contract clause to signed boundary
83% fewer hours than by hand, and 77% fewer than with a DLP tool and people: 422 hours, about $63K.
See the hours by activity, and how each is reached
| Activity | By hand | A DLP tool and people | Agent and people |
|---|---|---|---|
| Trace where CUI lives20 systems. By hand: 16 h each of interviews, share listings and mail searches. DLP: about 8 h each to write patterns and clear alerts. Agent: 2 h each to check its trace. | 320 h | 160 h | 40 h |
| Judge unmarked CUI and false alarmsDLP raises more alerts to clear than a person would look at, and cannot judge an unmarked drawing. Agent: a person confirms its verdicts. | 40 h | 60 h | 10 h |
| Draw the boundaryAgent: it proposes the smallest set that holds CUI; your security lead approves it. | 40 h | 40 h | 8 h |
| Map gaps to the 110 requirementsAgent: each gap comes from a fact it found and names the requirement and the fix. | 60 h | 60 h | 12 h |
| Write the boundary memo and planAgent: it drafts; a person edits and approves. | 40 h | 40 h | 8 h |
| Review and signThe same in every column. | 8 h | 8 h | 8 h |
| Total | 508 h | 368 h | 86 h |
Estimates for the fictional Corvane Aerostructures (20 systems) at an assumed $150 an hour, on the same assumptions as the business case above. Categories, not products: individual tools vary. Verify against your own environment.
How it works
- Trace
Every location, one pass
Repositories, file-share listings, email DLP exports and AI-tool logs, read in place. Plain patterns find CUI banners, portion markings, distribution statements, export-control phrases, contract numbers and contract drawings.
- Mask
Markings, never contents
Each finding carries its marker type, its location and a short masked snippet. Nothing marked reaches the agent or the report in clear text.
- Judge
What the markings can’t settle
The agent recognizes an unmarked drawing pasted into a chatbot as CUI, a leftover banner on a public brochure as a false alarm, and an invoice’s contract number as contract information rather than CUI. Each verdict shows its reason.
- Boundary
The smallest set that holds it
Plain code follows CUI along the system inventory’s connections, adds the systems that protect them, and leaves the rest out. Services that received CUI without authorization are marked to block, not to assess.
- Gaps
Mapped to NIST SP 800-171
Each gap comes from a fact: a public link, a personal address, a cloud service without FedRAMP Moderate, a USB transfer, an unencrypted backup. Each names the requirement and the fix.
- Memo
A boundary someone signs
What to fix first, what to do in 30 days, and the enclave design to approve, including which AI service may sit inside.
Where it runs
- In your environment
- Next to your file shares, mail gateway and repositories, with read-only access.
- Masked by default
- Markings and locations leave the scan, contents do not.
- Your inventory
- Systems, owners and connections from your records, corrected by your people.
- People decide
- The boundary, every verdict and every fix are proposals until your security lead approves them.
On your data
The demonstration above runs on a fictional company. On yours, the agent connects to your file shares, mail, repositories and AI-tool logs, inside your environment. Design partners can pilot it before their assessment.
The engagement
CUI Shield Audit™Boundary in 3 weeks; ready for assessment typically in 3–6 months
- 1 · Week 1
Connect
- Agents
- Installed beside your file shares, mail, repositories and AI-tool logs, read-only.
- Our people
- Deploy the agent, check the masking, and plant marked test documents to measure what it finds.
- Your people
- Grant access and name a CUI owner.
- 2 · Weeks 2–3
Trace and enclose
- Agents
- Trace CUI, propose the smallest boundary, map gaps to NIST SP 800-171.
- Our people
- Confirm every flow with your architects and design the enclave.
- Your people
- Approve the boundary.
- 3 · Months 1–5
Close
- Agents
- Re-trace as systems change, track every gap, draft SSP sections.
- Our people
- Run the POA&M and write the SSP with you.
- Your people
- Build the enclave and close the gaps.
- 4 · Before assessment
Prove and hand over
- Agents
- Assemble evidence for each requirement.
- Our people
- Rehearse the assessment, train your team, and watch the agent monthly.
- Your people
- Keep the map current.
What agents don’t speed up: Building the enclave, and the assessor’s (C3PAO) schedule.
What you keep
- CUI data-flow diagram
- Including AI tools and agents.
- Boundary decision record
- What is in scope, and why.
- SSP and POA&M
- Current and defensible.
- SPRS score
- Recalculated and ready to post.
- Accuracy record
- Every verdict and every override, logged. Planted-answer scores on every run.
Dozens of applications, more than two hundred stakeholders, and no map of where CUI went. A year later, a smaller boundary and a clean pre-assessment.Defense-sector contractor · CMMC Level 2 readinessRead the case study
Questions
Does the agent see our CUI?
No. Plain code finds the markings in your environment and masks the contents. The agent works from marker types, locations and masked snippets, and anything it needs beyond that stays with a person.
What about CUI with no marking?
That is the hard part, and where the agent earns its place. It reads the context (a title block, tolerances, a part on the contract) and proposes a verdict with its reason. A person confirms it.
Can our engineers use AI with CUI?
Only inside the boundary, on a service authorized for it, typically FedRAMP Moderate or equivalent. The agent shows where CUI already reached a public tool, so it can be blocked and handled as a spill.
How does it relate to the CUI Shield program?
It is the engine of the program’s trace. The program sets the boundary, the enclave and the plan with your team; the agent keeps the map current as work moves.