Opens in a new tab
Back

How do you run a data loss risk assessment and find out where sensitive data is actually going?

How to run a two-to-four-week data loss risk assessment: observe where sensitive data really goes, score it, and report to the board.

Table of Contents

A woman in a jazz club sound booth ticks cells on a hand-ruled ledger grid under a brass lamp, beside glowing amber VU meters, as a trio plays behind the glass.

A data loss risk assessment shows where your sensitive data can leave, where it actually leaves, and which of those exits you can’t see or stop. At a 200 to 2,000-person company it fits in two to four weeks. Name five to ten crown-jewel data types, each with a business owner. List the exits: web uploads, personal webmail and cloud, AI prompts and uploads, USB, print, screenshots, sharing links, sync clients and email. Then watch what really moves for a sample period, because questionnaires miss the busy exits. In Verizon’s 2026 DBIR, 32% of non-malicious insider DLP events went to file-sharing sites, 20% to personal webmail and 12% to shadow AI. Map each exit against today’s controls. Score likelihood from observed events and impact from record counts and regulation, then write a risk register with owners and treatments. Finish with a one-page board summary and the triggers for the next assessment. Tenant audit logs and proxies show sharing links and destinations, but they miss what happens on the laptop: pastes, uploads to personal accounts, desktop AI apps, screenshots and anything done off the network. An endpoint agent closes that gap. On managed Windows and macOS laptops, Jazz’s endpoint agent sees those actions at the operating-system level without a browser extension and tells a personal account from a corporate one, and Melody investigates each risky movement, so register rows start as explained events.

Updated October 23, 2026.

Which tools can see where sensitive data leaves, and what each one misses

Approach What it shows for this assessment What it misses
Questionnaires and owner interviews Which data matters, who owns it and which workflows are legitimate The busy exits nobody mentions
Microsoft 365 or Google Workspace audit logs Downloads, sharing links and external shares, with history you can backfill Everything after the file reaches the laptop
Proxy, secure web gateway and DNS logs Which destinations get traffic, including AI services and personal webmail Content without TLS inspection, a personal account on the same domain as the corporate one, off-network laptops, USB, print, pastes and screenshots
Browser extension or managed browser Pastes and uploads inside that browser Desktop AI apps, other browsers, USB and print
Endpoint agent on managed laptops (Jazz) The action itself, on or off the network: paste, upload, screenshot, AI prompt, personal or corporate account. Jazz sees it at the operating-system level in any browser or desktop app, and Melody investigates each risky movement Linux machines, phones and unmanaged devices
DSPM or API scanning of SaaS and cloud storage Sensitive data sitting at rest Movement, which is what this assessment measures

What a data loss risk assessment covers, and how it differs from a cybersecurity risk assessment

A cybersecurity risk assessment looks at every threat to every system. NIST SP 800-30 Rev. 1 (September 2012), the method most regulators point to, runs it in four steps: prepare, conduct, communicate and maintain. Ransomware, unpatched servers and phishing all belong there.

A data loss risk assessment uses the same method on one outcome: sensitive data reaching a person, account or service that shouldn’t have it. Most of that movement is done by employees and contractors with legitimate access, usually to get work done, plus AI tools and the occasional departing employee who means it. Keep vulnerability scans and patch status in the broader assessment and reference it. A PII risk assessment is this method limited to personal data. A DLP assessment is the same exercise run before you buy or replace DLP; the Canadian Centre for Cyber Security’s ITSM.40.110 (April 2023) advises “extensive testing” before selecting a DLP solution.

NIST’s draft SP 800-154 (March 2016, still a draft) calls data-centric threat modelling “a form of risk assessment” and gives four steps that line up with this one: identify the data of interest, select the attack vectors, characterize the controls, analyze. It also warns that anyone with access to data could place it “in any accessible location, authorized or not.” That is why the exit list has to cover where data can go, beyond where it is supposed to live.

Two neighbouring jobs stay out of scope. Scanning cloud stores for data at rest is DSPM territory, covered in what is DSPM. Grading the maturity of an insider risk program is what CISA’s free Insider Risk Mitigation Program Evaluation does, and our insider risk assessment answer walks through it. Run the IRMPE alongside this assessment if you’re building that program; it doesn’t produce a likelihood and impact register.

The seven steps in two to four weeks, with a time budget for a team of one

The observation window sets the calendar, so start it on day one and do the desk work while data accumulates. Plan for about 35 to 50 hands-on hours across about three weeks, or four if your crown jewels include finance data, because a 14-day window can miss month-end close.

Step When Hands-on hours Output
Start the observation sources (log exports, endpoint agent, privacy notice) Day 1 5 to 8 Data collecting from every source you plan to use
1. Name 5 to 10 crown-jewel data types, each with a business owner Days 1 to 5 6 to 8 One row per data type: owner, systems, rough record count, regulation
2. List the exits each data type can take Days 3 to 5 3 to 4 Exit list per data type
3. Observe what actually moves Days 1 to 14 or 28 6 to 8 (about 30 minutes a day) Event log by data type, exit, account and person
4. Build the exit-by-control gap matrix Days 15 to 17 4 to 6 One row per exit: control, what it sees, what it stops, gap
5. Score likelihood and impact Days 17 to 19 4 to 6 Scored risks with the evidence behind each score
6. Write the risk register with owners and treatments Days 19 to 21 4 to 6, including owner sign-off Register signed by each owner
7. Write the one-page board summary and set re-assessment triggers Days 21 to 22 3 to 4 Board page and trigger list

Step 1 runs on interviews. Give each likely owner 30 minutes and ask what would hurt in a competitor’s or reporter’s hands, where it lives, roughly how many records exist, who handles it daily, and which exits they use legitimately. The owner should be the business leader who would answer for a leak, since ISO/IEC 27001 clause 6.1.2 asks you to “identify the risk owners.” If you need to find sensitive data without a full inventory, this answer covers interviews, targeted scans and watching real flows.

If you are the first security hire, this assessment is also how you decide where DLP sits among your priorities. Where DLP fits for a first security hire covers the order against MFA, EDR and backups. The register you produce here is the evidence for moving it up or down.

Which exits should a data loss risk assessment list?

List every route a crown-jewel data type could take out of your control, then mark which ones it plausibly uses. A practical starting list for a company of this size:

  • Web uploads to file-sharing and file-transfer sites
  • Personal webmail
  • Personal cloud storage, including sync clients that mirror a corporate folder to an unmanaged device
  • AI prompts, pastes, file uploads and AI browser extensions, on corporate and personal accounts
  • USB drives and other removable media
  • Print
  • Screenshots and screen sharing
  • Sharing links, especially “anyone with the link” and external shares
  • Corporate email to external recipients
  • Personal code repositories and AI coding tools, if you write software

Verizon’s 2026 DBIR, published 19 May 2026, shows why the personal-account exits come first. Across 4,280,149 non-malicious insider DLP events sent to untrusted destinations, the top targets were file-sharing sites (32%), personal webmail (20%) and shadow AI (12%) (Figure 64, p. 60). The same page reports that 45% of employees now use AI on corporate devices at least once every 15 days, up from 15% a year earlier, and that 67% of users reach AI services from corporate devices with non-corporate accounts. Source code was the largest data type in AI-bound events at 28% (Figure 65), and the average company had more than 15% of users with unauthorized AI browser extensions. The report’s example of a convenience leak is an employee who wants to work from home and emails company data to a personal account (p. 59).

Government guidance names the physical exits too. NIST’s SP 1800-28 (February 2024), on data confidentiality, includes a privilege-misuse scenario in which an insider copies files to a personal USB drive and prints them. ITSM.40.110 lists “physical drives or peripheral devices” and “cloud or web storage platforms” among its 13 exfiltration methods.

Which cyber security risk assessment tools show where data actually goes?

No single source sees every exit, so the assessment combines three you probably already have access to and records what each one missed. The table below gives the detail you’ll need when exporting from each one. Questionnaires tell you what people believe and which workflows are legitimate; use them in step 1, and use observed events for step 3.

Source What it shows during the window What it can’t show
Microsoft 365 unified audit log Downloads (FileDownloaded), full downloads through the OneDrive sync app (FileSyncDownloadedFull), anonymous links (AnonymousLinkCreated) and external shares, per user and site. Audit (Standard) keeps 180 days, so you can backfill history Anything that happens after the file reaches the laptop: pastes, uploads to personal accounts, USB writes. Endpoint events appear only if Microsoft’s separate endpoint DLP is deployed
Google Workspace Drive log events Views, downloads, prints, uploads, and visibility changes such as shared externally or public on the web. Kept for 6 months Downloads through Google Takeout, offline browser caches and Drive items emailed as Gmail attachments. Outside users show as anonymous unless explicitly shared. Nothing that happens inside a personal Google account
Proxy, secure web gateway and DNS logs Which destinations get traffic: AI services, file-sharing sites, personal webmail, and how much is uploaded Content, unless TLS is inspected. A personal account on the same domain as your corporate one, unless the proxy decrypts and reads the tenant. Laptops off the network. USB, print, pastes and screenshots never cross the proxy
Endpoint agent on managed laptops The action itself: paste, upload, screenshot, USB write, desktop AI app, personal or corporate account, in any browser, on or off the network Unmanaged devices, phones, operating systems the agent doesn’t support, and data sitting untouched in SaaS

The Microsoft figures come from Microsoft’s Purview auditing documentation (activity list updated October 2026, retention page updated June 2026). Audit (Premium) keeps Exchange, SharePoint, OneDrive and Entra records for a year, but only for E5-licensed users. CISA’s February 2024 announcement confirms the default moved from 90 to 180 days. The Google details come from Google’s Drive log events and data retention help pages (undated, accessed October 2026); Google says administrators cannot change the six-month retention, and most Drive events are logged only for files owned by users on supported editions.

Check DNS before you trust a destination list. The NSA’s guidance on encrypted DNS (January 2021) says that when browsers use DNS over HTTPS, the enterprise resolver is bypassed and “logging and visibility at the enterprise gateway is usually not possible.”

To measure how much sensitive data goes into AI apps, count every prompt, paste or upload that carries a crown-jewel data type to an AI destination, split by corporate or personal account, and divide by headcount to get events per 1,000 staff per week. Record which sources fed the count, since a network-only number misses desktop apps and off-network laptops. Our GenAI risk assessment goes deeper on the AI slice.

Tell staff before the window opens what is visible, why and who sees it. If any of them are in the EU, the Article 29 Working Party’s DPIA guidelines (WP248 rev.01, October 2017) list an employer systematically tracking staff workstation and internet activity as an example that needs a data protection impact assessment, so do it before you deploy an agent. Is it legal to monitor employees for data loss? covers notice by jurisdiction.

Where Jazz fits in the observation window, and where it stops

Jazz is the endpoint row of that table. Its agent sees copy and paste, screenshots, GenAI prompts, screen sharing and file uploads at the operating-system level, with no browser extension to install, so a second browser doesn’t create a gap, and it runs at under 1% CPU.

During the window you can ask Jazz a plain-language question such as “Where is customer PII going?” and get a map of the answer. Jazz’s Posture view surfaces risky user behaviours, workflows where sensitive data regularly leaves sanctioned systems, and unsanctioned destinations such as personal drives and external GenAI tools. Discovery counts often surprise security teams. One customer discovered 447 GenAI tools running across its organization, and in another environment where security had sanctioned a handful, the agent found 1,223.

The bigger saving comes in steps 4 to 6. Melody, Jazz’s agentic investigator, investigates each risky movement across the data, the systems, the people and the business context, so every candidate register row arrives as an investigated incident: what the data was, where it went, who moved it and whether that fits a sanctioned process. Your scoring week starts from explained events. In the worked example below, the four USB writes were IT staff imaging laptops and the file-transfer uploads were deliveries a customer had asked for. After the assessment, the same view becomes your re-measurement, the monthly posture cycle of picking the top issues, fixing the driver and watching the trend.

The limits belong in your coverage statement. Jazz runs on managed Windows and macOS devices only, so Linux machines, unmanaged or personal devices and phones are outside it. It does not scan data at rest inside SaaS tenants or cloud storage, which is DSPM or API-scanning work. Keep the tenant audit log in the assessment for sharing links, and record any device you can’t cover as an evidence gap on the register.

How do you score data loss risk, and what does a cyber risk score miss?

Score each risk on two five-level scales, write the scales down before you score, and keep the evidence next to each number. ISO/IEC 27001 clause 6.1.2 requires assessments that “produce consistent, valid and comparable results,” and you can only compare next year’s register with this one if the scales stay the same.

Base likelihood on how often the event happened in the window, annualized. NIST SP 800-30’s Table G-3 gives frequency anchors for non-adversarial events, which suits mostly accidental leaks: Very High is more than 100 times a year, High 10 to 100, Moderate 1 to 10, Low less than once a year but more than once every 10 years. Multiply the window’s count by 52 and divide by the number of weeks observed.

Score Likelihood (NIST SP 800-30 Table G-3 anchor) What that means in a 2-week window
5 Very high More than 100 a year 4 or more events
4 High 10 to 100 a year 1 to 3 events
3 Moderate 1 to 10 a year None seen, exit open and the data type present
2 Low Less than once a year None seen, exit partly controlled
1 Very low Less than once in 10 years None seen, exit closed and the closure verified

A two-week window can’t tell 1 event a year from 20; one event annualizes to 26. That is why zero observations fall back on judgment about whether the exit is open, and why a longer window sharpens the low end.

Base impact on what the data is and how much of it moved, in the event or plausibly in the next one.

Score Impact anchor (tailored for data loss)
5 Very high More than 100,000 regulated records, or IP whose loss would change the company’s competitive position
4 High 500 to 100,000 regulated records, exposed credentials, or source code and pricing useful to a competitor
3 Moderate Fewer than 500 regulated records, or confidential contracts, pipeline or board material
2 Low Internal-only data with limited harm
1 Very low Public or near-public data

The 500-record line matches HIPAA’s threshold for notifying HHS at the same time as patients under 45 CFR 164.408; adjust the bands for your regulators. NIST calls its own scales “exemplary” and expects tailoring.

Combine the two by multiplying (1 to 25) and band the result: 15 to 25 high, 8 to 14 moderate, 1 to 7 low. NIST’s Table I-2 combines them with a lookup matrix instead, and it is not symmetric: Low likelihood with High impact comes out Low. Either method is defensible; choose one and write it down. Break ties on impact, then on how many rows one fix would close.

Security ratings services build their “cyber risk score” largely from what they can observe outside your network, so the score can’t reflect an employee’s upload to a personal drive. If the board wants dollars, convert the top rows into expected loss with the method in how to make the business case for DLP, which uses observed counts as the frequency input.

A worked example: a data loss risk assessment for a 300-person company

Tallowmere Health Software is fictional, and so are its numbers. It has 300 staff and sells claims-processing software to US health plans, which makes it a HIPAA business associate, and its customers require a SOC 2 Type II report. The fleet is 270 managed Mac and Windows laptops, 22 Linux engineering workstations and personal phones. It runs Microsoft 365 E3 with no DLP. The observation window ran 14 days with Microsoft 365 audit logs, DNS filtering logs and an endpoint agent (Jazz) on the 270 laptops, which is 92% of company computers.

Step 1: crown-jewel data types and owners

Data type Business owner Where it lives Volume Regulation or contract
Member claims data (PHI), including support extracts VP Customer Operations Production database, support ticketing tool, CSV extracts on analysts’ laptops 2.4 million members across 11 health plans HIPAA, business associate agreements
Source code and deployment secrets CTO Git hosting, laptops, CI pipeline 40 repositories Trade secret, customer security terms
Health-plan contracts and pricing General Counsel Legal SharePoint site, CRM 11 customers, 60 open deals Contract confidentiality
Employee HR and payroll records Head of People HR system, payroll provider, HR SharePoint 300 current and 900 former staff, with SSNs State breach-notification laws
Board and financial materials CFO Board portal, finance SharePoint Fundraising round in progress Investor confidentiality
Customer list and pipeline CRO CRM 11 customers, 60 prospects Competitive harm

Steps 3 and 4: what moved, and the exit-by-control gap matrix

Counts cover crown-jewel data only, after each event was checked against who moved it and why.

Exit Control today Sees it? Stops it? Observed in 14 days Gap
File-sharing and transfer sites DNS category block on a few sites Domain only Partly 4 uploads of claims extracts, all to a customer’s own transfer portal at the customer’s request Low
Personal webmail None No No 3: two claims extracts (one with 140,000 member rows), one contract High
Personal cloud storage None No No 1: a sales rep in their notice period synced a CRM export to a personal drive High
Sync client to unmanaged devices OneDrive sync allowed from any device Audit log No 3 full syncs of the finance site to a home PC Moderate
AI prompts and uploads Approved enterprise AI workspace; personal accounts open Workspace admin log only No 12 PHI pastes into personal AI accounts by 4 support staff; 7 source-code pastes into personal AI chats, one containing a production API key High
USB and removable media Write blocked on Windows; Macs open Windows only Windows only 4 writes, all IT staff imaging devices Low
Print None No logging No Not measured Evidence gap
Screenshots and screen sharing None No No 2 screenshots of claims records shared with a customer during a support call, as the workflow intends Low
Sharing links Tenant allows anonymous links Audit log No 48 anonymous links created; 6 on folders holding contracts and pricing High
Corporate email to external recipients Gateway quarantines SSN patterns Partly Partly 1 claims file (220 members) sent to the wrong customer domain Moderate
Personal code repositories None Laptops only No None on covered laptops; the 22 Linux workstations were not observed Evidence gap

Steps 5 and 6: the scored risk register

ID Risk Evidence (14 days) L I Score Owner Treatment
R1 Claims extracts with PHI sent to personal webmail 3 events, one extract of 140,000 members 4 5 20 VP Customer Operations Mitigate: block regulated data to personal webmail on managed laptops; limit bulk export to two roles
R2 PHI pasted into personal AI accounts by support staff 12 events, 4 people; largest paste about 1,900 member rows 5 4 20 VP Customer Operations Mitigate: give support the approved AI workspace; block PHI to personal AI accounts
R3 Source code and secrets pasted into personal AI chats 7 events, one production API key 5 4 20 CTO Mitigate: rotate the key now; block secrets to personal AI; approved coding assistant
R4 Anonymous links on contract and pricing folders 6 links 5 4 20 General Counsel Mitigate: internal-only default and expiring external links on those sites
R5 Customer list taken to personal cloud by a leaver 1 event 4 3 12 CRO with Head of People Mitigate: activity review at resignation; block CRM exports to personal cloud
R6 Finance and board files synced to unmanaged devices 3 events 4 3 12 CFO Mitigate: allow sync only on managed devices
R7 Claims file emailed to the wrong customer 1 event, 220 members 4 3 12 VP Customer Operations Mitigate: recipient check for claims attachments
R8 USB writes on Macs 4 sanctioned IT writes; Macs open 2 3 6 IT lead Accept until the next hardware refresh; review annually
R9 Print and Linux workstations Not observed n/a n/a Not scored CTO and IT lead Close the evidence gap: print logging, then decide on Linux coverage

The API key in R3 was an incident the day it was found, and it was rotated that day. The register records the pattern behind it. R5 follows a control the DBIR describes on p. 60: an activity review when someone with access to sensitive data resigns.

Step 7: the one-page board summary

Data leak risk at Tallowmere Health Software, October 2026

Our data leak risk is high in four places, and all four are fixable this quarter: patient claims data going to personal email and personal AI accounts, source code with an embedded key pasted into personal AI chats, and contract and pricing folders shared through anonymous links.

How we know: we observed 14 days of real activity on 270 of 292 company computers (92%), plus Microsoft 365 and DNS logs. Staff moved crown-jewel data to personal destinations 23 times, about 38 events per 1,000 staff per week; 19 of those went to AI tools. We did not observe the 22 Linux engineering machines, phones or printing.

Top risks, all scored 20 of 25: R1 claims data to personal webmail (owner VP Customer Operations), R2 PHI into personal AI accounts (VP Customer Operations), R3 code and secrets into personal AI chats (CTO), and R4 anonymous links on contract and pricing folders (General Counsel). The exposed key has been rotated, and R4 closes when sharing defaults are tightened on those sites.

What we accept: USB use on Macs until the next hardware refresh.

What we need: approval for AI workspace seats for the support team and an approved coding assistant.

Next: re-measure 30 days after the fixes and report the change in personal-destination events. Full re-assessment annually, or sooner on any trigger: a new AI tool or AI feature in a core app, a reorganization or layoff, a new regulated data type or customer contract, an acquisition, or a data leak incident.

How do you turn the scored register into the first three fixes?

Sort by score, break ties on impact, then look for a single control that closes several rows. At Tallowmere, four risks tie at 20; R1 ranks first on impact, and the other three each sit under a fix that also closes a lower row. Fix the workflow that leaks by design before you block anything, or the work moves to the next open exit. For each fix, name the owner, the expected residual score and the date you’ll re-measure.

At Tallowmere, the first three fixes were:

  1. Stop regulated data and secrets reaching personal accounts on managed laptops. Block PHI, secrets and CRM exports going to personal webmail, personal AI accounts and personal cloud, and nudge people moving other company data there. One control lowers R1, R2, R3 and R5. On managed laptops, Jazz applies that range of responses, from a prompt or a request for justification to a targeted block, based on the data, the user and the workflow.
  2. Move the work inside the boundary. Support staff pasted claims data into personal AI accounts because the approved tool wasn’t available to them, so extend the AI workspace to support, add an approved coding assistant, and cut bulk-export rights to the two roles that need them.
  3. Close the tenant-side exits. Set sharing to internal-only with expiring external links on sites holding PHI, contracts and pricing, and allow sync only on managed devices. That closes R4, the fourth high risk, and R6.

Then re-run the observation for 14 days with the same sources and scales and report the before and after. If you have no DLP at all, where to start if you have no DLP today lays out the 30-day rollout that follows an assessment like this.

Which frameworks require a written risk assessment, and what must it contain?

None of the major frameworks prescribes a data-loss-specific assessment or a two-to-four-week method. All of them expect a documented, repeatable risk assessment with written criteria, and a data loss assessment built as above supplies the data-handling part of it. Checked against the primary text in October 2026:

Framework Citation What the written assessment must contain How often
ISO/IEC 27001:2022 Clauses 6.1.2 and 8.2 Risk acceptance criteria and criteria for performing assessments; consistent, comparable results; identified risks and risk owners; consequences and “realistic likelihood”; prioritized evaluation; documented results retained At planned intervals or when significant changes are proposed or occur
SOC 2, AICPA 2017 Trust Services Criteria (points of focus revised 2022, published September 2023) CC3.2, with CC3.4 for change Risks identified across the entity and analyzed as the basis for managing them; points of focus include intentional and unintentional acts, criticality of information assets, likelihood and magnitude No interval in the criterion; CC3.4 covers significant changes
HIPAA Security Rule 45 CFR 164.308(a)(1)(ii)(A) An “accurate and thorough assessment” of risks to ePHI. HHS guidance expects you to identify where ePHI is stored, received, maintained or transmitted, plus threats, current measures, likelihood, impact, risk level and documentation No fixed frequency; periodic review and after changes such as new technology or an incident
GLBA Safeguards Rule (FTC) 16 CFR 314.4(b) Written; criteria to evaluate and categorize risks; criteria to assess confidentiality, integrity and availability, including the adequacy of existing controls; how each risk will be mitigated or accepted Periodically, under 314.4(b)(2). Firms with fewer than 5,000 consumers are exempt from the written-content rule under 314.6
NYDFS 23 NYCRR 500.9 (amended 1 November 2023) Documented, under written policies with the same three criteria as the FTC rule; must consider the nonpublic information you collect or store At least annually, and whenever a business or technology change materially changes cyber risk
PCI DSS v4.0.1 Requirement 12.3.1 A targeted risk analysis for each requirement that allows a flexible frequency: assets, threats, factors in likelihood or impact, and the resulting justification. PCI calls an enterprise-wide assessment “recommended, but is not required” Each analysis reviewed at least every 12 months; mandatory since 31 March 2025
NIST CSF 2.0 (February 2024) ID.RA-03 to ID.RA-07 Threats recorded; impacts and likelihoods recorded; inherent risk used to prioritize; responses chosen, tracked and communicated; changes and exceptions recorded Voluntary; no interval
GDPR Article 35 (DPIA) Where processing is likely to be high risk: a description of the processing and purposes, necessity and proportionality, risks to individuals, and the measures to address them Before the processing starts, and again when the risk changes

Read the GDPR row carefully. A DPIA assesses risk to the people whose data is processed, and the processing in question may be your own observation tooling. It complements the company-risk assessment and doesn’t replace it.

HIPAA has a second “risk assessment” that is easy to confuse with the first. The breach-notification test in 45 CFR 164.402 runs after a specific incident and asks whether there is a low probability the PHI was compromised, weighing four factors: the nature and extent of the PHI, who received it, whether it was actually acquired or viewed, and how far the risk was mitigated. Does HIPAA require DLP? covers that test. HHS’s January 2025 proposal to add an asset inventory and annual written risk analysis to the Security Rule was still not final as of October 2026, so the current text governs.

When an auditor asks how you prevent data exfiltration, the written assessment, the gap matrix and the register with owner sign-off are what they ask for first, and SOC 2’s CC6.7 points of focus name “data loss prevention processes and technologies” outright. For the rest of the evidence pack, see what evidence to show your auditor.

When should you re-assess data loss risk?

Re-measure 30 days after the first fixes, using the same sources, window length and scales, so the board sees a change it can trust. Then re-assess at least yearly. NYDFS sets that minimum outright, and ISO clause 8.2 and SOC 2 CC3.4 tie re-assessment to significant change. HHS’s risk analysis guidance names security incidents, changes in ownership, turnover in key staff and new technology as reasons to revisit.

For data loss specifically, write these triggers into the board page: a new AI tool or an AI feature switched on inside an approved SaaS app; a reorganization, layoff or acquisition; a new regulated data type or a customer contract with new data terms; a new exit such as a sync client or file-transfer service; and any data leak incident. Each one changes either the crown-jewel list or the exit list, which are the two inputs everything else in the assessment depends on.

Frequently asked questions

The board just asked about our data leak risk. Where do we start?

Run a scoped data loss risk assessment. Name five to ten crown-jewel data types with a business owner each, list the exits, observe activity for at least 14 days, score and register the risks, then give the board one page. One person needs roughly 35 to 50 hands-on hours across two to four weeks.

How is a data loss risk assessment different from a cybersecurity risk assessment?

A cybersecurity risk assessment, such as one following NIST SP 800-30, covers every threat to every system, including ransomware and unpatched servers. A data loss risk assessment applies the same method to one outcome: sensitive data reaching a person, account or service that shouldn't have it, usually moved by employees with legitimate access.

Can tenant audit logs and proxy logs alone show where sensitive data goes?

Only partly. Microsoft 365 and Google Workspace logs record downloads, sharing links and external shares, but not what happens after a file reaches the laptop. Proxy and DNS logs show destinations, not content unless TLS is inspected, and miss off-network laptops, USB, print, pastes and screenshots. Jazz's endpoint agent covers that gap on managed Windows and macOS laptops.

How can I measure how much sensitive data is actually going into AI apps?

Count every prompt, paste or upload that carries a crown-jewel data type to an AI destination during the observation window, split by corporate or personal account, and express it as events per 1,000 staff per week. Record which sources fed the count, since network-only data misses desktop AI apps and off-network laptops.

How can we see who pastes company data into personal ChatGPT accounts during the assessment?

Tenant logs don't record it, and a proxy can't tell a personal ChatGPT account from the corporate one on the same domain unless it decrypts traffic and reads the tenant. Jazz's agent on managed Windows and macOS laptops sees prompts, pastes and uploads in any browser or desktop app, without a browser extension, and tells a personal AI account from the corporate one, so the count splits by account. Linux machines and phones stay outside it.

Can we run a data loss risk assessment without a proxy rollout?

Yes. Combine the tenant audit logs you already have with an endpoint agent on managed laptops. The agent records the action on the laptop on or off the network, so you don't need to route traffic through a proxy to see uploads, pastes and AI prompts. Jazz does this on Windows and macOS. Record Linux machines, phones and unmanaged devices as evidence gaps on the register.

Our auditor asked how we prevent data exfiltration. What evidence should we show?

Start with the written risk assessment, the exit-by-control gap matrix and a risk register signed by each owner. SOC 2's CC6.7 points of focus name data loss prevention processes and technologies outright. What evidence to show your auditor covers the rest of the evidence pack.

Which frameworks require a written risk assessment?

ISO 27001 clauses 6.1.2 and 8.2, SOC 2 CC3.2, the HIPAA Security Rule at 164.308(a)(1)(ii)(A), the FTC Safeguards Rule at 314.4(b) and NYDFS 500.9 all expect a documented assessment. PCI DSS 12.3.1 requires narrower targeted risk analyses. Only NYDFS sets an explicit annual minimum for the full assessment, and none prescribes a data-loss-specific method.

Is the HIPAA breach risk assessment the same as the Security Rule risk analysis?

No. The Security Rule risk analysis at 45 CFR 164.308 is a standing assessment of risks to ePHI across the organization. The breach-notification test at 164.402 runs after a specific incident and weighs four factors to decide whether a leak is reportable. A data loss assessment feeds the first and doesn't replace the second.

How do you turn a scored risk register into the first three fixes?

Sort by score, break ties on impact, then choose controls that close several rows at once. Fix workflows that leak by design before blocking, for example by giving support staff an approved AI workspace. Name an owner and an expected residual score for each fix, then re-measure for 14 days with the same sources and scales.

Does Jazz cover every device and data store the assessment needs?

No. Jazz runs on managed Windows and macOS laptops, where its agent sees pastes, uploads, screenshots and GenAI prompts in any browser and Melody investigates each risky movement. It doesn't cover Linux, unmanaged devices or phones, and it doesn't scan data at rest inside SaaS tenants, so keep tenant audit logs and record those gaps on the register.

Do we need a DPIA before observing how employees move data?

For EU staff, plan on one. The Article 29 Working Party's WP248 guidelines list an employer systematically tracking staff workstation and internet activity as an example that needs a data protection impact assessment, so complete it before deploying an agent. In every country, tell staff what is visible, why and who sees it before the window opens.

Keep exploring