Talk to the Data answers business questions across your systems in plain language. You get connected records, your own business rules and the evidence behind each answer.
Start with sample data. Connect your systems when you are ready.
The problem
Your answers exist. Finding the full picture takes too long.
Your records sit in separate systems. Teams join the facts and rules by hand before they can decide.
01
The answer is scattered
Finance, operations, customers and contracts each describe a piece of what happened.
02
The context gets lost
Without business rules and relationships, a correct number may still lead to the wrong conclusion.
03
The moment to act passes
By the time teams assemble and reconcile the picture, the business may have moved on.
What is Talk to the Data?
Talk to the Data reads your systems, applies your business rules and shows its evidence.
What you can ask, grid by grid.
Pick a grid, then a question teams ask there.
What makes OpenKnowra different
Reading data is not understanding your business.
Getting an answer is useful. Knowing how facts, relationships and rules fit together makes it actionable.
An AI assistant connected to your data
Finds content across your documents and systems
Summarizes what the sources say
May cite the documents it used
Needs extra setup to apply decision logic
Talk to the Data with OpenKnowra
Connects records through shared business entities
Interprets questions with your context and rules
Shows the records, relationships and rule applied
Supports governed, repeatable business decisions
Capabilities vary by data sources, configured integrations and business rules. Other enterprise AI products may also offer retrieval, citations, graphs or governance.
Built around your enterprise
Your systems stay. Your business context connects.
Your systems stay where they run, and your own rules decide what comes back.
Start on sample data. Connecting your own uses the usual integrations and permissions.
Which degrading assets get a preemptive work order, in what order, and in which window - for asset managers and the NOC.
AgentMaintenance Planner
Decides onNetwork asset
Records in the graph210
Connected5 systems
Recommends
Recommend which network assets need a work order this week
Order of work
Assets with the shortest remaining useful life against the SOP’s thresholds, weighted by the customers they serve
Weighs
Remaining useful lifethe alarms and observationsthe weather and grid noticesthe vendor advisories
Hard limit
Assets under a regulatory hold, or outside their maintenance window, are deferred.
11 rules
1 An asset whose RUL days are below 90 and whose state is DegradedState is flagged as a maintenance candidate (already failing).
1
2 An asset whose state is Failed is flagged as a maintenance candidate (already failing).
2
3 An asset whose RUL days are below 7 and that is chronically failing is flagged as a maintenance candidate (a predictive catch).
3
4 An asset at a coastal site that is chronically failing and whose RUL days are below 14 is flagged as a maintenance candidate (an environmental risk).
4
5 An asset that is a single point of failure and whose RUL days are below 90 is flagged as a maintenance candidate (a single point of failure).
5
6 An asset with one or more recall advisories is flagged as a maintenance candidate (a vendor recall).
6
7 An asset whose consumed life share is 1 or more is flagged as past its derated design life.
7
8 An asset with 3 or more critical alarms is flagged for NOC review.
8
9 An asset with one or more regulatory holds at its site is flagged for permit clearance before a crew is sent.
9
10 An asset with one or more freeze windows at its site is flagged: its site is under a change freeze.
10
1 Work is taken by urgency points, highest first; then by ACV at site, highest first; then by priority score (the number, not the band), highest first.
The build portfolio for the year’s budget - each proposed build valued (NPV, payback, the downside), the unfundable kept out, the rest funded by return per dollar - and which need the CFO.
AgentCapital Plan Advisor
Decides onPlan cycle
Records in the graph1
Connected3 systems
Recommends
Fund the plan areas of each cycle, in priority order, within the budget
Order of work
Highest return in subscribers, margin and NPV over the planning horizon, balanced against payback and downside risk
Weighs
NPVpaybackdownside
Hard limit
Builds that can't be funded are excluded: NPV not positive, permits not ready, or payback longer than 7 years.
3 rules
4 A plan area whose capex is 2000000 or more is flagged for the CFO's review.
1
5 A plan area whose downside NPV is below 0 is flagged for a risk review: it loses money at the downside take rate.
2
6 A plan area whose competitor is fibre overbuilder is flagged: plan a competitive response.
3
Entities it uses
Plan areaTake-rate studyPlan cycle
Source: Capital Plan Procedure - network builds for the plan year, revision 1 (sample). Sample data.
The credit the service agreement pays for an outage, and when the account’s context - its credits, its household, its churn, a chronic site - overrides it - for care and save-desk agents.
AgentCredit & Refund Advisor
Decides onCredit request
Records in the graph53
Connected5 systems
Recommends
Clear each credit request, or hold it for fraud review
Order of work
The SOP’s limits per account and household, then the churn risk
Weighs
8 rules
1 A credit request whose SLA eligible is 1 and whose credits this year are 4 or more is flagged to hold for fraud review: too many credits on the account this year.
1
2 A credit request whose premium siblings are 1 or more and whose own plan premium is 0 is flagged to credit under the household's stricter plan.
2
3 A credit request whose SLA eligible is 0, whose churn risk is 0.7 or more and whose chronic outages are 3 or more is flagged for goodwill: a customer likely to leave, on a site that keeps failing.
3
4 A credit request whose SLA eligible is 0, whose churn risk is 0.7 or more and whose lifetime value is 12000 or more is flagged for goodwill: a valuable customer likely to leave.
4
5 A credit request whose SLA eligible is 1, whose chronic outages are 3 or more and whose family lines are more than 1 is flagged for a chronic uplift.
5
6 A credit request whose SLA eligible is 1 is flagged to pay the SLA credit.
6
7 A credit request whose SLA eligible is 0 is flagged to deny: not eligible under the agreement.
7
1 The requests are reviewed by SLA credit, highest first.
Which partner sites to onboard, onboard with conditions, hold or return to the partner - and what each partner fixes first.
AgentPartner Onboarding Advisor
Decides onPartner site
Records in the graph90
Connected3 systems
Recommends
Recommend for each partner site: onboard, onboard with conditions, hold, or return to the partner
Order of work
Gold partners first, then the earliest submission
Weighs
Data completenesscoordinatespermitselement support and end of supportpower backupoverlap with our sitespartner security score
Hard limit
Sites without an active partner contract are held.
11 rules
A partner site whose partner has no active contract is held: nothing is onboarded without a signed agreement.
1
A partner site with a data completeness below 90% is returned to the partner for the missing fields.
2
A partner site whose coordinates are not valid is held for a site survey.
3
A partner site whose permits are not valid is held until the partner provides them.
4
A partner site with one or more elements not supported by our network management system is referred to the integration team.
5
A partner site with one or more elements reaching end of support within 180 days is onboarded with conditions: the partner replaces or upgrades them first.
6
A partner site with power backup below 4 hours is onboarded with conditions: a battery upgrade to our 4 hour standard.
7
A partner site within 0.5 km of one of our own sites is reviewed as shared infrastructure, not as new coverage.
8
A partner site whose partner has a security score below 70 is held for a security review.
9
A partner site that passes each check above is onboarded.
10
Partner sites are worked Gold partners first, then the earliest submitted.
A store whose customers made 20 or more calls in the last 30 days and whose repeat call rate is 30% or more is flagged for a process review by the regional manager.
1
A store whose first-contact resolution is below 50% is flagged for a coaching visit.
2
A store whose average sentiment is below -0.2 is flagged for the customer experience lead.
3
A store whose average wait is more than 25 minutes and whose weekly footfall is above 1,500 is flagged for a staffing review.
4
A store whose top call reason is Bill shock is flagged for a billing explanation refresher with its staff.
5
Stores are worked highest repeat call rate first, then the most calls from customers.
6
Entities it uses
StoreSubscriberInteraction
Source: XYZ Telecom - Store call insights standard operating procedure (sample). Sample data.
A contact centre agent who handled 3 or more contacts in the last 30 days and whose first-contact resolution is below 60% is flagged for resolution coaching by their team leader.
1
A contact centre agent whose repeat call rate is 30% or more is flagged for a call review with the quality team.
2
A contact centre agent who transfers 25% or more of their contacts is flagged for product knowledge training.
3
A contact centre agent whose average customer sentiment is below -0.2 is flagged for soft-skills coaching.
4
A contact centre agent with less than 6 months of tenure and any of the flags above is paired with a buddy from their team.
5
Agents are coached lowest first-contact resolution first, then the highest repeat call rate.
6
Where the quality team wants the detail of a call, the call's transcript is scored with Call Quality Assessment, and the agent's weakest factors set the coaching plan.
7
Entities it uses
Contact centre agentInteraction
Source: XYZ Telecom - Contact centre agent call coaching standard operating procedure (sample). Sample data.
Which invoices are overdue in approval, due a discount, or need a second look.
AgentPayables Advisor
Decides onInvoice
Records in the graph59
Connected3 systems
Recommends
Flag the invoices overdue in approval, due a discount or needing a second look
Order of work
Earliest invoice date first, then largest amount
Weighs
Statusinvoice dateamountthe supplier’s payment terms and segment
5 rules
An invoice whose status is Approving and whose invoice date is more than 30 days ago is escalated to the approver’s manager.
1
An invoice whose status is Rejected and whose invoice date is more than 45 days ago is returned to the supplier for a credit note.
2
An invoice from a supplier whose payment terms are 2/10NET30 and whose status is Approving within 10 days of the invoice date is prioritised for the early-payment discount.
3
An invoice above 50,000 EUR from a supplier in the Tail segment is checked by the accounts payable lead.
4
Invoices are worked earliest invoice date first, then largest amount.
Which findings are escalated to the CISO, ticketed or fixed first: exploitability, the deadline, whether endpoint security sees the device, its user’s admin roles and MFA, and whether a ticket exists.
AgentVulnerability Priority Advisor
Decides onVulnerability finding
Records in the graph90
Connected5 systems
Recommends
Flag the vulnerability findings to escalate to the CISO, ticket or fix first
Order of work
Critical first, then exploitable, then the highest priority rating, then the longest open
Weighs
Severityexploit availabledays openendpoint coveragean open ticketthe device user’s admin roles and MFAapplication control mode
10 rules
A critical vulnerability finding with an exploit available that has been open more than 7 days is past its deadline and is escalated to the CISO.
1
A critical vulnerability finding open more than 15 days is past its deadline and is escalated to the CISO.
2
A high vulnerability finding open more than 30 days is past its deadline and is flagged for the infrastructure lead.
3
A critical or high vulnerability finding on a device that endpoint security for Endpoint does not see is flagged as a coverage gap for security operations.
4
A critical or high vulnerability finding whose vulnerability has no remediation ticket is flagged to raise a ticket.
5
A critical or high vulnerability finding whose remediation ticket is Done while the finding is still open is flagged to reopen the ticket.
6
A vulnerability finding on a device used by a person who holds an admin role is escalated to the CISO.
7
A vulnerability finding on a device used by a person who has not registered for MFA is escalated to the CISO.
8
A vulnerability finding on a computer that application control does not hold in secured mode is flagged for the endpoint team.
9
Vulnerability findings are worked critical first, then those with an exploit available, then the highest priority rating, then the most days open.
10
Entities it uses
Vulnerability findingendpoint security deviceVulnerabilityManaged deviceUserMFA registrationAdmin role assignmentProtected computerTicket
Source: XYZ Bank Cyber Security - Vulnerability remediation standard operating procedure (sample). Sample data.
Which dependency alerts must be fixed first, and who owns them?
Which vulnerable dependencies are fixed now, within their deadline, mitigated or accepted, and which team owns each: runtime or development, severity, EPSS, a fix available, internet-facing or archived repositories, a valid exposed secret.
AgentDependency Triage Advisor
Decides ondependency alert
Records in the graph23
Connected1 systems
Recommends
Flag the dependency alerts to fix now, fix within their deadline, mitigate or accept, or close, each with its owning team
Order of work
Runtime before development, then critical first, then the highest EPSS, then internet-facing repositories, then the longest open
Weighs
Severityruntime or development scopeEPSSa fix availabledays openthe repository internet-facing or archiveda risk acceptance past its expirya valid exposed secret in the repositorythe team that owns the repository
13 rules
An open dependency alert in runtime scope that is critical and has an EPSS of 0.1 or more is escalated to the Head of Application Security and fixed now.
1
An open dependency alert in runtime scope that is critical, in an internet-facing repository, and open more than 7 days is past its deadline and escalated to the Head of Application Security.
2
An open dependency alert in runtime scope that is critical and open more than 15 days is past its deadline and escalated to the Head of Application Security.
3
An open dependency alert in runtime scope that is high, in an internet-facing repository, and open more than 15 days is past its deadline and flagged to the owning team's maintainer.
4
An open dependency alert in runtime scope that is high and open more than 30 days is past its deadline and flagged to the owning team's maintainer.
5
An open dependency alert in runtime scope that is medium and open more than 90 days is past its deadline and flagged to the owning team.
6
An open dependency alert in development scope that is open more than 90 days is flagged to the owning team for its next planned upgrade.
7
An open dependency alert with no fix available is sent to the owning team to mitigate it (remove or replace the package, or block the vulnerable path) or to record a risk acceptance with an expiry.
8
A dependency alert accepted as a tolerable risk whose acceptance has expired, or that names no expiry, is reopened and escalated to the Head of Application Security.
9
An open dependency alert in an archived repository is not fixed: the repository is confirmed retired and the alert closed, or, when the code still runs somewhere, the repository is unarchived and the…
10
An open dependency alert in a repository that also has an open secret scanning alert whose secret is still active is escalated to the Head of Application Security: the secret is rotated now, before a…
11
An open dependency alert in a repository that no team holds with admin or maintain is flagged to the Head of Engineering to name an owning team.
12
dependency alerts are worked runtime before development, then critical first, then the highest EPSS, then internet-facing repositories first, then the most days open.
Which devices with a live endpoint security alert also opened or downloaded confidential engineering documents, carry an exploitable finding or had an execution blocked: isolate, escalate, or leave to the alert queue.
AgentIP Exposure Advisor
Decides onendpoint security device
Records in the graph66
Connected5 systems
Recommends
Flag the devices to isolate pending investigation, escalate, or leave to the alert queue
Order of work
The highest alert severity first, then the highest exposure level, then the most recently seen
Weighs
Alert status and severityconfidential documents the device’s user opened or downloaded in the last 7 daysan exploitable critical findingexecutions application control deniedwhether the user is a guest
7 rules
A endpoint security device with an alert that is New or In Progress whose user opened or downloaded a confidential document in the last 7 days is isolated pending investigation and escalated to the CISO as a…
1
A endpoint security device with an alert that is New or In Progress and a critical vulnerability finding with an exploit available is isolated pending investigation.
2
A endpoint security device with an alert that is New or In Progress on which application control denied an execution in the last 7 days is escalated to the security analyst to review the blocked program.
3
A endpoint security device whose user downloaded more than 3 confidential documents in the last 7 days is flagged to the data protection lead, whether or not an alert is open.
4
A endpoint security device whose user is a guest and opened or downloaded a confidential document is flagged for the export control review.
5
A endpoint security device with an alert that is New or In Progress and none of the above is left to the normal alert queue.
6
Devices are worked by the highest alert severity first, then the highest exposure level, then the most recently seen.
7
Entities it uses
endpoint security deviceSecurity alertVulnerabilityVulnerability findingManaged deviceUserDocument activityProtected computerApplication control event
Source: XYZ Energy Cyber Security - IP exposure response standard operating procedure (sample). Sample data.
Which assets are missing from the register or from a security tool, missing the data the asset standard requires, gone quiet, or retired but still on the network.
AgentAsset Coverage Advisor
Decides onAsset
Records in the graph77
Connected5 systems
Recommends
Flag the assets to register, complete, onboard to the EDR, enrol in device management, locate or retire, or investigate
Order of work
High criticality first, then servers and plant equipment, then the most coverage gaps, then the longest unseen
Weighs
Statusdevice classin the registerin the EDRmanagedscannedmissing fieldsdays since seencriticality
3 rules
1 Each asset must be listed in the asset register. An asset seen by a security tool but not in the register must be registered within 5 working days; a server not in the register is escalated to the…
1
2 Each active laptop, workstation and server must run the EDR agent (endpoint security for Endpoint). An active asset without the EDR is onboarded within 2 working days; a server without the EDR, or any asse…
2
2 An asset whose register status is Retired, Disposed or Lost but which a security tool has seen in the last 30 days is investigated by Security as a priority: a retired device still on the network i…
Which people and files show signs of data leaving the company?
Which people show signs of data leaving the company, and what is done first: anonymous links to labelled files removed, external forwarding disabled and the user interviewed, unverified apps with mail scopes revoked, DLP alerts on confidential files escalated.
AgentData Exposure Advisor
Decides onUser
Records in the graph80
Connected3 systems
Recommends
Flag the users to act on: remove a link or give it an expiry, disable forwarding and interview, revoke an app, escalate a DLP alert, review a guest
Order of work
A DLP alert on a confidential file first, then forwarding outside the company, then anonymous links to labelled files, then high-risk app consents, then the most recent
Weighs
DLP alerts: severitystatusthe file’s label; Links: anonymous or to a guestthe file’s labelthe expiry; Inbox rules forwarding outsidedeleting or hiding security mail; App consents with high-risk scopesthe publisher verified or not; What the person touched in the window: confidential or IP filesthe programs they belong toa program ahead of its milestone or export controlled; The kinds of sign of data leaving the person showscounted on their record
12 rules
A user with an open DLP alert (new or in progress) on a file labelled confidential is escalated to the CISO the same day, with the policy matched and the recipient.
1
A user with a DLP alert on card numbers is escalated to the CISO and the card data is removed from the file or the file deleted, as PCI DSS requires.
2
A user who shared a labelled file by an anonymous link has the link removed now; the user confirms who received it.
3
A user who shared a file by an anonymous link with no expiry has the link given an expiry of 14 days at most, or removed when the file is labelled.
4
A user with an inbox rule that forwards or redirects mail outside the bank has the rule disabled now and is interviewed by the security analyst; when the rule also moves mail to RSS Feeds, Archive or…
5
A user with an inbox rule that deletes or hides mail from the security team, or mail about phishing, passwords or suspicious sign-ins, has the rule removed and the account checked for compromise.
6
A user who consented to an unverified app with a mail scope (Mail.Read, Mail.ReadWrite or Mail.Send) has the consent revoked and the app blocked; the user is told why.
7
A user who consented to any app with Files.ReadWrite.All or Sites.Read.All is reviewed by the identity engineer, who revokes the consent unless the business owner approves the app.
8
A guest given a link or a grant to a labelled file has their access reviewed by the file owner within 5 working days; access with no expiry is given one or removed.
9
A user with two or more signs of data leaving is escalated to the CISO as one case.
10
A user with at least one sign of data leaving who touched an IP file in the last 7 days (IP files touched above zero) is escalated to the CISO as one case, with the places and the programs named; whe…
11
Users are worked by a program ahead of its milestone touched first, then a DLP alert on a confidential file, then forwarding outside the bank, then anonymous links to labelled files, then high-risk a…
Which machines would hurt most if compromised, and what do we fix on them first?
Which machines to harden first, before anything happens: the risk of compromise (no EDR, not managed, unencrypted, exploitable findings, out-of-support software, an admin owner) against the IP the owner can reach (the IP register’s sites, the files they touch, the repositories their teams hold).
AgentBlast Radius Advisor
Decides onAsset
Records in the graph77
Connected7 systems
Recommends
Flag the machines to harden first: onboard to the EDR, enrol and encrypt, patch the exploitable findings, remove out-of-support software, take the admin role off the daily account
Order of work
The largest blast radius first, then the highest risk of compromise, then the most IP within reach
Weighs
Blast radius and its bandthe risk of compromise and its factorsthe IP within reach and its factorsthe device classthe criticalitythe owner’s department
8 rules
An asset with a high blast radius is stopped and hardened this week: each factor on it is closed, in the order below, and the CISO is told what remains.
1
An asset with no EDR and a medium or high blast radius is stopped and onboarded to endpoint security for Endpoint within 2 working days; no other change waits for it.
2
An asset with an exploitable or known exploited finding and a medium or high blast radius is stopped and patched within 7 days, whatever the finding's own date.
3
A laptop not encrypted or not managed, with a medium or high blast radius, is stopped and enrolled and encrypted within 5 working days.
4
An asset with out-of-support software and a medium or high blast radius is stopped and has the software removed or upgraded within 30 days.
5
An asset whose owner holds an admin role, with a medium or high blast radius, is stopped and has the owner moved to a separate admin account within 10 working days: the daily account keeps no admin r…
6
An asset with a medium blast radius and none of the above is reviewed in the monthly hardening cycle.
7
Assets are worked by the largest blast radius first, then the highest risk of compromise, then the most IP within reach, each largest first.
Which incidents need containing now, and what leaves with them?
Which open incidents show confidential or IP data leaving, or about to: a device under attack whose user touched confidential files, signed in from somewhere new, uploaded or copied files out, grew a forwarding rule, gave an app each file, cloned repositories - and what is contained first: isolate, revoke sessions, block sign-in, disable the rule, remove the links, revoke the consent and the tokens, wipe the phone, sweep the other devices.
AgentBreach Containment Advisor
Decides onSecurity case
Records in the graphNot in the graph
Connected7 systems
Recommends
Flag the open incidents to contain: isolate the device, revoke the user’s sessions and block sign-in, disable the forwarding rule, remove the links, revoke the consent and the tokens, wipe the phone, sweep the other devices, escalate to the CISO and export control; and for the business: patent counsel, the partners within their notice, the lab lead, the impact written onto the endpoint security incident
Order of work
The highest IP business impact first, then confirmed exposure, then the highest severity, then the most confidential files touched, then the most recent
Weighs
Exposure: a confidential or IP file that left (an uploada USB copya DLP alerta repository cloned)confidential files touched since the startsign-ins from an unusual countrya forwarding rule made sincerisky app consentstokens madethe same indicator on other devicesthe device’s classphones with company data; Business impact: the programs whose files were touched or lefta program ahead of its milestone (days to a patent filing or a launch)export-controlled data leftpartners to notify and the shortest noticeopen experiments relying on the datathe IP business impact
Hard limit
A plant controller or any other agentless device is not isolated by a rule: it is contained at the network with the engineer who runs it.
14 rules
A case whose exposure is confirmed (confidential data left) is stopped and contained now: the device is isolated, the user's sessions are revoked, and the CISO and the Data Protection Officer are tol…
1
A case with a sign-in from an unusual country since the start is stopped and treated as an account in someone else's hands: the sessions are revoked and sign-in is blocked until the password is reset…
2
A case whose mailbox grew a rule forwarding outside, or hiding or deleting mail, since the start is stopped: the rule is disabled and the account treated as compromised.
3
A case where the person consented to an app since the start is stopped: the consent is revoked and the app blocked.
4
A case where a token or key was made since the start, or a repository cloned with one, is stopped: the tokens are revoked and each secret the repositories hold rotated.
5
A case where confidential files were touched since the start but nothing is seen leaving is isolated pending investigation, and any live anonymous link the person holds is removed.
6
A case whose indicator was denied on other devices has those devices swept the same day: the same script, the same sender, the same download.
7
A case on an agentless device is not isolated by a rule: it is contained at the network with the engineer who runs it, and the CISO decides.
8
A case where the person has a phone with company data (one or more phones), and the account is in someone else's hands (a sign-in from an unusual country) or confidential data left, has the company d…
9
A case whose data left from a program ahead of its milestone is stopped: the program's sponsor and legal counsel are told within 24 hours with the files that left, and decide whether the negotiation…
10
A case where export-controlled data left is stopped: the compliance officer is told the same day, and decides on the report the regulations ask.
11
A case whose data left from a program with partners is stopped: each partner is told within the notice its agreement asks (the shortest notice on the case), by the program lead with legal.
12
A case with a Critical IP business impact is flagged for the SOC: its impact, its programs and its partners are written onto the endpoint security incident, so the incident queue and its playbooks act on them.
13
Cases are worked by the IP business impact first (critical, high, medium, low), then by confirmed exposure, then by severity (critical, high, medium, low), then by the number of confidential files to…
Which AI agents could put our IP together and send it out?
Which AI agents could reconstruct the company’s IP: the places each can read with its own identity, the kinds of knowledge those hold, the sensitive combinations that makes, and what it can send out (an external model, email, a web request) - and what is narrowed first: the outbound permission, the sites beyond its approval, the broad permission, the unregistered agent.
AgentAI Exposure Advisor
Decides onAI agent
Records in the graphNot in the graph
Connected4 systems
Recommends
Flag the AI agents to narrow: remove the outbound permission, take back the sites beyond its approval, replace the broad permission, disable the unregistered agent, escalate an interaction that went outside, send the rest to the owner for review
Order of work
The highest IP exposure first, then the most sensitive combinations within reach, then the most IP places within reach
Weighs
IP exposure and its factorsthe sensitive combinations within reachwhat it can send outthe IP places beyond its approvala broad permissionin the register and approvedthe interactions that combined restricted knowledge and went outsidethe model
8 rules
An AI agent with a Critical IP exposure is stopped: the permission that lets it send outside is removed and the sites granted beyond its approval are taken back within 2 working days; the owner and t…
1
An AI agent that is not in the AI agent register and reaches IP is stopped and disabled until its owner registers it and the CISO approves it.
2
An AI agent with a broad permission that reaches IP beyond its approval is stopped: the broad permission is replaced by the sites it was approved for, granted one by one, within 10 working days.
3
An AI agent whose interactions combined restricted knowledge and went outside is stopped and escalated to the CISO and the knowledge owners the same day: what was read, by whom, and where it went.
4
An AI agent with a High IP exposure and none of the above goes to its owner for review within 30 days: the reach confirmed against its business purpose, or narrowed.
5
An AI agent that uses an external model and reaches IP is flagged: the bank's tenant model is used for that work, or the owner records why not.
6
A code assistant installed on each repository is flagged: its installation is narrowed to the repositories it was approved for.
7
Agents are worked by the highest IP exposure first, then the most sensitive combinations within reach, then the most IP places within reach.
Which AI interactions put restricted knowledge together and sent it out?
Which AI interactions put restricted knowledge together, and which may leave: denied when a sensitive combination goes to an outside model or tool, held for the knowledge owner when IP goes outside, allowed otherwise - the same answer an AI gateway gets before the transfer.
AgentAI Transfer Advisor
Decides onAI interaction
Records in the graphNot in the graph
Connected4 systems
Recommends
Flag the AI interactions to deny or hold for review: a sensitive combination sent outside, an unregistered agent reading IP, IP sent outside, a sensitive combination kept inside
Order of work
A sensitive combination sent outside first, then the most kinds of IP knowledge read, then the most recent
Weighs
The sensitive combination put togetherwent outsidethe kinds of IP knowledge readthe IP places readthe agent in the register and approvedthe agent’s IP exposure
6 rules
An interaction that puts a sensitive combination together and goes outside is stopped: the transfer is denied, and the knowledge owners of both kinds of knowledge are told.
1
An interaction by an AI agent that is not in the AI agent register, and reads IP, is stopped.
2
An interaction that reads IP and goes outside is flagged for review by the owner of the data read.
3
An interaction that puts a sensitive combination together and stays inside is flagged: the owners of both kinds of knowledge confirm the person needs it.
4
An interaction by an AI agent with a Critical IP exposure that reads IP is flagged for review.
5
Interactions are worked by a sensitive combination that went outside first, then the most kinds of IP knowledge read, then the most recent.