AI agents

Data security for AI agents

If AI leaks data, the result is customer complaints, audits, and a launch that stalls. The agent sees what your employees see – orders, customer records, internal policies – so before go-live we agree with you on where it is hosted and which data it may open. Hosting is either our US/UK infrastructure or hardware you own. The agent follows the access rights of whoever it is talking to, logs each action, and learns from conversations with names and phone numbers removed.

Get a quote Pricing

CONTACT.WORKData pathdemo data

Path of a customer question

Questionreceived in channelplatform

Searchknowledge base and CRMin platform

To modelmessages and 2 excerptsphone hidden

Replysaved to historylog

CONTACT.WORKWhere data isdemo data

Where AI agent data lives

CallsUS data center or your serveryour choice

Knowledge basenext to the platformsame place

Modelcloud or your ownyour choice

Trainingno personal datacleaned

CONTACT.WORKRestrictionsdemo data

Restrictions in instructions

Other people's datanever disclosedforbidden

Prices and discountsonly from knowledge baserule

Sensitive topicshands off to staffrule

Promisesnone without approvalforbidden

US, UKhosted serversUS and UK data centers

Rolesshape every answerCustomer, partner, employee

Logof every AI answerWhat it opened and what it said

Your serverif you preferWorks without internet

Experience

Details we lock down during setup

A leak through AI almost always comes from excess access or excess information, not from a "clever" model. So that's what we lock down.

Minimum data outThe cloud model gets only the messages in the current conversation and a couple of knowledge base snippets. The full knowledge base stays in the platform.

MaskingPhone numbers, addresses, and ID numbers are replaced with placeholders before anything is sent to the model provider.

Same permissions as a personThe AI agent has the same restrictions as an employee in the same role, and nothing beyond them.

Restrictions in the instructionsA list of topics the AI stays away from and cases where it brings in one of your employees right away.

A separate user accountThe AI connects to your CRM and ERP as a separate user with permissions defined operation by operation.

Retention periodConversations and transcripts are kept for the period you choose and then deleted.

Use cases

What we protect in practice

Three everyday contact center use cases and how each one keeps unnecessary information out of reach.

The AI agent sees only the person who reached out

The AI consultant identifies the customer by phone number or messaging app account and opens only their orders and tickets. It simply has no information about other customers for this conversation. If a cloud model drafts the answer, the phone number and address are replaced with placeholders before anything is sent. The customer gets an immediate answer, and your company never puts anyone else's data at risk.

Only their own orders
Contact details hidden from the cloud
No access to other customers

CONTACT.WORKAccessdemo data

Customer call · what AI sees

Customerrecognized by numberphone

Visibletheir orders and ticketsown only

Other customersunavailablehidden

To cloudphone and address – placeholdershidden

Use case 1 of 3

Each role gets its own information

Agents get policies and their own customers' data from the AI assistant, while purchase prices, management reports, and other departments' information stay open only to people with access. The AI assistant identifies employees by their platform login and works within the same limits. You can safely give the AI assistant to every employee – it won't show anyone more than they should see.

Role-based permissions
Restricted sections stay hidden
Sign-in with a user account

CONTACT.WORKStaff accessdemo data

AI assistant · role permissions

Agentpolicies and own customersopen

Purchase pricesmanagers onlyhidden

Team reportsto department headby role

Loginplatform accountchecked

Use case 2 of 3

Every answer can be traced step by step

A customer disputes an answer, or your security team reviews how the AI works. The log shows the question, which records and articles the AI opened, and what it answered. Errors are found in minutes, the reviewer gets facts instead of secondhand accounts, and the issue is closed without a long email thread.

Question and answer
Information it opened
Steps in order

CONTACT.WORKLogdemo data

AI answer log

Questionhow do I return an item?logged

Visiblearticle and order in CRMbooked

Replysent to customerlogged

Accesssecurity teamby role

Use case 3 of 3

How recordings are stored, who controls what, and how consent is collected are covered on the platform security page.

How it works

Where data lives and what the AI learns from

Cloud

US and UK data centers, your access controls

Chats, audio recordings, and reference materials are stored on our servers in the US and the UK, and your administrator decides who sees them. If your policies limit where data can be processed, answers come from a model provider that meets those requirements. You don't need your own servers or admins.

Hosted in US and UK data centers
You grant access
Model provider that fits your policies

CONTACT.WORKClouddemo data

AI agent in the cloud

Platformservers in the UScloud

ModelRussian on requestyour choice

To modelmessages and excerptsno desk phones

Accessgranted by your adminon your servers

Your server

Everything stays in-house

We install everything on your side: the platform, the knowledge base, and an open-source model. No internet connection is needed for answers, and your security team controls the AI agent like any other internal system. Companies that cannot let data leave the building choose this option: banks, clinics, and government agencies.

Platform on your servers
Open-source model on your servers
No internet access needed

CONTACT.WORKYour networkdemo data

Everything on your server

Platformyour serveron your servers

Modelopen model, on a GPUon your servers

Internetnot needed for answerswithout it

Datanever leave the companyinside

Training

Training on de-identified conversations

Before training the AI on real chats and calls, we replace names, phone numbers, addresses, and contract numbers with placeholders. An expert spot-checks that nothing was missed, and only then do the conversations go into training. The AI learns from your real work, with no personal data in the training set.

Placeholders instead of names
Expert review
Only scrubbed conversations

CONTACT.WORKDatasetdemo data

Training prep

Calls12,400 selectedsample

Names and phonesreplaced with placeholdersremoved

Addresses and contractsreplaced with placeholdersremoved

Reviewspot-checked by expertpassed

Ready for auditsCustomer data stays in our US and UK data centers or on your servers, is accessible by role, and very little leaves the platform – which helps with CCPA/CPRA and other privacy reviews.

Customers trust the AIEveryone sees only their own data, so customers are happy to resolve issues with the AI instead of waiting on hold for an agent.

Complaints resolved in minutesThe log shows how the AI reached its answer, so you settle disputes with facts, not guesswork.

No launch delaysThe access model is ready up front, so your security team signs off sooner and savings start earlier.

Launch

How we set up AI agent security

01RequirementsWe review your internal data protection policies and the list of restricted data.
02DeploymentWe decide: our cloud or your servers, and a cloud model or an open-source model on your side.
03RolesWe define what customers, partners, and each department are allowed to see.
04RestrictionsWe write down which topics the AI avoids and when it brings in an employee right away.
05TestingWe test answers across different roles and check the log against expected results.
06OversightAfter launch, we give managers and security specialists access to the log.

Requirements

What we need from you

Policies: Your internal rules for handling data
Roles: Who sees what in each conversation
Deployment: Our cloud servers in the US and the UK or your hardware
Model: A cloud model that fits your policies, or an open-source model on your servers
User accounts: Employee sign-in through the platform
Cost: Setup is billed in support hours

Preparation

What to decide before launch

What counts as restricted?Information the AI never opens for anyone: salaries, purchase prices, internal reports. We configure restrictions from this list.

Can data leave your company?Whether you can send data to a cloud provider and whether it must stay in a specific country. The answer decides between our cloud and a server inside your company.

Who is allowed to see what?What customers, partners, and employees in each department see. This usually mirrors the roles already set up in the platform.

Which topics are off-limits for the AI?Discounts beyond policy, complaints, legal questions – the AI doesn't handle these conversations and hands them to an employee.

Who will review the log?Usually the contact center manager, the quality team, and a security specialist.

Do you need documentation?Whether your security team needs a diagram of the deployment and data flows to approve the launch. If so, we will prepare it.

Q&A

Frequently asked questions

Does the cloud model see our entire knowledge base?

No. It gets only what is being said right now and a couple of relevant snippets. Everything else – the knowledge base, conversation history, customer data – stays in the platform. If even that cannot leave your company, we install an open-source model on your servers.

Where are customer conversations with the AI stored?

Wherever the platform itself is hosted: on CONTACT.WORK servers in America or Britain, or inside your own infrastructure. Viewing rights are set by role, the same as for ordinary conversations.

Is data encrypted in transit?

Data travels over secure connections between the platform, communication channels, and the cloud model.

Can CONTACT.WORK employees see our data?

Only when you request it, for example to troubleshoot an error. Every such access is recorded in the log.

How long are records kept?

As long as you decide. When the retention period ends, the records are deleted.

What happens to our data when the contract ends?

We delete it from the cloud when the contract ends, or earlier if you ask.

Does the model learn from our conversations?

Only if you ask for it. If you decide to train the AI on your conversations, we use them only after removing names, phone numbers, and addresses. Otherwise, conversations are not used for training.

Who owns the model after training?

Your company does. Whoever paid for the training owns the resulting model, the approved answers, and the knowledge base.

Can we run completely offline?

When the entire system, including an open-source model, is installed inside your company, it needs no internet access to answer.

How does the AI assistant know what it can show an employee?

From their platform login: the AI assistant can access only the sections open to that employee.

Who can read the answer log?

Whoever you give access to – usually the contact center manager, the head of quality assurance, and a security officer.

Do you help get the launch approved by our security team?

We describe the deployment, data flows, and storage locations in a document for your reviewers.

What if our policies restrict where data is processed?

We use a cloud model from a provider that meets your data residency requirements, or install an open-source model on your servers so nothing leaves your company.

How is the platform itself protected?

Recordings, conversations, and customer records are covered by the platform's roles, logs, and retention rules.

Platform security →

How much does security setup cost?

Setting up roles, restrictions, and deployment is billed in support hours, and the platform has no license fees.

Pricing →

Show all questions (15)

Glossary

Terms on this page

Cloud deployment (cloud): The platform runs on the vendor's servers, and agents connect over the internet. Learn more →
Knowledge base: All answers, policies, and prices in one place – for agents and AI. Learn more →
Hosting: You rent a server from a provider instead of buying one. Learn more →
Data center: A building with servers, backup power, cooling, and network links. Learn more →
Customer record: The screen of customer data an agent sees during a conversation. Learn more →
CRM – customer relationship management: Where a company keeps its customers, deals, and full communication history. Learn more →
ERP – enterprise resource planning: Orders, inventory, and finance in one system. Learn more →
On-premises deployment: The system runs on the company's own servers, not in the vendor's cloud. Learn more →
Personal data: Anything that points to a specific person: a name, a number, a voice. Learn more →
Encryption: Encrypted data can't be read without the key. Learn more →
Contact center quality management: Calls and chats are scored against a scorecard to find mistakes and coach agents. Learn more →

Blog

Fresh from the blog

All articles →

Services · 7 min read Custom development of your contact center platform for your processes When contact center settings are not enough: integrations, modules, AI workflows. How to tell development from configuration and estimate the payback. Read the article → Management · 9 min read How many agents does your contact center need? Erlang C in plain English Calculate contact center staffing with Erlang C: workload, service level, shrinkage, and headcount by shift – all in one worked example. Read the article → Integrations · 7 min read CRM and phone system integration: what it does and how it works How to connect your phone system to your CRM: calls and recordings in the record, screen pop, click-to-call. What to decide first and how to check it works. Read the article →

Updated: November 11, 2026