Agencies and consultancies

You hold other people''s confidential data for a living.

Client strategy, unreleased campaigns, pricing and personal data from their customers — all on your laptops, all under their NDA, and all one paste away from a tool your client has not approved.

CLIENT A · 4 rules CLIENT B · 6 rules CLIENT C · 3 rules ONE LAPTOP, SEVERAL NDAS
Built for Creative and media agencies Management consultancies Design studios PR and comms Outsourced marketing

The threat

The brief is the client''s. So is the risk.

Every path below is somebody delivering work a client commissioned. The difference from an in-house team is that the consequence of a leak is contractual as well as regulatory, and it belongs to a relationship rather than to a system.

  1. Rewriting a client brief

    The brief is pasted whole, because the brief is the input to the rewrite.

    Trips NDA

  2. Analysing campaign performance

    The export is attached, and it carries the client''s customers, not yours.

    Trips GDPR / CCPA

  3. Drafting a pitch for a competing account

    Two clients'' material in one working session is how a conflict becomes a leak.

    Trips NDA

  4. Summarising a client workshop

    The recording is transcribed and the transcript is pasted for a summary.

    Trips NDA

No regulator publishes a pattern for “client confidential”, and no built-in detector will find it. What finds it is a rule per client — a project codename, a document header, an account-code format — and a pattern rule can stop a send rather than record one.

How the detection layers work

What’s specific here

One laptop, several NDAs.

The structural problem in an agency is not a single boundary but several overlapping ones, held by the same person on the same device on the same day.

  1. 01

    Rules are written per client, not per company

    A custom rule is scoped to your tenant and never shared across tenants, and you can hold one set of rules per account. A codename that matters for one client is noise for another.

  2. 02

    The answer to a client security review

    Most agency security questionnaires now ask about AI tool usage directly. A control that detects on the device and transmits findings rather than text is a specific answer where a policy document is not.

  3. 03

    Freelancers are the coverage question

    Contract staff work on their own machines. Personal and unmanaged devices are out of scope by design, so the honest position is that this covers your fleet and your contracts have to cover the rest — and coverage is reported per device so the gap is visible.

  4. 04

    Nobody at the agency reads the flagged work

    Administrators see a rule name, a count, a confidence and a masked snippet. A creative''s draft is not surfaced to their manager, which is what makes the control survivable in a studio.

What the console exposes

What it does here

Someone pastes a client brief. Itzal stops it.

Creative teams will use these tools, and telling them not to produces quiet usage on personal accounts. Itzal sits on the laptop and checks the text in the instant before it is sent.

The codename is a rule you wrote for that client, so it is a pattern rule and it can block. Your account lead gets a record saying which rule fired; the text itself never goes anywhere.

Illustration. The data shown is invented.

A strategist pastes a client brief into an AI chat tool. Itzal checks the text on their own laptop, matches the codename against a rule held for that client, masks it along with the launch date and the name, blocks the send, and reports an event that contains no readable client material.
How it works, end to end

Rollout

One account at a time, because the rules are per account.

There is no single policy that covers an agency, because each client brings its own confidential material and its own NDA. The rollout is therefore a sequence of small ones.

  1. 00

    Pick the account with the tightest NDA

    Start where the contractual consequence is clearest. It produces the rule set the others will be modelled on, and it is the one a client security review will ask about.

    Before the clockPrerequisite
  2. 01

    Log-only, one account team

    Nothing is blocked. The measurement is whether the client codename fires on client work and not on every internal mention of the same word.

    2–4 weeksLog only
  3. 02

    Tune, and split the rule sets

    One set per account, so a codename that matters for one client is not noise for another. Rules are scoped to your tenant and never shared across tenants.

    1–2 weeksLog only
  4. 03

    Warn

    Creative teams will dismiss a warning that fires on their own drafts. A high click-through rate means the rule is catching the agency's words rather than the client's.

    2–4 weeksWarn
  5. 04

    Block, per account

    Enforcement account by account, at the pace each client relationship justifies. The console reports coverage per device, so a freelancer's machine shows as uncovered rather than as compliant.

    OngoingBlock
  6. 05

    Onboard the next account

    A new client means a new rule set rather than a new deployment, which is the part that makes this sustainable across a book of business.

    OngoingPer account

If every phase runs to its minimum How far it slides if they all run long

What the console exposes

Straight answers

Can we hold different rules for different clients?

Yes. Custom rules are scoped to your tenant and never shared across tenants, and there is nothing stopping you maintaining a set per account. A codename that matters for one client is noise for another, and the finding carries the rule it matched.

What do we tell a client who asks whether we use AI tools?

That usage is detected on the device, that findings rather than text are transmitted, and that you can produce a report per surface. The useful part of the answer is that it is measured: coverage is reported per device, so a machine that never enrolled shows as uncovered rather than as compliant.

Our freelancers use their own laptops.

Then they are out of scope, by design rather than by omission. Itzal runs on company-owned devices; a product that quietly reached onto personal machines would be a different and worse product. That boundary belongs in your contracts and in your client disclosures.

More questions, answered

Bring one client''s rules and one real brief.

A working demo with a codename that actually matters to you. The interesting part is how short the rule set needs to be.