Multilingual Support AI Agent

What this agent does

Translates the reply, not which version governs

A reply is drafted, and the language owed is read from where the customer lives, not what they prefer.

A language is owed under Quebec s. 50.2, which binds an enterprise offering goods or services.

Reason

A reply carries cancellation terms, and s. 55 wants French remitted before another language binds.

A reply goes out in the wrong language, and s. 204.23 leaves it dead in your hand, live in theirs.

A language rule reaches the help centre, since s. 52 now opens Regardless of the medium used.

Decide

A reply crosses into Poland or Flanders, where the local text is the operative one.

A reply is approved by a named localisation owner, never by this one, and the version sent is kept.

Out

A source and its translation stay paired, so the version the customer received is on the record.

Execute write actions only inside the approval boundaries agreed during implementation.

→ Product statement

Translating, pairing the source with the version sent and flagging a language a rule refuses belong to the agent. Which version governs belongs to a named localisation owner.

Example workflow

One reply, inbound to sent

AgentHuman

1Message arrives at the deskWeb form, email, chat, an in-app session or a thread routed in from another queue

2Language and jurisdiction readWhere the customer resides, the language owed there and the language the message came in

3Reply drafted and translatedThe source text, the translation, the flagged lines and confidence

4Controls appliedJurisdiction rules, document-class checks, source-to-translation checks and confidence threshold

No human action required

Stages 1 to 4 run unaided, and nothing is sent at any of them — the agent is translating, and the owner lane opens at the confidence gate.

5DecisionBranches at the confidence threshold

High confidence

Goes to the named localisation owner.

Low confidence

Adds a native-language read first.

Localisation owner review

The reply is held with its source text, its translation and the language owed.

Approve · Amend the translation · Send to reviewer

Approved — by the named owner▼

6Conversation record updatedOnly where write access and records policy allow it

7Outcome evaluatedLanguage coverage, amendment outcomes, version pairing and what review found

Amendments

Each localisation-owner amendment is counted in the evaluation.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required8 items

Deciding which language version governs.

Signing off a translated contract term.

Certifying that a translation is accurate.

Issuing a legally operative document.

Automation boundaryAgent acts unaided

✓Translate the reply and keep the source beside it.

✓Hold the version the customer actually received.

✓Flag a reply about to leave in a language the configured rule refuses.

✓Mark anything a named owner must approve before it leaves the desk.

Nothing operative is sent here; a named localisation owner decides every one of those.

Judging whether a French version was remitted.

Telling a customer which text a court will read.

Choosing the language a jurisdiction owes.

Changes to language, jurisdiction or approval rules.

Example output

One translated reply, annotated

This serves a desk that may have to show which language version a customer actually received, long after the thread closed; below is one reply exactly as the agent leaves it.

Translation output · single replyIllustrative example

Conversation

Source line

Language owed

Version of record

Confidence

Approved by

Inbound, Quebec resident

Your subscription ends at the close of the current term

French, read from residence

French drafted, English held

Held unsent

The named localisation owner, by name

As receivedTaken from the thread as it arrived — nothing on this side is written by the agent.

What the record holds Source text Paired translation Jurisdiction rule

Why nothing left hereWhich version governs is a call a named localisation owner makes.

ActionApproveAmend the translationSend to reviewer

What the score decidesBelow the configured threshold the reply picks up a native-language read before the owner.

Value

Where AI adds value

The same four claims, placed at the point in the workflow where each one applies.

Where the value landsValue 01 – 04

Every reply that leavesFrom the language owed

03Translation

Read the rule, then hand over

The media-entertainment accessibility certification agent works on captions and access modes; this one is about which language version a court reads.

01Approved path

A translation can be the text

A document that breaches the Charter may not be invoked by its author, but may be invoked against the author — dead in your hand, live in theirs.

02Human review

What was checked, and not found

Checked across the jurisdictions surveyed: no case, statute or regulator statement anywhere decides whether a machine-translated communication binds the company, and nothing obliges you to label a reply as machine-translated. No translation is signed, certified, filed or entered in any register.

04Build an evidence trail

The reply, the language it was owed in and the version the customer received stay together.

Integrations

Typical integrations

Five system groups connect to the same agent. Which of them are in scope is decided in discovery.

Ticketing and support desksZendesk · Intercom · Salesforce
Conversation and locale fields

Translation servicesDeepL · Google Cloud Translation
Machine and human output

Knowledge and help centreZendesk Guide · Document360
Published articles and their translations

Agent

Multilingual support

Reads the language
Pairs source and translation
Holds for the owner

Contract and document systemsDocuSign · Ironclad · Conga
Terms, warranties and related documents

Observability & evaluationOpenTelemetry · Langfuse
Supported monitoring/evaluation sources

Integration availability depends on the client's existing systems and API access.

Agent controls

Six layers between the model and the reader

Six locks down one staircase, the last the deepest. What descends is set out in the map below.

L6 · Outermost — last line of defenceInward → L1 · closest to the model

L6Rollback / safe modeFall back to the required language alone when evaluation or production signals degrade.Roll back

L5Version monitoringTrack model, prompt, glossary and jurisdiction-rule configuration changes.Track

L4TraceabilityRecord the source text, the translation, the language owed and the version sent.Record

L3Owner approvalHold the reply for a named localisation owner; the hold governs sending, not whether the translation reads right.Gate

L2Policy guardrailsTest each reply against the configured jurisdiction rules; a language a rule refuses returns the reply.Restrict

L1Confidence thresholdsRoute a weak translation to a native-language reviewer before the owner sees it.Require review

Model coreReply translated — the source text, the translation, the language owed and confidence

L1 – L2Test whether a reply may be sent

L3Puts the sending in an owner's hands

L4 – L5Keep the reply and the language behind it

L6Answers in the required language only when signals degrade

How Nestack evaluates it

Evaluate the whole translation — not only the sentence that goes out.

Coverage runs the whole depth of the workflow, and every layer is cut by slice.

Surface — the reply the customer reads

Depth of coverage ▼

E1Final-output evaluationDid the reply leave in the language that jurisdiction owes?

E2Step-level evaluationDid the agent read the right residence, the right rules and the live glossary?

E3Tool evaluationDid it call the correct translation route and write the correct thread?

E4Confidence calibrationDo weak translations actually attract more owner amendments?

E5Slice evaluationHow does performance change across specific jurisdictions?

E6Business outcomeHow many replies needed an amendment before the owner approved?

Floor — the version the company answers for

Failure modes

Where each failure originates in the agent

Seven failure modes, each placed where the language first goes wrong.

Agent lifecycleDirection of processing →

01 · Retrieval1 mode

PK-03

Rule read from preference

The rule is picked by preference, not by residence.

Stage gathersThe thread, the residence and the language owed

02 · Reasoning2 modes

PK-04

Translated term drifts from source

The source moved and the translation did not.

PK-06

Weak translation passed forward

A thin rendering moves on without the read.

PK-05

Version bound to the wrong thread

The translation is filed against another reply.

Stage writesOnly where write access and approval policy allow it

04 · Output1 mode

PK-01

Reply sent in the wrong language

The language owed there was not the one used.

Stage returnsThe version the customer received, and its source

05 · Change / Version1 mode

PK-07

Silent re-translation

A reply is regenerated and the sent version is gone.

Stage tracksModel, prompt, language rules and glossaries

Sev-1 · a document sent in the wrong language Sev-2 · a translated term no longer matches Sev-3 · signals degrade, reply held back

See the complete Customer Support failure-mode catalogue →

Affected slices

Quebec documents absorb the amendments

A jurisdiction-level coverage figure can read clean while Quebec adhesion documents carry most of the amendments. Nestack reports the owner-amendment rate by jurisdiction, not only in total.

Slice performance — reported separately, not only in aggregateIllustrative example

Slice	Failure rate	Lift	Status
Quebec adhesion documents	12.6%	3.7×	Review
France consumer replies	9.0%	2.6×	Review
Poland and Flanders threads	5.6%	1.6×	Watch
Single-language markets	2.6%	0.8×	Normal

Bar: amendment-rate lift vs. single-language-market baseline · scale 0–4.0× · tick marks the 2.0× review threshold 2 of 4 slices over threshold

Evidence-linked improvement

What a wrong-language reply costs

A cycle shuts when the reply sent in the wrong language is a regression case. That suite is what the next message issued is measured against.

Improvement cycle · five stagesSwitchback — the path turns at Improve and returns at Learn

01Detect

Amendment rate rises in one jurisdiction.

02Diagnose

The cancellation terms the customer read in their own language, which the company cannot now rely on, are worked backwards until one cause is left standing.

03Improve

A change goes out numbered, and the replies that forced it travel with it.

04Verify

Nothing ships while one language case remains red.

05Learn

It is retained permanently, and the language rules move with it.

Learn → DetectThe return edge. The next reply is measured against a suite one case longer.

Typical build scope

Twelve workstreams across six weeks

The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, sources, translation workflow, evaluation, integration, then production validation and handover.

Workstream Week 1Week 2Week 3Week 4Week 5Week 6

01Language-obligation and automation-boundary mapping.

02Ticketing, translation and document feeds.

03Jurisdiction-to-language and document-class mapping.

04Thread and language reading.

05Translation routing and source pairing.

06Confidence scoring and reviewer routing.

07Localisation owner review workflow.

08Ticketing-and-document integration.

09Language and version cases.

10Guardrails and translation controls.

11Language-trail instrumentation.

12Deployment, documentation and Agent Care handover.

12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel Final scope and sequence confirmed in discovery

Engagement tiers

What each tier includes

Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.

Capability✓ in scope · — not at this tier PilotOne language pair, one queue ProductionProduction support systems AdvancedMultiple languages / jurisdictions

Introduced at Pilot

Translation to your jurisdiction rules✓✓✓

Localisation owner approval✓✓✓

Language-coverage baseline✓✓✓

Introduced at Production

Reporting by jurisdiction—✓✓

Owner review workflow in your systems—✓✓

Approved write-back—✓✓

Translation-service integration—✓✓

Introduced at Advanced

Multi-jurisdiction document rules——✓

Multi-stage owner approvals——✓

High message volume——✓

Multi-jurisdiction language controls——✓

Build price From $5,000 From $8,000 Custom quote

Final build priceConfirmed after discovery based on integrations, workflow complexity, message volume, approval controls and deployment requirements.

Separate from buildBuild pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.

What we need from you

What you bring, and what we build with it

Each input maps to a piece of build scope and a week in the delivery timeline.

You bringWe build with it

01Your markets and the language each one owes → Jurisdiction mapping and language bindingWeek 1

02Representative threads from each language → Thread ingestion, translation routing and the language baselineWeek 2

03The documents your desk sends and who owns each → Jurisdiction mapping, document classes and the automation boundaryWeek 1

04Access to relevant APIs, feeds or exports → Ticketing, translation and document assessment, then integration setupWeek 2

05Replies you would not want construed → Version cases and the evaluation runWeek 4

06What no translation may correct → Confidence scoring, reviewer routing, guardrails and release controlsWeek 3

07A named localisation owner who approves the reply → Localisation owner approval workflow, then pilot and production validationWeeks 5–6

Nothing else is required Deployment, documentation and Agent Care handover are ours.

Delivery timeline

Four phases across six weeks

Read the bands as effort rather than layout; the fifth week carries a pair because the work does.

Phase W1W2W3W4W5W6

Discovery W1

Build W2 – W3

Evaluate W4 – W5

Pilot & Launch W5 – W6

Week focus W1Language-obligation discovery, jurisdiction mapping and the automation boundary W2Ticketing and translation integration and the language baseline W3Translation routing, confidence logic and approval controls W4Evaluation suite, language cases and failure-mode testing W5Document-system integration, pilot threads and targeted corrections W6One support quarter run under the localisation owner, then Agent Care handover

Reading the bandEach bar covers only the weeks its own work is named for, and week five is shared by design.

At the end of W6Once the language record validates, Agent Care takes the agent on.

DurationSix-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Customer Support AI agent

Build a multilingual support agent around the version your customer actually received.

Show us the languages your desk answers in and the documents those replies carry. In Quebec the French version is the one that governs, and a document breaching the Charter may not be invoked by its author. Bring one reply you could not now reproduce as it went out.

Discuss this agent →

Nestack Agents · Multilingual supportAGT-SUP-07 · Agent Care available after launch