Support AI agent · Multilingual support
Multilingual Support AI Agent
Serve each customer in the language the jurisdiction owes them: translate the reply, keep the source paired with it, and leave which version governs to a named localisation owner.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Named owner
- Version held
- After launch
- Agent Care
What this agent does
Translates the reply, not which version governs
In
- 01
A reply is drafted, and the language owed is read from where the customer lives, not what they prefer.
- 02
A language is owed under Quebec s. 50.2, which binds an enterprise offering goods or services.
Reason
- 03
A reply carries cancellation terms, and s. 55 wants French remitted before another language binds.
- 04
A reply goes out in the wrong language, and s. 204.23 leaves it dead in your hand, live in theirs.
- 05
A language rule reaches the help centre, since s. 52 now opens Regardless of the medium used.
Decide
- 06
A reply crosses into Poland or Flanders, where the local text is the operative one.
- 07
A reply is approved by a named localisation owner, never by this one, and the version sent is kept.
Out
- 08
A source and its translation stay paired, so the version the customer received is on the record.
- 09
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
- 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.
Goes to the named localisation owner.
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 required
8 items
- Deciding which language version governs.
- Signing off a translated contract term.
- Certifying that a translation is accurate.
- Issuing a legally operative document.
Automation boundary
Agent 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 reply
Illustrative example- Conversation
- Inbound, Quebec resident
- Source line
- Your subscription ends at the close of the current term
- Language owed
- French, read from residence
- Version of record
- French drafted, English held
- Confidence
- Held unsent
- Approved by
- The named localisation owner, by name
As received
Taken 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 here
Which version governs is a call a named localisation owner makes.
What the score decides
Below the configured threshold the reply picks up a native-language read before the owner.
Action
ApproveAmend the translationSend to reviewerValue
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value lands
Value 01 – 04
Every reply that leaves
From the language owed
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.
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.
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.
Build 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 desks
- Zendesk
- Intercom
- Salesforce
- Conversation and locale fields
Translation services
- DeepL
- Google Cloud Translation
- Machine and human output
Knowledge and help centre
- Zendesk Guide
- Document360
- Published articles and their translations
Agent
Multilingual support
- Reads the language
- Pairs source and translation
- Holds for the owner
Contract and document systems
- DocuSign
- Ironclad
- Conga
- Terms, warranties and related documents
Observability & evaluation
- OpenTelemetry
- 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 defence
Inward → L1 · closest to the model
- L1 – L2
- Test whether a reply may be sent
- L3
- Puts the sending in an owner's hands
- L4 – L5
- Keep the reply and the language behind it
- L6
- Answers 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
- 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 lifecycle
Direction of processing →
01 · Retrieval
1 mode
- PK-03
Rule read from preference
The rule is picked by preference, not by residence.
Stage gathers
The thread, the residence and the language owed
02 · Reasoning
2 modes
- PK-04
Translated term drifts from source
The source moved and the translation did not.
- PK-06
Related document sent unremitted
Another language goes out with no French first.
Stage proposes
The source, the translation and the language owed
03 · Tool / write
2 modes
- PK-02
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 writes
Only where write access and approval policy allow it
04 · Output
1 mode
- PK-01
Reply sent in the wrong language
The language owed there was not the one used.
Stage returns
The version the customer received, and its source
05 · Change / Version
1 mode
- PK-07
Silent re-translation
A reply is regenerated and the sent version is gone.
Stage tracks
Model, 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
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 aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | 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 stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Amendment rate rises in one jurisdiction.
- 02
Diagnose
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.
- 03
Improve
A change goes out numbered, and the replies that forced it travel with it.
- 04
Verify
Nothing ships while one language case remains red.
- 05
Learn
It is retained permanently, and the language rules move with it.
Learn → Detect
The 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.Week 1
- 02Ticketing, translation and document feeds.Week 1, Week 2
- 03Jurisdiction-to-language and document-class mapping.Week 1, Week 2
- 04Thread and language reading.Week 2
- 05Translation routing and source pairing.Week 2, Week 3
- 06Confidence scoring and reviewer routing.Week 3
- 07Localisation owner review workflow.Week 3, Week 4
- 08Ticketing-and-document integration.Week 5
- 09Language and version cases.Week 4, Week 5
- 10Guardrails and translation controls.Week 4
- 11Language-trail instrumentation.Week 5, Week 6
- 12Deployment, documentation and Agent Care handover.Week 6
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 price
- Confirmed after discovery based on integrations, workflow complexity, message volume, approval controls and deployment requirements.
- Separate from build
- Build 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 bring
We build with it
01Your markets and the language each one owes
→Jurisdiction mapping and language binding
Week 102Representative threads from each language
→Thread ingestion, translation routing and the language baseline
Week 203The documents your desk sends and who owns each
→Jurisdiction mapping, document classes and the automation boundary
Week 104Access to relevant APIs, feeds or exports
→Ticketing, translation and document assessment, then integration setup
Week 205Replies you would not want construed
→Version cases and the evaluation run
Week 406What no translation may correct
→Confidence scoring, reviewer routing, guardrails and release controls
Week 307A named localisation owner who approves the reply
→Localisation owner approval workflow, then pilot and production validation
Weeks 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
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – 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 band
- Each bar covers only the weeks its own work is named for, and week five is shared by design.
- At the end of W6
- Once the language record validates, Agent Care takes the agent on.
- Duration
- Six-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.
- Typical delivery
- 4–6 weeks
- Deployment
- Your languages
- Named owner
- Version held
- After launch
- Agent Care available
Nestack Agents · Multilingual support
AGT-SUP-07 · Agent Care available after launch