Faster answers for citizens and customers, with a person within reach.
We start with the people who answer: suggested replies, case context and summaries for contact-centre staff. Self-service follows where it is safe, with clear limits and an escalation path that hands over the full conversation.
For service organisations answering people in several languages and countries, on every channel at once.
A contact arrives on any channel. The system detects intent, language and sensitive topics and looks up the case. Sensitive or out-of-scope contacts are escalated to a specialist with full context. Otherwise the assistant drafts a reply and proposes next actions, and a service employee edits and sends it. Everything is recorded in the CRM and sampled for quality review.
Before and after
What changes in the working day
Today
With the system
Today
Today: Searching while the caller waits
Service staff switch between the CRM, the case system, the knowledge base and old emails to answer a single question.
With the system
With the system: Context on arrival
The case status, recent contacts and the relevant knowledge articles are on screen before the employee says hello.
Today
Today: Wrap-up eats the day
Every contact ends with minutes of typing a summary and choosing codes, often from memory.
With the system
With the system: Drafts, not decisions
The assistant drafts the reply and the summary; the employee edits, sends and stays responsible.
Today
Today: Bots that trap people
Scripted chatbots loop on questions they cannot handle, and people arrive at a person already frustrated, repeating their story.
With the system
With the system: No dead ends
Self-service hands over to a person with the full conversation and the detected intent, whenever it reaches its limits or someone asks.
Today
Today: One language at a time
Queues are split by language, so waiting times depend on who is on shift.
With the system
With the system: Multilingual by design
Replies are drafted in the language of the contact, in plain language, and checked by the employee before sending.
Workflow
How the workflow runs
Assisted replies come first: the assistant prepares, a person sends. Self-service is added later for well-defined questions, under the same controls and with the same escalation path.
Illustrative workflow
7 steps · 2 human checkpoints
Legend
System
AI agent
Person
Human checkpoint
Exception path
Service assistance: illustrative workflow with escalation path
Read the diagram as text
The main path runs from a contact arriving on any channel, through intent and context detection and the assistant drafting a reply, to a service employee reviewing and sending it. The contact is then recorded in the CRM and sampled for quality review.
Before anything is sent, a person reviews the draft and approves any proposed action: that is the human checkpoint on the main path.
Contacts flagged as sensitive, complex or out of scope leave the main path after intent detection and go to a specialist with the full context. They rejoin the main path when the contact is recorded.
01
Step 1: Contact on any channelSystem
Phone (with live transcription), chat, email, portal messages and scanned letters arrive in one queue. The contact is identified through the portal login or eID where the request needs it.
02
Step 2: Intent and contextSystem
Intent, language and urgency are detected; topics that must always reach a person (complaints, vulnerability, legal disputes) are flagged. Case status is looked up through authenticated APIs.
03
Step 3: Escalate to a specialistPerson
Sensitive, complex or out-of-scope contacts go to a specialist queue with the transcript, detected intent and case data attached, so nobody has to repeat their story.
Exception path: Sensitive, complex or out of scope· Returns to step 6
Human checkpoint
The assistant never gives legal advice or takes decisions on entitlements. Escalation rules are set by the service owner, not by the model.
04
Step 4: Draft reply and next actionAI agent
The assistant drafts a reply from approved knowledge, in plain language and the contact’s language, and proposes actions it is allowed to prepare: an appointment, an address change, a status update.
05
Step 5: Service employee reviews and sendsPerson
The employee edits the draft, confirms or rejects the proposed actions and sends the reply on the original channel.
Human checkpoint
Nothing is sent and no action is executed without a person’s approval. Edits are captured to improve the drafts.
06
Step 6: Record and wrap upSystem
A summary, contact reason and outcome are written to the CRM or case system, and the conversation is retained under the agreed retention rule.
07
Step 7: Quality reviewPerson
Team leads review sampled conversations and drafts, mark issues and update knowledge articles. Changes go through the evaluation set before release.
Components
What we would build
01
Channel and telephony integration
Connections to the contact-centre platform, chat, email and portal, with live transcription for calls and one conversation record across channels.
Case data is fetched only after authentication and only for the person in the conversation. The drafting assistant works with a short list of actions, each with its own permission, and cannot change records directly.
Human review
In assisted mode, every reply and action is approved by an employee. In self-service, escalation triggers are explicit: sensitive topics, low confidence, repeated questions or a simple request for a person.
The assistant does not give legal advice and does not decide on entitlements, benefits or claims.
Auditability
Conversations, drafts, edits, proposed and executed actions are logged with the model and prompt version, so a complaint can be reconstructed step by step.
Data protection
Transcripts and logs follow retention rules agreed with your data protection officer. Personal data is masked in analytics and never used to train external models.
AI transparency
People are told when they are interacting with an AI system, as the EU AI Act requires, and can ask for a person at any point. Drafted replies sent by employees remain the organisation’s communication.
Measures
What we would measure
We agree these measures with you during discovery and compare them with the same queues before the assistant is introduced.
What we would measure
Metric
Why it matters
How we would measure it
01First-contact resolution
Why it mattersThe clearest sign that people got what they needed without coming back.
How we would measure itRepeat contacts on the same case within an agreed window, from CRM data.
02After-contact work time
Why it mattersSummaries and coding are where assistance saves effort without touching the conversation itself.
How we would measure itWrap-up time per contact from the contact-centre platform, by queue and channel.
03Escalation quality
Why it mattersAn escalation is only good if the specialist does not need to start again.
How we would measure itSpecialist rating of the handover context and the share of escalations needing a call-back for missing information.
04Satisfaction
Why it mattersSpeed means little if people feel unheard.
How we would measure itPost-contact surveys and complaint themes, compared per channel and language.
No targets are set before a baseline exists.
Rollout
How we would roll it out
Phase 01
Discovery
Contact reasons, channels, languages and systems. We pick the high-volume, low-risk intents and define topics that must always reach a person.
Exit criteria
Intents and escalation rules agreed
Baseline measured per queue
Data protection review started
Phase 02
Assisted replies pilot
Context panels, drafts and summaries for one team, on one or two channels. Employees send everything.
Exit criteria
Draft quality meets agreed criteria
Staff trained and asked for feedback
Logging and retention in place
Phase 03
Controlled self-service
Selected intents answered directly on chat or the portal, with escalation to the same team and full context.
Exit criteria
Escalation tested end to end
AI disclosure in place
Containment reviewed against complaints
Phase 04
Scale
More intents, channels and languages, each passing the same evaluation gate.
Appointment, reimbursement and administrative questions, with medical questions always routed to qualified staff.
Illustrative example
One service desk, five languages
Situation
A pan-European service desk answers customers in five languages from three locations. Queues are split by language and after-contact work takes a large part of each shift.
System
The assistant drafts replies and summaries in the customer’s language from one shared knowledge base, inside the existing CRM console. Self-service on the portal covers order and status questions.
Human control
Employees send every assisted reply. Complaints, cancellations and vulnerable customers always go to a person, with the conversation attached.
What we would measure
After-contact work time and first-contact resolution per language, and specialist ratings of escalation handovers.
International
Service across languages and countries
Multinational service teams and EU-wide public services answer the same questions in several languages, under different national rules. We design one assistance layer with knowledge and escalation rules per country, so a contact in German, French or Dutch gets the same quality of answer and the same route to a person.
The EU AI Act requires people to be told when they interact with an AI system, and the European Accessibility Act sets requirements for customer-facing digital services. Both shape the design from discovery onwards, together with GDPR retention rules for transcripts and logs.
No. We start with assisted replies, so employees spend less time searching and typing and more time on the conversation. Self-service is added only for well-defined questions, and a person stays one step away.
02How do you balance self-service and escalation?
We design escalation first. Every self-service intent has explicit exit conditions and a handover with full context. We measure the quality of escalations and repeat contacts, not just how many conversations the assistant closes.
03Can it serve people in Dutch, French, German and English?
Yes. Drafts are produced in the contact’s language and in plain language. For public services we check readability, and where required we support easy-to-read variants, reviewed by your staff.
04Do we have to tell people they are talking to an AI system?
Under the EU AI Act, people must be informed when they interact directly with an AI system. We design the disclosure, the option to reach a person and the logging in from the start.
05Does it work with our existing CRM and contact-centre platform?
That is the intention. We build inside your service console where possible, for example as components in Salesforce Service Cloud, and integrate with your telephony and case systems through their APIs.
06Where do we start?
With your contact reasons. The most frequent, low-risk intents make a good first scope for assisted replies. See how we approach AI Agents & Automation.
Discuss your service channels
Share your top contact reasons, channels and languages. We will propose where assisted replies help first and where a person must stay in front.