A public service phone line is a rationing mechanism. It always has been. Every queue, hold time and opening hour decides who gets through and who gives up, and that decision has never been evenly distributed.
This matters because of a framing that has become common: that adding an automated or multilingual channel is an efficiency improvement, a neutral upgrade that helps everyone a bit. It is not neutral. A change to the front door is a change to who gets through it. When a new channel works better for some callers than others, it does not just add capacity — it redistributes access to whatever sits behind the switchboard.
The honest question is therefore not "can AI answer the phone in more languages?" It is: when we automate the front door, who ends up on the other side?
The funnel, and where people fall out
Access to a public service is not one step. It is a sequence, and every stage loses people.
1. Awareness. You cannot use a service you have not heard of. IRCC's own 2023 Settlement Outcomes Report found 32% of newcomers were unaware of the free settlement services available to them.
2. Channel access. The channel has to be reachable. The CRA's ethnographic research found low computer literacy made the online channel inaccessible for many newcomers it interviewed — only three of 48 had ever used the CRA website.
3. Language. The CRA study found 69% of the newcomers it interviewed required an interpreter to complete the conversation at all.
4. The queue. The Auditor General of Canada reported in October 2025 that the CRA received more than 32 million calls in 2024-25 and that just over 10 million reached an agent. 8.6 million calls were deflected — no option to speak to a person was offered. Callers who did connect waited an average of about 31 minutes, and roughly 12% waited more than an hour.
5. Comprehension of the answer. Getting through is not the same as getting the answer. The Auditor General placed 167 test calls to CRA agents and found responses to general individual-tax questions were accurate 17% of the time.
Two observations follow. First, the funnel leaks at every stage, and language is only one leak. Second, an organisation can improve stage three while doing nothing about stages one, two, four or five. A multilingual voice assistant addresses the language step. It does not create awareness, it may not fix digital exclusion, and it does not make the queue shorter or the answer correct.
Why better access is not the same as fairer distribution
Here is the trap.
Introducing a new channel usually serves, first and best, the people who were already getting through. They know the service exists. They already call. They have a phone, a signal, and the confidence to try. A smoother channel makes their experience better — which is genuine value, and worth having.
But the people who were falling out of the funnel at stage one or two are not reached by improving stage three. You have upgraded the experience of the people inside the system, not widened the system.
The CRA evidence makes this concrete. Only one of 48 newcomers interviewed had ever personally called the agency. They routed instead through settlement workers, private sponsors, friends and family — because telephone removes the non-verbal cues that help when confidence is low, and because a trusted intermediary is easier to trust than an institution. Build a better telephone line and those 47 people still never call it.
So the first honest test of any deployment is not "did call volume rise?" It is: did the people who previously fell out now get through?
The two-tier risk
Now add a legal detail that shapes everything.
Canada's Official Languages Act obliges federal institutions to serve the public in English and French, with equal quality and an active offer. There is no equivalent duty for any other language. Multilingual service beyond those two is discretionary.
That combination produces a specific hazard. If an agency introduces an automated multilingual channel while its legal obligation remains bilingual, the likely shape of the service is:
- A fast, cheap, automated channel — handling the people whose languages the system performs well in, plus anyone the agency would rather route away from a human.
- A slow, expensive, human channel — still delivered in English and French, because that is what the law requires.
Whether that is progressive or regressive depends entirely on who lands where. If the automated channel serves people who would otherwise have waited 31 minutes, it is a win. If it becomes the place where low-English-proficiency callers are quietly parked while English and French speakers reach a person, then a two-language legal guarantee has become a two-tier service, and the tier you are on correlates with how recently you arrived.
That is not a hypothetical. It is the default outcome of optimising for cost without measuring for equity — which is why the measurement question below is not a nice-to-have.
What Canada's rules require — and where they stop
Canada is not starting from nothing. The Treasury Board's Directive on Automated Decision-Making applies to federal automated decision systems developed or procured after 1 April 2020, and it requires a genuinely serious set of things:
- an Algorithmic Impact Assessment before production, published on the open government portal;
- testing for unintended data bias before launch;
- validation that the data used is relevant, accurate and current;
- a notice of automation delivered through every service channel in use — explicitly including telephone;
- a meaningful explanation to affected individuals of how and why a decision was made;
- recourse for people affected by a decision; and
- peer review by a qualified third party.
It was amended on 25 April 2023 to extend scope to internal services, require the impact assessment to be published before launch rather than after, strengthen bias testing and data governance, and add questions about impacts on people with disabilities.
That is a stronger framework than most countries have. It also has three gaps that matter for voice automation specifically.
First, it excludes the Canada Revenue Agency — the operator of the country's largest contact centre, and the subject of the Auditor General's most alarming 2025 findings. The agency handling 32 million calls a year sits outside the directive written for automated systems.
Second, it does not bind provincial or municipal governments. Healthcare, education, social services and 311 lines are largely provincial and municipal. A city deploying a multilingual voice agent for service requests is, in most cases, outside this framework entirely.
Third, and most subtly, it covers decisions rather than routing. The Directive applies to systems that make or assist in making an administrative decision about a client. Treasury Board's own scope guidance puts a chatbot that merely "directs a client to government websites" out of scope, while a tool officers use to recommend a course of action is in scope. A voice agent that sorts callers, answers routine questions and decides who waits in which queue sits in that grey zone — doing work with real distributional consequences, potentially without triggering any of the safeguards above.
The precedent worth remembering
The cost of getting this wrong is not abstract. The Dutch childcare benefits scandal is the case study every public body should study before automating a benefit or service channel.
Beginning in 2013, the Dutch tax authorities used an algorithmic system to flag childcare benefit claims as potentially fraudulent. Nationality was used as a risk factor, so non-Dutch nationals received higher risk scores. A self-learning mechanism then reinforced the pattern — the algorithm adapted on its own experience, with no meaningful human oversight, producing what investigators described as a discriminatory loop.
Tens of thousands of parents and caregivers, most of them low-income, were falsely accused of fraud. Ethnic minorities were disproportionately affected. Benefits were suspended and families were subjected to hostile investigations governed by rigid rules and aggressive recovery. The documented consequences included debt, unemployment, forced evictions, mental health harm and family breakdown.
Two design failures made it worse. The system was opaque — civil servants were required to review flagged cases but were given no reason for the risk score, so the review was nominal. And there was a perverse incentive to pursue recovery aggressively, because the authority had to demonstrate the system's efficiency. Amnesty International's assessment described a system that produced what it called "a black hole of accountability."
The Dutch government resigned in January 2021.
Note what is not in that story: no misbehaving model, no dramatic technical failure. The system did what it was built to do. The failures were in what it was asked to do, what data it was fed, and the absence of anyone able to explain or override it.
What the automation layer does and does not own
There is a version of "we'll connect it with a workflow tool" that quietly assumes the orchestration layer is also the system of record. It is not, and the gap between those two things is where public deployments get into trouble.
A workflow tool such as n8n is an integration and orchestration layer. It moves a request from a conversation into a system, and it can be very good at that. It is not an identity provider, an audit platform, a records system or a case-management database — and treating it as one produces failures that look like technical glitches but are actually governance gaps.
A public deployment normally needs, around the orchestration layer rather than inside it:
- durable queues and retries, so a dropped webhook does not lose a citizen's request silently;
- idempotency, so a repeated or retried submission does not create duplicate cases;
- role-based access and audit logs, recording who and what touched a record, and when;
- secrets management, monitoring, and a disaster-recovery plan; and
- formal change control, because a workflow edited on a Friday afternoon is a change to a public service.
The integration itself is usually the underestimated part. Legacy systems may have no API at all, inconsistent identifiers between departments, rate limits, or owners in another business unit entirely. Screen automation against a system built in the 1990s is not a workflow-building exercise. Every connection needs error handling, reconciliation, duplicate prevention, a rollback or compensation step for partial completion — and a named person who owns it when it breaks.
One rule deserves to be absolute. A citizen should receive a reference number only once the system of record has confirmed the case was created. A confirmation sent because the workflow reached the end of its chain — rather than because the government system durably accepted the request — is worse than no confirmation at all. It hands someone a number that does not exist, and they stop chasing the thing they actually needed.
What would make a deployment defensible
If a public or publicly funded organisation is going to automate its front door, the following are the minimum conditions — and they are conditions, not features.
1. Measure outcomes per language and per group, not in aggregate. Reaching a human, wait time, and whether the caller obtained what they were entitled to. An aggregate "calls handled" figure will conceal exactly the inequality the deployment was meant to reduce.
2. Test accuracy per language before launch, with real speakers. Not the vendor's supported-language list. Real speakers, real accents, your actual caller population.
3. Publish what the system does and how it is assessed. Canada's Directive already requires an impact assessment and a notice of automation. Do it even if it is not technically mandatory for your body — the CRA exclusion and the provincial gap are loopholes, not permission.
4. Guarantee a human path, and instrument it. Callers must be able to reach a person, and you must measure how often each group manages to. Handoff rate is the honest signal that a language is underperforming.
5. Never let a system learn on people without oversight. No self-reinforcing scoring, and no automated decision that a reviewer cannot understand and override. The Directive requires explanation and recourse; the Dutch case shows precisely what their absence produces.
6. Do not use automation to reduce human capacity faster than measured demand justifies. CRA agent headcount fell from 7,782 in 2022-23 to 4,547 in 2024-25 while call volumes rose 30%. Whatever the causes, the result was 8.6 million callers offered no path to a person. Automation deployed into that gap becomes the only door, and the only door is a rationing decision by default.
7. Disclose the machine, and handle the voice data lawfully. Callers should be told they are speaking to an automated system. Voice recordings are personal information under PIPEDA, and under BC's PIPA, Quebec's Law 25 and Ontario's PHIPA in those jurisdictions.
Measuring the wrong thing
Containment — the share of calls resolved without a person — is the metric these systems are sold on, and on its own it is misleading. A high containment rate is entirely compatible with a service that is failing the people it was meant to help.
What would actually indicate whether access improved:
- Successful completion rate, per language and per channel — not calls handled, but outcomes achieved.
- Misunderstanding and repeat-contact rate. A caller who has to call back is a caller the system did not serve, and the rate should be tracked separately by language.
- Time to case creation, plus the incorrect-routing and duplicate-ticket rates — the operational cost of getting intake wrong.
- Escalation and abandonment rates, and resolution time after escalation. The queue does not disappear; it moves.
- Language and accessibility parity. If one language escalates far more than another, the two-tier problem becomes visible here first.
- Staff correction and rework time, because a saving on the phone line that reappears as case repair is not a saving.
- Citizen satisfaction and trust, and privacy or security incidents.
The discipline is the one running through this whole article: measure the outcome per group, or the aggregate will conceal exactly the inequality the deployment was supposed to reduce.
The costs that move rather than disappear
A business case for automating a front door is usually built on call deflection — calls handled without a person, multiplied by the cost of an agent. That arithmetic is real, and it is incomplete, because most of the saving is not deleted. It relocates.
The budget that actually matters includes:
- Volume and peak concurrency — calls per day, average call duration, and how many must be handled at once on the worst day of the year rather than the average one.
- Per-minute and per-message costs — speech recognition, telephony, model inference, transcript storage, SMS, and interpreter minutes where a human interpreter joins the call.
- Integration and its maintenance — building each connection, then keeping it working through upstream changes, for the life of the service.
- Human review and escalation staffing. Escalations are not free, and a system that escalates more often than expected can cost more than the calls it deflects.
- Language and accessibility testing, per language, on an ongoing basis.
- Procurement and security assessment — privacy impact, security review and legal sign-off can be a substantial share of a public-sector deployment's first-year cost.
- The cost of getting it wrong — correcting mis-created cases, handling complaints, staff time repairing bad data, and missed service standards.
That last item is the one most business cases omit, and the Dutch case is the reason it belongs there. A system that answers 80% of calls but creates incorrect cases, discourages vulnerable callers, or leaves staff repairing more work than before is not delivering a better service. It is moving the labour somewhere harder to see and more expensive to fix.
The uncomfortable question
There is a version of this that no efficiency analysis can answer.
If a service is rationed by a queue — and the queue is the mechanism by which a state decides how much attention any one person gets — then automating the queue is not simply making it faster. It is changing the mechanism of rationing. Sometimes that is clearly good. A routine status enquiry answered instantly at 2 a.m. is a real improvement for everyone, and freeing a human to handle a complex, distressing case is better for that person too.
The question worth asking before deployment is not "will this be faster?" It is: when this system decides who gets a human, who wins and who loses, and will we know?
If the answer is "we'll find out from the complaints," that is not a governance framework. It is the Dutch scenario with better intentions.
For the newcomer-specific evidence behind this piece — including the language statistics, official-language obligations and the accuracy limits of multilingual voice tools — see Multilingual AI Voice Assistants and Newcomer Service in Canada. For the privacy obligations that apply to recording and processing voice data, see our guide to AI voice assistants and Canadian privacy.