Shared centres, separated clients

In 2021 a security review by Pellerath Health Software, a prospective client seeking second-line support for its clinic software, found two weaknesses at our Bathurst centre. Screen recordings from four programmes sat in one repository, divided only by application permissions, and agents who moved between an insurer's and a utility's programmes could copy text from one session into the other. Nothing had been misused, but the arrangement was not what Pellerath expected, nor probably what several existing clients assumed. The Pellerath contract was worth about C$10.6 million a year.

I proposed separating every programme: its own virtual-desktop pool, recording store, encryption keys, identity groups and network segment. The chief operating officer disagreed strongly. Blended agents let our schedulers cover peaks across clients, and he estimated that ending the practice would cost three points of occupancy. Our chief financial officer accepted the design but wanted every client to pay for it. The chief executive approved the full programme at C$5.8 million of capital over twenty months, with about C$1.4 million of additional annual running cost.

We kept blending only where both clients agreed in writing, which left three pairs. Two programmes moved centres to simplify the network design. My team built the new environments one programme at a time and tested isolation by attempting cross-programme access from ordinary agent accounts, sharing the results with each client.

By June 2023 all 36 programmes then running were separated. Occupancy at the three most blended centres fell from 83 to 80 per cent, then recovered to 82 per cent after the workforce team rebuilt its forecasts programme by programme. Pellerath signed in March 2023. Only two existing clients accepted a segregation charge; the others regarded separation as something they had assumed already, and we absorbed most of the running cost.

Separation is now the starting point of every new programme design. I would put its cost into renewal pricing before building, rather than asking clients afterwards to pay for a protection they believed they already had.

Our on-premises telephony platform, hosted in data centres in Moncton and Winnipeg, was due to lose its manufacturer's support at the end of 2024. It carried about 4,300 seats and 41 client programmes. The cloud contact-centre provider we selected proposed moving everything across three weekends and offered a C$900,000 incentive for finishing by September 2024. Several clients wanted the new reporting and messaging features quickly.

I recommended migrating one programme at a time over fifteen months. Each client would have its own cutover date, its own test calls and a rollback that could return its traffic to the old platform within an hour. That meant running both platforms in parallel at a cost of about C$2.3 million, plus C$700,000 for extended support of the old system. The chief financial officer opposed both the cost and the lost incentive. Our sales leader worried that clients would read caution as delay. The chief executive accepted the phased plan after we showed which programmes carried card payments or regulated recordings and could not absorb a failed weekend.

The first cutover took place in May 2024. Two programmes rolled back, one because recording metadata failed to reconcile with client reporting and one because a scheduling feed dropped intervals. Both moved successfully a month later. In February 2025 three Winnipeg programmes lost inbound routing for 52 minutes after a carrier configuration error that our failover testing had not covered. We paid C$186,000 in service credits and told each affected client what had happened before they asked.

The last programme moved in July 2025. Annual telephony cost falls by C$3.1 million from 2026, mainly through retired hardware and released data-centre space, and measured availability for 2025 was 99.96 per cent including the outage.

Clients now receive a cutover plan with a named stop authority on both sides. I would test failover with every carrier, route by route, before the first cutover instead of relying on the platform's own redundancy tests.

Early in 2025 three clients asked whether our new platform could transcribe calls in real time and suggest answers to agents. The feature existed, but by default it processed audio in a region outside Canada. Our sales team proposed switching it on across all programmes as a standard feature, and the chief operating officer wanted the transcripts to score every call automatically for quality.

I recommended a narrower path. Agent-assist would run only for a client that instructed it in writing, after a privacy assessment led by our privacy officer, with a caller notice in the recorded greeting, processing in a Canadian region, a defined transcript retention period and a contractual bar on using transcripts to train any model. Quebec's private-sector privacy law requires an assessment before personal information is communicated outside the province, and federal privacy law expects consent proportionate to sensitivity; counsel advised on both. I also asked that transcript and suggestion data not be used to score individual agents automatically. The chief operating officer accepted aggregate quality reporting only after a Sherbrooke pilot showed agents ignoring suggestions when they believed they were being scored.

The narrower path cost us. Canadian processing arrived four months after the general release. Two prospective clients, worth about C$4.2 million a year combined, wanted transcripts used to train their own models outside Canada without changing their caller notices, and we declined both.

By September 2026 eleven programmes for eight clients use agent-assist. Average after-call work on those programmes fell from 74 to 49 seconds, and handle time by about 6 per cent. In one insurer programme a suggested answer quoted the wrong waiting period for a policy benefit; a quality reviewer caught it, and suggestions now draw only on client-approved knowledge articles.

Agent-assist is now sold as a client decision with its own assessment rather than a feature switch. I would involve agents and team leaders from the first design meeting, not after a pilot had already taught them to distrust the tool.