Methods for systems that serve other companies' callers

An outsourcer's efficiency comes from sharing buildings, networks, platforms and sometimes agents. A client's security reviewer starts from the opposite assumption. I design for separation first and then decide, client by client, what may be shared. Each programme at Tidelorne has its own virtual-desktop pool, identity groups, recording store, encryption keys and network segment that cannot reach another client's systems. Where two clients have agreed in writing that agents may move between their programmes, we document the pairing and review it every year. I test isolation the way a reviewer would, by asking an ordinary agent account from one programme to reach another programme's data and recording what happened. The design costs money and some scheduling freedom. I would rather a client sees that cost in the price than discovers later that separation meant a permission setting in a shared application.

Telephony failures are unusually public. A caller who hears silence does not care whether the fault sat in our platform, a carrier or the client's own number. When we change platforms, I plan around the programme rather than the technology. Each client has its own cutover date, test calls, rollback point and a named person on both sides who can stop the change. I insist on testing the routes from each carrier into the platform, not only the platform's internal failover, because the weakest link is often a number translation configured years earlier. Recording, reporting and workforce data must reconcile before and after every move. A phased migration costs more and runs longer. It also keeps one programme's problem confined to that programme, which matters to a client whose callers' patience it is paying us to protect.

Workforce management sits between three groups who read it differently. Clients see service levels, schedulers see forecast intervals and agents see whether a promised break arrives. I run the systems that hold forecasts, schedules, adherence and time records, together with their links to payroll and client reporting. My test is whether an agent's paid time, the scheduler's record and the client invoice describe the same day. When they disagree, the agent's pay is corrected first and the cause is traced afterwards. I also keep adherence data out of disciplinary use until a supervisor has examined the context, because a platform outage can look exactly like an agent who stopped working. Accurate schedules matter commercially as well: a forecast built on unreliable history produces understaffed mornings that no service credit fully repairs.

Real-time transcription and suggested answers can shorten calls and help newer agents. They also turn every conversation into lasting text, which changes the position for callers who expected to speak with a utility or an insurer rather than with software. I treat agent-assist as a client decision supported by our evidence, never as a default feature. Each programme needs written client instruction, a caller notice, Canadian processing, a defined retention period and a contractual bar on using transcripts to train any model. Quebec's private-sector privacy law and federal privacy law both shape that assessment, and our privacy officer and counsel lead the legal view. I also commit to agents that suggestion data will not be used to score an individual's calls automatically. Our first pilot showed that without that commitment, agents simply ignore the tool.

Clients rely on our SOC 2 Type II report, ISO/IEC 27001 certificate and PCI DSS assessment almost as much as on our agents. I treat those documents as summaries of daily practice rather than annual projects. Control owners keep evidence as they work, and our security team tests a sample every month instead of assembling records before assessors arrive. Availability is handled the same way. Our contracts set monthly availability levels with service credits, and I report the measured figure to clients and to our audit and risk committee, including minutes we would prefer not to count. In 2025 the figure was 99.96 per cent across programmes after a 52-minute routing failure in Winnipeg. I would rather explain a lower honest number than defend a definition that excludes the incident a client remembers.