Canada put AI market power on the competition agenda. Start with switching costs

THE SIGNAL IN ONE SENTENCE

Canada's competition watchdog has put AI market power on its list of questions for the next generation of communications policy. In an October 20 speech in Ottawa, Interim Commissioner of Competition Jeanne Pratt connected lessons from Canadian telecommunications to emerging AI markets. Her argument was not that every large AI company has broken the law. It was that competition has to be maintained through deliberate choices, and that the ability to switch is part of the machinery. The speech asks four direct questions about AI. How will access to data affect competition? Will AI reinforce advantages held by large incumbents? Could algorithmic pricing make it harder for consumers to understand offers or switch? How should authorities preserve competitive markets while encouraging innovation? Those are policy questions, not findings. The Competition Bureau did not announce a new investigation, accuse a specific company, issue a market study or propose a new AI law. Pratt described the problem as an approaching policy conversation and said the objective is to keep markets open, contestable and competitive. That boundary matters. Competition speeches have a habit of being reported as if someone just kicked down a server-room door. Nobody did. But the questions are useful because AI products are becoming sticky in ways that ordinary subscription comparisons barely capture. The plain signal is that Canada should measure AI competition with a switching-cost receipt. The Bureau's telecommunications example provides the basic logic. A market can have several providers and still offer weak competition if customers cannot act on the apparent choice. Long contracts, locked devices, opaque fees and difficult transfers can make a rival technically available but practically unreachable. AI services add new layers of attachment. A business may begin with a chat tool, then connect employee identities, upload documents, build retrieval indexes, save prompts, create agents, define evaluations, attach customer systems and train staff around one interface. Its invoices may come through a cloud contract. Its security team may rely on one audit trail. Its developers may use provider-specific APIs and tool formats. The model is only one piece. Leaving can mean rebuilding the workplace around it. That is why a model list is not a competition scorecard. Ten logos in a dropdown can still lead to one cloud, one identity system, one data store, one billing relationship and one set of proprietary controls. Start by mapping the layers. At the application layer, ask whether a customer can export conversations, files, prompt libraries, agent definitions, evaluations, user roles and feedback. Record the available format, the time required, the completeness of the export and whether another service can import it without manual reconstruction. At the model layer, test whether a workflow can move to another model without losing required capability, safety controls or output quality. Some differences are real product distinctions, not artificial barriers. The receipt should show which dependencies are necessary and which exist because a provider chose a proprietary interface. At the infrastructure layer, record data-egress fees, reserved-capacity commitments, identity dependencies, network architecture, logging formats and the engineering work required to move. A free model export is not much comfort if the surrounding system needs six months of surgery. At the contract layer, list minimum terms, automatic renewal, usage commitments, termination fees, credit expiration and any discount that disappears when a customer moves part of its workload. Discounts can be efficient. They can also make the real price of trying a rival difficult to see. Then run the exit test. Pick a representative workload before signing a large contract. Export its data and configuration. Recreate it with a second provider. Compare setup time, engineering hours, security review, output quality, latency, unit cost, error rate and missing features. Repeat the test annually because portability can decay as a system grows. This is not a demand that every product become identical. Competition depends on firms building distinct things. A provider that develops a genuinely better tool should benefit from that work. The problem is not that customers prefer one service. The problem is that unrelated friction can trap them after the reasons for choosing it disappear. The Bureau's January report, Your Data, Your Control, makes a similar point about personal information. It defines data portability as the ability to obtain and transfer personal data easily and securely between providers. It also says portability works only when privacy, security and interoperability are designed into the system. Those conditions belong in an AI switching test. An export that drops permissions can expose confidential material. An export that omits provenance can make a document collection unreliable. A transfer that breaks audit records can create compliance risk. A portability rule that ignores security may lower one barrier by opening a trapdoor. So publish both success and damage. For every migration test, record which objects transferred, which did not, what permissions survived, what data was duplicated, what the old provider retained and when deletion was verified. Note whether users had to consent again and whether administrators could inspect the final state. Small firms need their own view of the receipt. Large enterprises can negotiate custom exports, migration help and contract exceptions. A small Canadian company may get a download button, a support article and a cheerful suggestion to check the documentation. If competition policy measures only the experience of customers with procurement teams, it will miss the market most vulnerable to friction. Create standard tests for small-business accounts. Time a complete export. Price the move. Count manual steps. Test whether commonly used accounting, commerce, customer-support and collaboration tools reconnect. Publish the result in plain language. Now add algorithmic pricing. Pratt asked whether algorithmic pricing systems could create barriers to consumer information or switching. That concern needs precision. Software that changes a price is not automatically anticompetitive. Dynamic pricing can respond to demand, capacity or risk. The competition problem appears when a customer cannot understand the offer, compare it with a rival or know whether the displayed price includes the cost of leaving. For AI services, the comparison should include more than the headline token rate. Show model charges, storage, retrieval, tool calls, networking, support, reserved capacity, minimum spend and migration costs in one view. Keep the quote stable long enough for a customer to compare it. Record when personalization, bundling or automated negotiation changes the offer. Regulators do not need every company's pricing algorithm. They do need evidence that people can see the material terms, compare realistic totals and move without a surprise bill. Data advantage is the harder layer. An incumbent may have more customer interactions, richer operational data, larger distribution and the ability to place an AI feature inside a product people already use. Some of that advantage reflects investment and scale. Some may come from contractual restrictions, default settings or data that customers cannot take elsewhere. A switching-cost receipt cannot settle every market-power question. It can expose where the advantage lives. Measure whether customers can authorize a rival to access their data, whether the interface is reliable, whether access is equivalent enough to support a competing service and whether the incumbent can degrade the connection without a visible record. Track approval time, uptime, latency, field coverage, rate limits and denials. Do the same for public procurement. Canadian agencies buying AI systems should require a tested exit plan before award. The contract should identify export formats, documentation, assistance, deletion, transition periods and the price of departure. A government should not discover portability by attempting it during a dispute. Finally, publish trends rather than a single score. Switching cost changes as a market matures. A provider may improve export tools while adding proprietary agent features. A new standard may reduce technical friction while a bundled discount increases contractual friction. Report the layers separately so a good API does not erase a punishing contract. Canada's Competition Bureau is right that competition is a process. In AI, that process needs instrumentation. Count the hours, fees, missing objects, broken integrations and performance changes between deciding to leave and operating somewhere else. Then repeat the test. Market power can be abstract. The exit bill is not.