AI Agent Development Company in Victoria

We quote fixed fee against a signed statement of work, and agent projects price differently from ordinary software because the cost sits in evaluation and integration rather than in the model. A scoped pilot, covering the action inventory, an evaluation set built from real historical cases, one agent against one or two tools and a shadow-mode run with measured accuracy, typically runs CAD 45,000 to 110,000 over six to ten weeks. A production agent integrated into a case management, ticketing or asset system, with human-in-the-loop gates, audit logging, retention and a supported handover, generally runs CAD 150,000 to 450,000. A multi-agent system with several specialised agents, long-lived memory and integrations across more than one back office runs higher and should be phased. Three costs are consistently underestimated by buyers. Building the evaluation set, which is unglamorous data work and the single highest-value artifact in the project. Ongoing inference cost at real volume, which we model before commitment rather than after the first invoice. And the human review capacity the design assumes, which is a staffing decision, not a software one.

Start with the fact that an agent creates far more personal information than it consumes. The prompt contains whatever context you retrieved, the trace contains the reasoning, the tool logs contain what was read and written, and any hosted model endpoint sees all of it. Section 30.1 of FIPPA, which used to require that personal information held by a BC public body be stored and accessed only in Canada, has been repealed and no longer appears in the Act. That has not made residency go away, it has moved it into procurement documents, contract schedules and privacy impact assessment findings, and for most Victoria public bodies the answer is still Canada because their own terms say so. The practical rules we work to are these. Treat prompts, traces and tool logs as in scope for the same residency terms as the database, because they routinely contain more identifying detail than the source record. Get a specific hosting region committed in writing from any model vendor rather than relying on a product page. Where the terms cannot be met, self-host an open-weight model on Canadian infrastructure and accept the capability trade-off knowingly. And record the decision in the privacy impact assessment so it is defensible later.

It can prepare one. It should not make one that affects a person's rights, benefits, licence or status, and this is a legal position rather than a cautious preference. A decision made by a public body has to be authorized, has to be explainable to the person affected, and is reviewable, and none of that is satisfied by a model output that nobody can reconstruct. Since there is no Canadian AI statute in force, the constraints come from the law that already applies: FIPPA governs the personal information the agent touches, including collection authority and use limitation, and administrative fairness duties attach to the decision regardless of what produced it. The federal Treasury Board's Directive on Automated Decision-Making binds federal institutions and not provincial ones, but it is the most useful published reference point for how to think about risk tiers and we design to its logic where it fits. So the pattern we build is a named human decision-maker, an agent that assembles the file and drafts a recommendation with its sources attached, a record of what the agent proposed and what the human actually decided, and an audit trail that lets any single decision be reconstructed years later.

On the engineering side, yes, and we are direct about the boundary. Ocean Networks Canada operates cabled ocean observatories from the University of Victoria that produce continuous instrument data, and the work in that domain divides cleanly. Novel method development, such as a new detection approach for a rarely observed acoustic or seismic signal, is research and belongs with the researchers who own the domain and the publication. Turning a working method into infrastructure that runs unattended for years is engineering, and that is where most ocean data projects stall after a grant-funded prototype. That second half is our work: reliable ingestion, backfill and gap handling, versioned models with reproducible training, drift monitoring, alerting that a human on shift can act on, and cost control on continuous inference. We will tell you in the first conversation which half your problem is in. Where a project genuinely needs both, the sensible structure is a research collaboration for the method and a contracted engagement for the production system, and Mitacs programs are a common way for Victoria organizations to fund the research portion.

For unclassified scheduling, supply, document and maintenance workflows, yes. Retrieval is what makes this question different from an ordinary software question, and it is the part suppliers underestimate. An agent pointed at a shared drive will index whatever it can read, and once controlled technical data is inside a vector index, it leaks through paraphrase rather than through file access, which no document-level permission model will catch. So the boundary has to be built into the index itself: separate stores, retrieval permissions tied to a person's registration and screening status, every retrieval logged with what was returned, and controlled content never included in a call to a hosted model. The staffing rule follows the same line. Controlled scope is worked only by Canadian-based screened personnel inside the client's environment under the client's own security plan; our India-based engineers are not eligible to see controlled technical data and we say so at scoping rather than after a contract is signed. Uncontrolled scope, which is usually most of the work, runs through the full team. We do not take classified work.

Ten to eighteen weeks for a first production agent, and the variance is almost never the model. Weeks one to three are discovery: the action inventory, the data inventory, the privacy impact assessment scope for a public body, and the evaluation set built from real historical cases with known outcomes. Weeks four to nine are build, integration against the systems the agent has to read and write, and guardrail wiring. Weeks ten to fourteen are shadow mode, where the agent proposes and humans execute, and where the accuracy number that matters is measured against the evaluation set rather than against impressions. From week fifteen the rollout is gradual, by user group or case type, with the promotion criteria agreed in advance. Two things stretch this schedule in Victoria specifically, and both are predictable. Privacy impact assessment and security review inside a ministry have real institutional lead times that cannot be compressed near the end, so we scope them in week one. And the availability of the domain experts who have to adjudicate the evaluation set is usually the binding constraint, which we plan around rather than discover.