SITELIVE What launched
The City of Toronto has opened a pilot called Building Permit Application Pre-Check. An applicant uploads plans, forms and supporting documents to CivCheck — a third-party platform which the City’s notice of collection attributes to Clariti and Comply AI d/b/a CivCheck — and gets advisory feedback before formally lodging with the City. The tool compares the documents against a defined set of rules, requirements and bylaws within the pilot’s scope and flags missing plans or forms, incomplete or inconsistent information, potential Ontario Building Code and zoning issues, and files that fail the City’s electronic submission formatting rules.
The workflow is the interesting part. Results are labelled “likely compliant”, “needs attention” or “not compliant”. For anything needing attention, the applicant can point to where the information actually appears in the plans, upload a corrected file for another check, or explain why the requirement does not apply. At the end they receive a pre-check report, the reviewed documents and a unique pre-check ID, and are redirected into the City’s permit system — where the ID and PDF report must be supplied with the application so staff can confirm the pre-check was completed. In other words the AI output is not a decision; it is an artefact that travels with the file.
Scope is deliberately narrow. The pilot covers residential buildings containing two dwelling units or fewer, additions and alterations to them, and accessory structures such as garden suites, laneway suites, decks, porches and detached garages. Within that, it offers a limited advisory check of some Ontario Building Code requirements and of zoning and applicable law for certificate reviews on properties in Residential Detached and Residential Multiple zones. Use is voluntary, it does not change the City’s formal application requirements, and the City states flatly that the tool does not approve, refuse or delay any permit and does not guarantee a particular review time.
Why this shape of AI deployment is the notable bit
Most AI-in-approvals talk imagines the machine doing the assessing. Toronto has done the opposite and put the model on the applicant’s side of the counter, checking completeness and citing the rule behind every finding, while the statutory review stays entirely human. That distinction is not cosmetic: it changes who bears the risk of a wrong answer. A false flag costs the applicant an explanation; a missed issue costs them nothing they were not already exposed to, because staff review every application regardless. Compare that with a system that automates the decision, where a hallucinated clause becomes a legal defect.
The governance record is unusually explicit for an AI pilot, and worth reading if you are trying to get one approved anywhere. The City names the legal authority — a decision of its Planning and Housing Committee, the City of Toronto Act 2006 section 8, the Building Code Act 1992 section 7, and the municipal code chapter on building construction and demolition — and states that data is collected under Ontario’s municipal freedom of information and privacy act, must be stored securely in Canada on a pre-approved platform, kept separate from other data, and encrypted at rest and in transit. A security assessment and a privacy impact assessment were completed and remedial measures implemented. The approach is framed by the City’s Digital Infrastructure Strategic Framework.
The mechanism it targets is real. Incomplete applications are one of the most common causes of permit delay, and each correction cycle adds weeks that the applicant absorbs. Grounding a language model on a fixed rule set and asking it to check documents against that set — with a citation for every finding — is the narrow, verifiable version of this technology rather than the open-ended one. Whether it works is an empirical question about first-pass approval rates, and Toronto has not committed to publishing those.
The honest limits
Every performance number attached to this launch belongs to the vendor and to another city. Clariti says applicants in Honolulu using CivCheck have saved more than 40 days per permit on average and that residential permits reach decisions 55% faster than non-CivCheck applications in the first quarter of 2026. No methodology, sample size, or control for project complexity has been published, and self-selection is the obvious confound: the applicants who volunteer to run an optional pre-check are plausibly the more organised ones whose applications would have moved faster anyway. Toronto has published no target, no baseline and no evaluation plan for its own pilot.
The scope limits also cap the impact. Two dwelling units or fewer, in specified residential zones, is the small end of the housing pipeline — real for laneway suites, garden suites and small infill, irrelevant to the apartment projects that carry most of the provincial target of 285,000 new homes by 2031 the City says it is working to meet. The City itself says the tool reviews only requirements inside the pilot’s scope and may not identify every issue. And a completeness check does not touch the parts of approvals that actually consume months on larger projects: zoning amendments, site plan control, heritage, committee of adjustment, utility approvals.
There is a practical dependency too. Applicants must create an account on a third-party site, upload unlocked readable PDFs, and accept an AI disclaimer; password-protected files cannot be reviewed. Documents go to a vendor platform, which is why the data-residency and assessment language matters. None of that is disqualifying — but it means the pilot’s benefit is conditional on applicants doing extra work up front, which is precisely the group least likely to be causing the correction cycles in the first place.
What it means for operators
For builders and designers working anywhere approvals are the constraint, the transferable idea is not the AI. It is that a completeness check against an explicit rule set, run before submission, is worth money — and you can run one yourself today with a checklist, a named reviewer and a record that it was done. The reason Toronto’s tool can exist at all is that the requirements were already written down as a defined set. If your own submission process has no equivalent list, no amount of model quality will help.
The second lesson is about citations. The design decision that makes this deployment defensible is that every finding points at the requirement behind it, so a human can check the machine in seconds. That is the standard to hold any AI tool to on your own projects — for takeoffs, for code checks, for programme advice. An answer you cannot trace to a document or a record is an opinion, and opinions do not survive an audit or a dispute.