Counsel now ask a question that did not exist three years ago: if the delay analysis was built with AI tools, will it be admitted? In Ontario the answer has two parts. One rule is already in force and applies to every expert report. Four more have been proposed and are not in the Rules. Both parts turn on the same thing, which is knowing exactly what the AI did and what it did not.
In force: O. Reg. 384/24, amending the Rules of Civil Procedure (R.R.O. 1990, Reg. 194), in force 1 December 2024. Read on e-Laws, with the consolidated Rules current to 28 September 2026. Proposed: the four proposals of the Artificial Intelligence Subcommittee of the Civil Rules Committee, as reproduced in the Law Commission of Ontario's submission to the Subcommittee dated 15 September 2025. I have not read the Subcommittee's own consultation paper, so the proposal wording here is taken from that submission. No case is cited in this note.
01 What is already in force
Since 1 December 2024 an expert report served under rule 53.03 has to contain two things it did not need before. They sit in subrule 53.03 (2.1), as paragraphs 6.1 and 6.2.
Paragraph 6.1 requires a statement, signed by the expert, certifying that the expert is satisfied as to the authenticity of every authority or other document or record referred to in the report. Three kinds of material are carved out: evidence in the action that the retaining party gave the expert to analyse, material cited only because another expert's report cited it, and anything whose authenticity the expert doubts. Paragraph 6.2 then requires the report to give details of those doubts, including any doubts about the material the retaining party supplied. Subrule 53.03 (2.1.1) presumes authentic, absent evidence to the contrary, anything published on a government website or by a government printer, in a scholarly journal, or by a commercial publisher of research on the subject of the report.
The same regulation added a matching certificate for lawyers, in subrules 4.06.1 (2.1) and (2.2): a factum must certify the authenticity of every authority it cites. The regulation does not use the words "artificial intelligence" anywhere. It does not need to. An invented authority is the best known failure of a language model, and the certificate makes the person who signs answerable for it.
Every recommended practice, protocol section, contract clause and decision the report refers to has to have been checked at its source by the expert, whatever tool drafted the sentence around it. And the carve-out for the client's own records is narrower than it looks. The schedule files are evidence supplied by the retaining party, so the expert does not certify them. But if the expert doubts one, for example an update whose logic or actual dates were changed after the fact, paragraph 6.2 puts those doubts in the report.
02 What has been proposed, and where it stands
In the summer of 2025 the Artificial Intelligence Subcommittee of the Civil Rules Committee, chaired by Justice Lauwers of the Court of Appeal, asked for comment on four proposed additions to the Rules.
- A definition. Artificial intelligence would mean a collection of technologies that replicate human intelligence in processing information, by performing tasks considered cognitive: perceiving, learning, reasoning, problem-solving, and understanding and generating language.
- Identification. A party who puts forward evidence generated in whole or in part by a computer system using AI would have to identify the software or program, identify the categories of data used to train it, and provide supporting evidence that its output is valid and reliable.
- Authenticity challenges. A party could challenge the authenticity of evidence generated or modified by an AI system. Where the court finds the evidence could reasonably be believed and could reasonably be fabricated, it would not be admissible unless the party tendering it shows, on a balance of probabilities, that its probative value exceeds its prejudicial effect.
- Expert AI evidence. Where the output of an AI system would be subject to rules 4.1.01 and 53.03 (2.1) if a human witness testified to it, the court would have to find six things. Three restate the usual admissibility requirements: the evidence is relevant and material, it is necessary to assist the trier of fact, and no other rule excludes it. Three are new: the evidence is based on sufficiently valid and reliable facts or data, it is the product of valid and reliable principles and methods, and it reflects a valid and reliable application of those principles and methods to the facts of the case. Admission would remain in the judge's discretion.
Where it stands: these are proposals. The consolidated Rules on e-Laws, current to 28 September 2026, do not contain the words "artificial intelligence". I have found no announcement that the proposals have been adopted, changed or withdrawn. The profession's responses were mixed. The Law Commission of Ontario supported all four and asked for a broader definition. The Ontario Bar Association filed its own submission, and the Toronto Lawyers Association questioned whether new rules are needed at all, as Law Times reported.
Prepare as though the fourth proposal applies now. Its three new tests ask about the data, the method and the application of the method to the case. Those are the same three questions a competent cross-examiner already asks a delay expert, with or without a rule. A report that can answer them loses nothing if the proposals are never adopted.
03 Which part of a delay analysis is "AI"
This is the question the proposals make counsel ask, and most delay reports cannot answer it today. A delay analysis prepared with AI tools has three layers, and they are not the same kind of thing.
The calculation
The dates, the float and the critical path come from a critical path method calculation: a forward pass and a backward pass through the logic, on the calendars. That is arithmetic. The same file and the same settings give the same answer every time, whether Primavera P6 runs it or another engine does. On the proposed definition I do not read a calculation of that kind as artificial intelligence. It does not learn, and it was not trained on anything. That is my reading, not a ruling, and a broader definition of the kind the Law Commission asked for would make the line harder to draw.
The AI-assisted steps
This is where a language model does real work. It may decide which analysis to run next, read correspondence and pair a letter with a slip in the schedule, summarise a set of updates, or draft the narrative. Output of that kind is what the second proposal is aimed at. If it goes into the report as evidence, the report should say so, name the software, and show why the output can be relied on.
The opinion
The opinion has to be the expert's. The fourth proposal is written for the case where an AI system's output stands in for what a human expert would otherwise say. A delay report should never be in that position. If the expert cannot explain every number in the report on cross-examination without the tool open, the opinion belongs to the tool, and the problem is larger than any rule.
Ask the expert to label the three layers in the report. Which figures were calculated, by what software and version. Which steps a language model performed, and who checked each one. Which conclusions are the expert's opinion. A report that keeps those apart can meet the proposals as drafted. One that cannot say which is which will struggle under the present law as well.
04 The three new tests, applied to a schedule analysis
| Proposed test | What it asks of a delay report |
|---|---|
| Valid and reliable facts or data | Which schedule files were used, identified well enough that the same files can be found again. Their data dates. Whether each update can be trusted as a record of what was known at the time, and what was done about the ones that cannot. |
| Valid and reliable principles and methods | A published method, named, with the reason it suits these records. Software that has been tested against something other than itself, with the results and the known limits available to the other side. |
| Valid and reliable application to the facts | The same inputs give the same outputs when the other side's expert runs them. Every setting that changes the answer is stated: calendars, constraints, scheduling options, and any change made to a file for the analysis. |
3 rows, complete. The three tests are items (d), (e) and (f) of the fourth proposal.
One item in the second proposal deserves a warning. It would require a party to identify the categories of data used to train the software. For a commercial language model, the expert knows only what the vendor publishes. That is a thin answer, and it is a good reason to keep the language model away from the numbers. A calculation engine has no training data, so the question does not arise for it. A model that drafted a paragraph the expert then checked against the record is a much smaller target than a model that produced a figure.
05 Questions to put to an expert who uses AI
- What calculated the dates and the float? Name the software and the version. If it is not P6, ask how it has been tested against P6.
- Did a language model produce any figure in the report? The safe answer is no. Figures should come from a calculation that can be re-run.
- Can the other side reproduce it? Ask what they would need: the files, the software, the settings. Ask whether the result is the same every time.
- What did the AI read? Project records are confidential, and documents obtained in discovery carry the deemed undertaking. Ask where the documents went and on what terms.
- Who checked the citations? Paragraph 6.1 is signed by the expert. Ask how each authority was verified, and against what.
- Which text was drafted by AI, and who reviewed it? Drafting help is not the problem. Unreviewed drafting is.
- Does the expert doubt any of the schedule files? If so, paragraph 6.2 requires the details, and counsel should hear them before the report is served.
- Whose opinion is it? Ask the expert to walk through one window of the analysis without the tool. If that cannot be done, stop there.
06 How Critical Path Partners is set up for this
CPP uses AI in its work and says so on the site. The practice is arranged along the three layers in section 03, and each part below can be checked.
- The numbers come from a calculation engine, not from a language model. The engine is open source, so the other side's expert can read it and run it. Its tests, including the cases built and scheduled in Primavera P6 for comparison, are published with their limits on the validation page.
- The language model works under the analyst. It sequences the analysis steps and drafts. The analyst reviews each result and signs the work.
- Every deliverable carries a reproducibility manifest: the hash of each schedule file, the parser and engine versions, and the method settings. That is the answer to the third test in section 04.
- The method is published and named. A forensic schedule analysis follows the AACE and SCL methods, and where the matter warrants it a second method is run as a check and the gap between the two is reported.
- The tools leave cause and responsibility blank. A schedule does not say whose fault a slip was. Those findings are made by the analyst from the record, which keeps the opinion where it belongs.
- Doubtful schedule files are found early. The claim workbench compares every update with the one before it and lists logic changes, constraint changes and edits to past dates. Its trust score is a signal, not a verdict, but it is how a paragraph 6.2 doubt gets found before the other side finds it.
If you are deciding whether a delay claim is worth an expert at all, the free Claim Check is the first step. If the other side has served a delay report and you want it tested against the questions in section 05, or you need an analysis that can answer them, the page for counsel sets out what an engagement delivers, or talk to CPP.
07 What is uncertain, and what this does not reach
- The four proposals may be adopted, changed or dropped. Check the current Rules on e-Laws before relying on anything in section 02.
- The proposal wording is taken from the Law Commission of Ontario's submission, which reproduces it. I have not read the Subcommittee's consultation paper.
- The Civil Rules Review has separately recommended wider changes to the Rules, including to expert evidence. This note does not cover them.
- Whether a critical path calculation falls outside the proposed definition is my reading. No court has decided it.
- Interim adjudication under the Construction Act and arbitration are not governed by these rules. The sensible working assumption is the same as for impartiality: prepare the evidence as though the court test applied, because it may later.
- No decision applying paragraph 6.1 or 6.2 is cited, because none has been verified for this note.
- Nothing here is legal advice. The rule numbers are given so the text can be checked at source.
08 Practical points
- Treat paragraphs 6.1 and 6.2 of subrule 53.03 (2.1) as the rule on AI that already binds your expert.
- Have every authority in the report checked at its source by the person who signs.
- Ask the expert to label what was calculated, what a language model did, and what is opinion.
- Keep language models away from the figures. Figures come from a calculation the other side can re-run.
- Name the software and the version for both the calculation and the AI tools.
- Get the reproducibility details into the report: files, settings, calendars, data dates.
- Ask early whether the expert doubts any schedule file, and why.
- Ask where the project records went when the AI read them.
- Prepare as though the three new tests in the fourth proposal apply now.
- Check the Rules again before the report is served. This area is moving.