A public service that buys AI before it trains its own people has bought the argument and skipped the work. The first hard file will still land on a desk where nobody can say who checked the draft, what rule allowed it, or who signs the answer when a resident pushes back.
Direct answer: Rebuild in 3 steps, in this order. First, train people on real files from their own service before any tool rollout. Second, write guardrails on 1 page that a front-line worker can read and apply at the desk. Third, keep a named human accountable for every answer a resident relies on. Skip the lesson that leads with headcount cuts. That opening poisons adoption and drives out the experienced staff your rebuild needs most.
I put the order first because Ottawa debates tend to reverse it. A procurement moves. Training follows later, if the budget holds. Accountability gets settled last, after an answer has already gone out. Run the right order in a federal department, a provincial ministry or a municipal service. The scale changes. The order does not.
Why does training come before the tool rollout?
Training on real files teaches a worker where AI helps and where it fails inside work they already know. A generic course teaches terms. A real file teaches judgement.
Build the training in 4 moves. First, pick files your service closed in the last year, with personal details handled under your own records rules. Second, have staff draft the same answer 2 ways, once by hand and once with AI help. Third, compare the 2 drafts line by line with the person who would sign the real answer. Fourth, write down where the AI draft saved time, where it added risk, and where it sounded right and was wrong. That last list is your real curriculum.
Run the training by role, because the risks differ by desk. A benefits worker, a permits officer and a call centre agent use different records and sign different things. Give each group its own file set and its own task list. Public workforce AI training in Canada works when it starts from Canadian files and the resident at the other end.
What belongs on a one page guardrail?
A guardrail a front-line worker cannot read in 1 page at the start of a shift is a policy document, not a guardrail. Keep it to 1 page, in plain words, and sort every task into 3 groups.
The first group covers what AI may draft. A summary of a long file. A first draft of a routine letter. A list of missing items in an application. Draft means draft. Nothing in this group leaves the building on the strength of the draft alone.
The second group covers what a person must sign. Any answer that sets a benefit, refuses a service, sets a fee, grants or denies a permit, or tells a resident what they must do next. A named person reads the draft against the source file, checks the facts and signs. The signature means that person owns the answer in full.
The third group covers what never enters a tool. Set this list with your own access and privacy officers, in concrete terms your staff can apply without a meeting. If a worker has to guess whether a record belongs in this group, the guardrail has failed where it had to work.
The discipline in When an agent buys for you, set the authority before the first order applies here in a wider form. Authority set in advance protects the worker as much as the resident. A worker who can point to a written rule has cover.
How do you test a guardrail before you trust it?
Test the page, not the intent. Hand the guardrail to 5 workers who did not help write it. Give each worker 10 short cases from real work, with personal details handled under your own rules. Ask each to sort each case into the 3 groups and name the line that decided it. Where all 5 land in the same group and name the same line, the page works. Where answers split, the gap sits in the writing. Rewrite that line and retest.
The check in Google now checks each agent action against your written rules first points the same way from the tool side. If 5 trained people read your rule 5 different ways, no tool setting will save it. Clear writing comes first.
Who stays accountable when AI helps with the answer?
A named human stays accountable for every answer a resident relies on. Named means a real person, in a real role, whose name or title sits on the answer and in the file. A team name is not accountability. A vendor is never accountability, because a resident cannot appeal to a vendor.
Set accountability in 3 places. On the answer itself, so the resident can see who stands behind it. In the file, so a reviewer can find who checked the draft. In the appeal path, so the person who hears the appeal can reach the signer.
This is where the Canadian checks belong, run as questions. Put PIPEDA and your provincial privacy rules to each vendor as direct questions about your records, and confirm the answers with your own access officers and your own counsel. Ask how long records stay in the tool and who can read them. Ask how your records retention rules apply to drafts, prompts and outputs. Ask how you serve a resident in English and in French with the same care when AI helps draft the answer, and who checks the second version. Confirm each answer inside your own service, because the right answer depends on your jurisdiction and your records.
The buying test in Build or buy your next AI tool. The premium has to earn its place holds for a public buyer as well. Price the exit before you sign, run the export, and open the files.
What does the kept table look like?
Keep 1 table for the whole rebuild and review it each month. The table is the kept asset from this piece.
| Area | What you count or check | The rebuild passes when |
|---|---|---|
| Training | Staff who finished training on real files from their own service, by role | Each role has trained on its own file set before its tool rollout date |
| Guardrail | Cases sorted the same way by 5 workers using the 1 page guardrail | All 5 workers sort each test case into the same group and name the same line |
| Accountability | Answers a resident relies on that carry a named signer and a file record | Every sampled answer names the signer, the check made and the appeal path |
Which lesson should Canada skip?
Skip the lesson that leads with headcount cuts. A cut is a number, and a number looks like progress on a slide. As an opening move in a workforce rebuild, it fails on the work itself.
Announce cuts first and your experienced staff hear the message in full, whatever the deck says later. The people with the deepest file knowledge have the most options outside your service, and they leave first. The people who stay stop sharing what they know, because knowledge shared in training can read as knowledge handed to a replacement.
You lose the exact asset the rebuild runs on. Training needs people who know which smooth draft hides an error. Guardrails need people who have carried the file. Accountability needs people willing to sign an answer and defend it. Count service quality instead. Count errors found before an answer goes out, appeals upheld, and the time a resident waits for a correct answer.
How do you start in the next 30 days?
Week 1, pick 1 service line and name the person who owns the rebuild for that line. Week 2, build the training file set and run the first session. Week 3, draft the 1 page guardrail and test it with 5 workers on 10 cases. Week 4, set the accountability record on each answer type, run the vendor questions in writing, and fix the gaps the tests found. Start with 1 line, not 5. One line done well gives you a tested guardrail you can carry to the next line.
Your deadline is not the rollout date. Your deadline is the first file a resident appeals after AI helped draft the answer. On that day, a reviewer will open the file and ask who was trained, which guardrail line applied, and who signed. You cannot move that day by working harder in the week it lands. You can only meet it with work you did before the answer went out. Train first. Write the page. Name the signer.