The five phases
Every project runs in the same five phases. Each one ends with something your finance lead or IT contact can check before the next one starts.
| Phase | What happens | What you get |
|---|---|---|
| 1. Readiness assessment | We map the job with the people who do it today: the steps, the systems, the documents and the exceptions. We measure where the hours go and check what your ERP, accounting system and other software allow. | Findings, a work plan with milestones, and a budget by fiscal year and cost category |
| 2. Build and test on your past jobs | We build the agents and their connections, then run them on a set of your past jobs (quotes, orders, loads, records) and compare their output with what your team actually did, until they do the job your way. | Test results on your own work, and an agent card for each agent |
| 3. Rollout | The agents go live on real work, with a person approving every step that goes out. They take on more only when the people who approve their work agree. | A go-live checklist, the access list and the action log |
| 4. Staff training | Hands-on sessions for the people who approve, use and own the agents, plus change management for the wider team. | Training material, your AI use policy and a who-approves-what list |
| 5. First-year support | During the project we watch the self-checks and the logs, fix what breaks and change rules as your work changes. | A record of every fix and rule change, reviewed at each milestone |
What an AI agent is in our setup
An AI agent is not a chat window. In our setup it is a worker with five parts, written down before it starts:
| Part | What it means | Example: a Quote Desk agent |
|---|---|---|
| Role | The one job it does | Drafts quotes from RFQ emails, drawings and spec sheets |
| Tools | The systems it may use, and how | Reads the quotes inbox, past jobs and rates in the ERP; writes draft quotes only |
| Schedule | When it works | Checks the quotes inbox every few minutes during the day |
| Rules | What it must always and never do | Never sends a price an estimator has not approved; flags any tolerance it has not seen before |
| A person who approves | Who says yes before anything leaves | The estimator on duty |
We run our own company this way. Aitoflo has 35 AI employees on SAGE, our own AI employee system, with more than 70 scheduled jobs between them, around the clock. Our Secretary works the team inbox every 3 minutes, keeps the CRM current and drafts replies; its own note from 25 Sep 2026 reads "Draft reply, nothing sent." Our chief of staff, an AI called Trejon, hands work to the right desk and checks with a person before anything goes to a customer. People approve anything outward-facing or paid.
A free 30-minute call with Gurleen. Bring your finance lead or IT contact. You leave with a one-page project plan.
Book a call with GurleenHow it connects to the software you run
- ERP and accounting. Many ERP and accounting systems can share data with other software. Where yours can, the agent reads and writes through that connection, with its own access limited to its job. Where it cannot, the agent works from exports and prepares entries for a person to key in. We check this in the readiness assessment, before anything is promised.
- Email. The agent works a shared inbox or a mailbox set up for its job, such as quotes or orders. Replies go out from the approving person's mailbox or a shared one, after approval.
- Phone. A voice agent answers on your line or on a number you forward to it, looks things up while it talks, and logs every call.
- Documents. Drawings, spec sheets, purchase orders, bills of lading and certificates are read from where they already live: the inbox, a shared folder or your document system.
- Dispatch boards and schedules. The agent reads the board or the production schedule and proposes changes. A person approves them before they are made.
- Custom tools. Where your team needs a screen your software does not have, such as a quote review queue or an owner dashboard, we build it as custom software next to your systems.
Guardrails
- Approval before anything goes out or is paid. Quotes, order confirmations, customer messages, supplier emails and payments wait for a yes from the person who owns that job.
- Access limited to the job. Each agent gets only the systems and actions its role needs. A quoting agent can read rates; it cannot change them.
- A log of every action. What the agent read, what it wrote, what it sent for approval and who approved it, with the time.
- A regular self-check. Our Inspector checks every one of our AI employees every 30 minutes. Its note from 4 Sep 2026 reads "All clear, previous issues resolved." Your agents get a scheduled check of their own, and a person is told when something is off.
- Written rules it never breaks, agreed with you before go-live and changed only with your approval.
Data handling options
Where your data is stored and processed is decided with you before the build and written into your data rules. These are the options we discuss, and we confirm in the readiness assessment which ones fit your systems:
- Accounts in your company's name, where the software allows, so the data and the access stay with you.
- Cloud hosting in Canada, for companies that want their data kept in Canada.
- A computer in your own building, for work that should not leave your premises.
- Only the data the job needs. Each agent sees the fields and documents its job requires, nothing more.
- AI model settings chosen so your business data is not used to train the models.
- How long logs and copies are kept, set by you, and what happens to them when the project ends.
The documents you receive
| Document | What it covers | When |
|---|---|---|
| Work plan with milestones | Phases, dates, who does what, and what "done" means for each milestone | Before the project starts |
| Budget by fiscal year and cost category | Every cost, split the way funding agencies and your accountant expect | Before the project starts |
| AI use policy | Which jobs AI does, who approves what, and what it must never do | During the build, final at training |
| Data rules | What data each agent sees, where it is stored and processed, how long it is kept | Before anything connects to your systems |
| Agent cards and test results | Each agent's role, tools, schedule, rules and approver, and how it did on your past jobs | Before go-live |
How this maps to what program reviewers look at
If you apply for the federal 0% AI loan, the agency assesses, among other things, your capacity to execute, data security and governance, and your plan for responsible AI. A technical review may be done by NRC IRAP. Here is where our setup gives you material for each. The agency decides, and nobody can promise approval.
| What reviewers look at | What our setup gives you |
|---|---|
| Capacity to execute | A phased work plan with milestones and a budget by fiscal year, a builder who runs its own company on 35 AI employees, and test results on your own past jobs before go-live |
| Data security and governance | Access limited to each agent's job, a log of every action, written data rules, and a data handling choice made with you |
| Responsible AI | A person approves anything that goes out or gets paid, an AI use policy, written rules for each agent, a regular self-check, and staff training with change management |
The other parts of the assessment, such as economic benefits, financial capacity and market demand, come from your business. We help you set a starting point for the job, such as hours spent or quotes sent each week, so its results can be measured.