Workflow
Buying, selling, leasing, content, follow-up, closing, books, and the people around a deal — eight tracks, laid out step by step, each step marked for how far AI actually goes. The third mark is what this page is about: the more a machine can do, the more precisely you have to say what it does not touch.
8 tracks · 55 steps · one mark each. Open a track for the detail.
The machine finishes it and I read the result. A step only earns this mark when a bad run is visible and reversible.
The machine writes it; I go through it line by line. What goes out carries my name, and the responsibility with it.
The preparation can run fully unattended; the motion that commits has to be a person. Not because the machine cannot reach it — because that motion should not be handed over.
What one transaction looks like from first call to keys. Commercial is missing on purpose — it goes up when I have real closings behind it, not a generic template.
What the client tells us comes back as a first-pass needs summary and a shortlist of areas — and it can be done by voice now, because voice already handles the customer-service half of this conversation. The real budget and the why-now get confirmed by a person afterwards.
Sweep everything on the market, drop the obvious no-gos, rank the rest. I read the list before it goes out — no model knows the elevator in that building breaks weekly.
Taxes, maintenance fees, sale history, neighbourhood numbers, what is within walking distance — one page per unit. Pure retrieval and arithmetic, so it runs unattended.
Time windows, driving order, and the back-and-forth with the booking system get planned; conflicts resolve themselves.
A person is in the room and the machine is on call: amenities answered on the spot, a quick read on the surrounding blocks, this unit compared against the last one while you are standing in it, feedback and a recap written up before you reach the car. Light, noise, a water stain, a smell — someone still has to stand there.
The machine pulls comparables and lines up the definitions. What we offer and how we pace it is my call.
Starts from an address or a listing number and produces a draft, clause by clause against the template. Drafting assistance — the agent reviews and decides.
Clauses broken down quickly, strategy suggested from options set in advance, e-signature sent out by the system. That only works because an experienced agent loaded the options first — the playbook is human-built and the machine picks from inside it. The call stays mine.
A detailed pricing strategy with the numbers behind it comes out of the machine. How it gets presented — which part leads, which comparison carries it, how the pushback gets absorbed — is designed in advance, by a person.
Same type, same range, same window — sold and active, in one table. Neighbourhood numbers refresh monthly on their own.
Declutter, repairs, staging, photography and measurement scheduling, laid out backwards from launch day. Which spending is worth it, I edit.
English and Chinese written separately, never translated across. Translated copy is the most recognizable kind of AI writing there is. I read it before it goes out.
Tab by tab, validated field by field, instead of hunting through a dozen screens — with the agent confirming each step before anything is saved.
Clicks, bookings, showing counts and recurring feedback words, summarized weekly — fuel for the next pricing conversation.
Several offers put side by side, clause by clause, risk spots flagged. Price is one line — conditions, deposit, closing date, how solid the financing looks. Ranking them is still judgement, not a sort function.
Same as the buy side: strategy drawn from a pre-loaded set of options, signing initiated by the system. Whether to accept, and when, is a person.
Leased and available comparables for the same layout, building and window, resolved into a range.
Both languages written separately; photo order and feature ranking drafted, then finalized by me.
Time windows and grouped showings get scheduled. Who is worth a separate appointment, I decide from the list.
Checklist generated, missing items chased, formats normalized. What we collect and how far we go is fixed by the compliance line, not by convenience.
Source verification, objective extraction, factual red flags: verify what can be verified, mark what does not reconcile. The system states facts; it does not reach conclusions.
Documents, comparisons and timeline arrive assembled. Who gets the unit is the landlord and the agent deciding, and carrying it — always a person, and always inside the tenant-selection boundaries of the Ontario Human Rights Code.
Documents, deposit receipts and the move-in checklist are all staged by the system. Signing, taking the deposit, handing over the keys — that trip is a person.
These belong to no single deal, and every deal leans on them. Five things running in the background year-round.
Topics come from questions clients actually asked and from the numbers this month — not from a trending list.
English and Chinese are written independently. The same fact deserves a different emphasis in each language.
Two rule gates: if a number cannot point at its source, or the wording touches a compliance line, the draft does not pass.
Same time every day, unattended, year-round. One round once failed to go out because a path had a space in it — and every monitor stayed green. So now the monitoring gets verified too.
Every neighbourhood number refreshes monthly and the pages grow themselves. Local pages carrying real figures with sources are also what machines are most willing to cite.
The same piece reshaped for each platform, read by a person before it goes anywhere.
Once content touches and follow-up records are wired together, attribution is a thing you can actually compute. Honestly: I built the content first and installed attribution later, and I am still catching up — so what I can give today is reach and real cases.
Deduplicate, restore last-contact dates, recover what they were looking for at the time.
Reply speed, how specific the questions get, open and revisit frequency. Someone asking about pre-approval, lawyers and closing dates is far closer than someone asking how the market is.
A job change, a family change, the real budget, the real reason for hesitating. Not in the data — only in the conversation.
Email, text, chat — all of it gets drafted. Every message has to carry something real: last month numbers for the neighbourhood they watched, a new listing that fits their criteria. Empty check-ins do more damage than any amount of AI phrasing.
Send one blast and some trust goes with it.
They replied, they opened it, they came back — the signal returns and the tier updates itself.
Routed to a person. No exceptions.
Condition periods, deposit deadlines, financing approval, lawyer cutoffs, closing day — scheduled backwards from the closing date.
When something expected has not arrived, the system says so early instead of in the final week.
Where the lawyer, the lender, the inspector and the insurer each stand, on one page. Who gets chased and how is my call.
The ledger, the receipts and the timing reminders are staged automatically. The deposit actually entering the brokerage trust account, confirmed one by one, is a person. Money changing hands does not go through an automated path.
Utility transfers, walkthrough checklist, key and fob counts, what was left behind — each one surfaced when it is due, so nobody has to remember it. The walkthrough itself is the next step.
Funds, keys and fobs, the final walkthrough list — all lined up in advance. This is the day the client has been waiting for. You show up.
Photograph it once; it files itself under the right category.
The driving that showings actually cost gets logged as it happens, so December is not a memory exercise.
Splits, deductions and payment timing checked per deal; anything that does not reconcile gets flagged first.
Filed quarterly so that what gets handed over is one clean package.
That goes to the accountant. My job is handing over clean records — I do not answer tax questions on their behalf.
Deal mix, cost structure, where the hours went. The machine pulls it; I draw the conclusion.
By transaction type and stage, the list of who needs contacting and what they need to receive.
What a lawyer needs is not what an inspector needs. Each brief gets drafted; I edit.
Availability, site visits and report turnaround land on the same timeline.
Profiles, quotes, availability and how they performed last time come assembled onto one page. Which one I recommend lends them my credibility — there is no automatic setting for that.
Who received what, who has not replied — status maintains itself.
The inspection finds something, financing stalls, the lawyers disagree. At those moments what a client needs to hear is a human voice.
Every step marked human only, gathered into five kinds. The rest of the page is not unattended — in the AI-drafts-and-I-review tier the person sits inside each step. And this tier is not machine-free either: the preparation can run fully unattended. It is the last motion, the one that commits, that has to be a person.
Preparation, unattended: receipts, the timestamp ledger, closing-day funds and key lists, lease documents — all staged. Execution, a person: the deposit entering the trust account, the keys changing hands, the signature going down.
Preparation, unattended: profiles, quotes, availability, how they performed last time, assembled onto one page. Execution, a person: who gets the tenancy, which lawyer, which inspector. Recommending someone lends them your reputation — a machine does not get to lend mine.
The real budget, the family change, why now. Never in the data — only in the conversation.
Complaints, bad news, an inspection that turns something up, financing that stalls. What a client needs then is a human voice. Legal and tax questions go to a lawyer or an accountant; I do not answer on their behalf.
Every message that actually reaches a client is read first, one by one. Blast once and trust drops a notch.
AI helps you see it. A person signs it.