Skip to content
Work→About→Stack→Contact→Get in touch
← All work
AI● In productionInteractive · sample data

The AI drafts. The team decides.

Five AI agents run inside the systems from the last two case studies. One matches homeowners to firms. One reads client documents into plans. One writes the monthly reports. One researches the niche and drafts content in our voice. The fifth is a set of scheduled jobs that keeps everything on time. None of them ship anything without a human, and all of them are below, working.

5 agents in production
Models sized per job
Every output human-approved
Runs while the team sleeps
The problem

Half the team's day was work that needed judgment, but not much of it. Pick 3 firms out of 104. Retype a client's plan into the system. Turn ad numbers into a readable report. Find content ideas in a niche you've been staring at for years.

That's drafting work, and AI is a good drafter. So the rule everywhere is the same: the agent proposes, a person approves. And every agent has a boring fallback, because production can't depend on a model having a good day.

01 · The matchmaker

104 firms. Three right ones.

Every qualified homeowner goes to 3 to 5 firms, and someone has to pick them. First, the feature as the caller sees it: one button, three picks with reasons. Then the same run below, told step by step, including what happens when the AI doesn't answer.

concrete.networksg.netTry itSample data
Callers AppSearch...⌘KHA
AI suggest · distribute
Pick 3 firms from 104
Mei TanCondo Resale$80–100kKeys: Oct 2026
Never bet production on a modelIf the AI times out or suggests a firm that doesn't exist, deterministic scoring answers instead. The caller can't tell the difference in speed. Only the tag changes.
the matchmaker · how it works, in plain wordsTry it
Step 1A lead is ready to go outMei is qualified and needs 3 interior design firms. There are 104 to choose from.
Step 2The obvious mismatches are removed firstSimple rules, no AI yet: wrong budget range, wrong housing type, firm not taking this kind of lead. 104 becomes 30.
Step 3The AI picks the best 3, and says whyIt reads the 30 candidates plus this week's priorities: who fits Mei best, and who is owed leads. Each pick comes with a one-line reason.
Step 4Its answer gets fact-checkedEvery suggestion is checked against the real client list. This run, the AI's 4th idea was a firm that doesn't exist. Deleted before anyone saw it.
Step 5 · a personA person presses sendThe caller reads the picks, swaps any they disagree with, and confirms. Nothing leaves the building on the AI's word alone.
Never bet production on a modelThe AI only sees a pre-filtered shortlist, its answers are fact-checked, and a person makes the final call. Flip the outage switch above.
02 · The reader

Paragraphs in, a plan out.

Clients send plans as WhatsApp paragraphs, PDFs, and call transcripts. Someone used to retype all of it. Run the feature first: extract, review, commit. Then read what actually happened, step by step, underneath.

concrete.networksg.netTry itSample data
Production AppSearch...⌘KMI
Import plan · AI extraction
From a rambly message to a plan

Clients send plans as paragraphs, PDFs, or call transcripts. Claude reads them into shoots and deliverables, dates inferred, children linked to parents.

June plan for Briq — two shoots this month. The Tampines 4-room (client offered the 14th or 15th) and the showroom refresh, late June. From Tampines we want a walkthrough reel, a before/after carousel, and a short with the designer voiceover. Showroom just needs the one hero film. Oh and carry over the testimonial edit from May.
Review before commit, alwaysThe AI proposes rows; nothing exists until a person approves them. And it's capped at 200 deliverables and 50 shoots, so a weird document can't flood the database.
the reader · how it works, in plain wordsTry it
Step 1A client sends a messy planA WhatsApp paragraph: two shoots, five videos, dates like “the 14th or 15th” and “late June”, plus one carry-over from last month. Normal client language.
Step 2The AI turns it into a checklistIt reads the paragraph the way a producer would: 2 shoots with actual dates (“14th or 15th” becomes the 14th), 5 videos, and each video attached to the shoot it comes from.
Step 3 · a personA person reads it before it becomes realThe checklist is a proposal. Everything on it can be edited or deleted. The AI has no power to save anything to the system.
Step 4One click creates all 7 recordsShoots first, then the videos linked to them, straight into the team's planning board. Nothing was retyped by hand.
The AI proposes, it never savesThere's also a safety cap. One import can't create more than 50 shoots or 200 videos, so a weird document can't flood the system.
03 · The analyst

Reports nobody had to write.

Every client used to cost an afternoon at month-end: pull the ad numbers, find the story, write it up. Generate the report the way the team does, then read how it's made: code does the math, the AI writes the words, a person signs off.

concrete.networksg.netTry itSample data
Production AppSearch...⌘KMI
Briq Design Studios · June 2026
The monthly report writes itself
Spend
$8,420
Leads
214
CPL
$39.35
IG engagement
4.8%

Pulled live from Meta the moment you ask: campaigns, engagement, posts. The AI only ever writes about real numbers.

Guardrails in the promptThe model is told what it can't say: budget moves stay between 5 and 15 percent, and “double down” is banned. Advice the team can't deliver is worse than none.
the analyst · how it works, in plain wordsTry it
Step 1The month's ad numbers are pulled, liveCampaigns, spend, leads, and social engagement come straight from Meta the moment the report is asked for. No copy-pasting from dashboards.
Step 2Plain code does the mathCost per lead, what changed since last month, which creative is tiring out. This part is ordinary code. The AI never gets to invent a number.
Step 3The numbers get checked for holesBefore any writing happens, the data is checked for gaps: a platform that returned nothing, a week that didn't load. This month: all clear.
Step 4The AI writes the story of the monthIt narrates the numbers it was handed, under house rules: budget advice stays between 5 and 15 percent, and hype like “double down” is banned. Advice the team can't deliver is worse than none.
Step 5 · a personA person edits, then it ships as a branded PDFThe account manager reads it, adjusts the tone, and sends. Every report is archived per client, per month.
The math is code; only the words are AIFlip the switch above to see what happens when a platform breaks mid-month.
04 · The content machine · n8n

It reads the niche. Then it writes.

A weekly n8n workflow scrapes what's ranking on Google, what Reddit is arguing about, and the ads competitors keep paying to run. It ranks all of it, merges the best with our tone and offer guides, and GPT drafts the copy in our voice, not the internet's. The real canvas is first. What it's doing, in plain words, comes after.

the real canvas · 30+ nodes · untouchedn8n
The production canvas. 30+ nodes of scraping, ranking, and writing.
Full screenshot landing here shortly.
The production workflow, not a diagramEvery green tick is a node that ran on the last schedule. Scroll sideways on small screens.
the content machine · what that canvas does, in plain wordsTry it
Step 1Every week, it wakes up on its ownNo one starts it. A schedule kicks the whole thing off with our six seed topics, the things homeowners actually search before a renovation.
Step 2It reads the whole niche in one sittingThree views of the same market: what ranks on Google, what Reddit is genuinely arguing about, and which competitor ads keep running. Nobody keeps paying for an ad that doesn't work.
Step 3361 findings become the best 19Everything it collected gets scored and ranked. Only the ideas with proven attention survive. The rest is noise and gets dropped.
Step 4It re-reads our voice rules before writing a wordOur tone and offer documents live on Google Drive: plain words, honest prices, no “dream home”, no exclamation marks. Update the documents, and every future draft speaks the new way.
Step 511 drafts land as Google Docs, for a personThe AI writes what's trending in our voice, not the internet's. A human edits, picks, and posts. The machine never publishes anything itself.
Research on a timer, taste on a contractAn afternoon of competitive research plus first drafts, done every week before the team sits down.
  • Three research angles. Search results, community threads, and paid ads see different things. The workflow reads all three.
  • 361 items → 19. Everything scraped gets scored and ranked. Only what earned attention becomes input.
  • Voice as a contract. Tone and offer documents live on Drive. The model writes inside them: plain words, honest prices, no clichés.
05 · The clockwork

Working while everyone sleeps.

Not every agent thinks. Some just never forget: hourly sweeps, portals provisioned for new clients, progress posted to Slack at 9am and the day's tally at 11pm. And when something needs judgment, it becomes two buttons in Slack. Tap one.

agents on a scheduleTry itSample data
The clockwork · every day
Every 15 minPoll the review platform for new homeowner enquiries
Hourly :00Sweep complimentary leads so firms can't claim replacements on free ones
Hourly :30Provision a portal for any client added outside the app
09:00 SGTPriority-list progress posts to Slack: sent vs promised, per firm
17:00 FriWeekly per-client digest drafted for review
23:00 SGTThe Allocation List: every lead sent today, counted by homeowner

War story: the hosting platform's own scheduler kept missing runs. So a plain Zapier clock hits these endpoints instead. Boring, reliable, done.

The human gate · Slack
Slack · #lead-concierge
CHome Match APP 10:39 AM

Replacement Request: Form Atelier / Esther Y.
Reason: uncontactable. Proof: 2 calls + 4 messages over 6 days, screenshots attached.

Tap a button. This is the entire review workflow. No login, no form.

Scheduled work, human decisionsEverything scheduled posts its receipts to Slack, and anything consequential waits for a human tap before it acts.
Under the hood

Claude for ranking, reading, and writing inside the CRM. GPT inside the n8n content pipeline. Six rules keep all of it predictable:

Models sized to the job

A fast, cheap model ranks firms in a second. A stronger one reads documents and writes reports. Nobody pays flagship prices for a sorting task.

A human before anything ships

Suggestions get confirmed, extracted plans get reviewed, drafts get edited, requests get a Slack tap. The AI proposes. A person decides.

Fallbacks, not faith

If the model is down or answers nonsense, deterministic scoring takes over and the feature keeps working. Production can't depend on a good day.

Hallucination guards

AI answers are mapped back against real records. A suggested firm that doesn't exist is dropped before anyone sees it.

Honest about missing data

When a platform returns zeros, the report says so in an auto-injected DATA GAP note, instead of presenting a hole as a trend.

Taste is versioned

The brand voice lives in tone and offer documents on Drive. Change the docs, and every future draft speaks the new way.

Next case study
Marketing Automation
Contact

Open to full-time roles.

Email me what's broken and how you want it to work. I'll reply within 24 hours with how I'd fix it.

Email meSend a brief