What an AI agent can do in an architecture office
An agent is a feature inside the studio's monday.com system, not a product beside it. It's a small program that watches one source (an inbox, a files column, a board, the calendar) and reads it with a language model. Then it does one bounded job, turning what a person reads, retypes and chases into a row, a draft or a summary where the office already works. Every agent below writes to boards our Foundation, Overhaul or Full System build sets up: the CA log, the Contracts board, the project boards and the overview board. That's why we sell agents after the build, never before: an agent with no CA log has nowhere to put an RFI.
For each agent below, we cover what it watches, what it does and where a person stays in the loop. None of them decides anything, and none can send mail or file anything outside the account.
The five agents
CA inbox agent
Watches the studio inbox, or a forwarding address on your domain that you ask the contractor and the city to copy.
Does: recognizes an RFI, a submittal or a change-order email and matches it to the project. It files a CA log row with the project, party, type, date out and owner, email attached. After seven days without an answer, it drafts the follow-up in the studio's own wording and puts it on the owner's task. That's the seven-day rule from the CA log page, with the draft already written.
The person: the project manager reads the draft, then sends it, edits it or picks up the phone instead. The agent sends nothing.
Consultant chaser
Watches replies from structural, MEP, civil, Title 24 and landscape consultants on threads it's copied on, matched to the contact and the project.
Does: reads the reply and sets the task's waiting-on value (Us, Consultant, City) and its new deliverable date, then posts a one-line summary on the project row. If the new date moves a plan-check submittal or a client milestone, it flags the task to the PM with both dates side by side.
The person: the PM confirms any date change that touches a submittal. The agent never emails the consultant, so a person still does the chasing, with the full record in hand.
Proposal drafter
Watches the Leads group on the overview for an inquiry past the first call, once the principal decides a proposal is worth writing.
Does: pulls the studio's fee template, with its own clauses, exclusions and fee phases. It also pulls the last three comparable projects (same project type, similar construction value) from the private Contracts board. Then it drafts the proposal on the lead's Contracts row, with scope by phase, assumptions and exclusions. It adds the fee as a range with its three comparables, plus the retainer and payment events. The architecture CRM page has the stages and the private board the draft sits on.
The person: the principal sets the fee and sends the proposal. The draft lives on a private board and never leaves it on its own.
Plan-check clearance reader
Watches a files column on the project board where the PM drops the city's Clearance Summary Worksheet or a correction letter as it arrives.
Does: reads the document and finds the agencies, the permit number and every clearance or correction line. Each line becomes a subitem, a small row nested under the right agency heading, with the permit number, the line text and a waiting-on value. Studio Yona, the Los Angeles studio in our case study, has the non-AI version today. It's a button that reads the structured worksheet PDF, and it got 6 of 6 right on a real City of Los Angeles worksheet. The agent version also reads the unstructured correction letter, where the layout changes from checker to checker and fixed rules give up.
The person: the PM checks the created items against the letter before any resubmittal. The agent never files anything with the city.
Weekly studio brief
Watches every project board, the CA log and, for the principal only, the Payments board, once a week.
Does: every Monday morning, writes one short paragraph per active project. Each covers the phase, what moved last week, who it's waiting on and for how long, plus which invoices are due or overdue. The principal gets it as an update on the overview, and each PM gets a version covering their own projects, without the money. It never emails a client.
The person: the principal reads it in five minutes instead of opening forty boards. If the boards fall behind, the brief says so, which is useful in itself.
How an agent is built
We pick one of two routes for each agent. For a job that starts and ends on a board, monday's own AI does the work inside the account. AI columns summarize a long update thread or classify an incoming item, and monday's AI agents run jobs across a board on a schedule. The weekly brief and the sorting half of the inbox agent run on monday's own AI. Nothing needs hosting, and the studio's monday admin can read every instruction the agent follows.
A job that reads an email or a PDF and applies the studio's own rules needs a small app on monday code, monday's own app hosting. The app calls Claude, Anthropic's model, with the studio's instructions and examples, and writes the result back through the monday API. The clearance reader, the proposal drafter and the consultant chaser work this way. Studio Yona's contract generator and clearance importer are the same kind of app without a model in the middle, so the hosting and ownership are proven. They run in the studio's own monday.com account, and the studio owns the code.
Every Hawkpoint build is made with Claude Code and checked by verification scripts, and an agent gets the same check before it touches a live board. A script feeds the agent twenty to fifty of the studio's real emails or letters. It compares the rows the agent would create with the rows a person created for the same documents. We read every mismatch, change the instructions and run the script again. The agent goes live only when the check passes on the studio's own documents, not a demo set.
Then comes a one-week pilot on one agent. For the first days it runs draft-only on the live inbox or board. It creates nothing, and we check its output against the real CA log or letter every day. Then it creates rows, and the PM reviews every one. At the end the studio decides whether it stays, and the second and third agents get the same pilot.
From the studio we need the forwarding address or inbox rule, the fee template and three past proposals. We also need a handful of real correction letters and about an hour of the PM's time in the second week of setup.
What it costs
Each AI agent is $900 to set up, then $450 a month. It comes after Foundation, Overhaul or Full System, because the agent writes to boards those builds create or rebuild. The CA inbox agent, the consultant chaser and the proposal drafter write to parts that come with Full System: the CA log, consultant coordination and the proposal pipeline. A studio without those parts starts with Full System, and the agents follow the month after. The setup covers the verification script, the one-week pilot and the agent's written instructions, and the studio owns the script and the instructions.
The monthly fee keeps the agent accurate as the studio changes: a new consultant, a new jurisdiction's letter layout or new follow-up wording. That fee is separate from our monthly care plan, which covers changes to the boards.
Two running costs are paid to others. monday meters its AI actions and bills the credits on your plan. The allowance and the price of extra credits are on monday.com/pricing, and we leave the figure there because monday changes it. A custom agent's model calls are billed at cost to your own account. For thirty CA emails a week, that's small next to one monday seat.
| Item | Price | When | Upkeep |
|---|---|---|---|
| AI agent setup, per agent | $900 | After Foundation or Full System; one-week pilot | $450 a month per agent |
For scale, a 10-person studio doing Full System, migration included, spends from $9,280 in year one with seats, per the pricing page. One agent adds $6,300 in its first year and $5,400 a year after that, before credits and model calls.
What stays with a person
Four things it is not
Not a replacement for the project list, the fee board or the CA log: it needs that structure first, so it comes after the build, never before. Not autonomous with clients or the city: nothing goes to a client, contractor, consultant or plan checker without a person's approval, and the agents are configured so they cannot send. Not free to run: monday bills its AI credits on your plan, and a custom agent's model calls are billed at cost. Not for a firm whose documents live only on paper or in unlabeled email threads: the free review tells you so before you buy anything. An agent reading a thread called "Re: Re: Fwd: stuff" learns nothing.
Where it needs a second look
Some inputs trip it up: a consultant who writes "end of next week", a contractor who puts three RFIs in one email, a correction letter with a page missing. On those, the agent's output is a draft to check, not a fact. The PM review in the pilot shows how often that happens on your documents. Buy the agent only if the studio can spare that review.
What has shipped so far
Only Studio Yona has agreed to be named, so the rest are described by industry.
- A monday AI agent that reads incoming email and turns it into CRM activities, for an interior design firm.
- AI columns on client boards that summarize long update threads and classify incoming items.
- A live drag-to-reschedule installation calendar for a cabinetry firm, built with monday's Vibe app builder.
- Custom monday code apps for Studio Yona that generate documents from the boards (contracts, proposals and invoice schedules), plus the clearance importer and the QuickBooks invoice sync.
- Like every Hawkpoint build, each of these was set up with Claude Code's help and checked by a verification script against the live account.
Key takeaways
- An agent is a feature of the studio's monday.com system, built after Foundation, Overhaul or Full System, never instead of them.
- Five chores: file CA email, track consultant replies, draft proposals, read plan-check letters, write the Monday brief.
- monday's own AI handles jobs on a board. A monday code app calling Claude handles jobs that read email or PDFs.
- Each agent first passes a verification script on the studio's own documents, then runs a one-week pilot, draft-only for the first days.
- A person approves anything that leaves the office, and the agents cannot send.
- Each agent is $900 to set up, then $450 a month to keep it accurate. monday credits and model calls are separate.
Limitations of this guide
These five agents come from our builds for California studios whose work goes through City of Los Angeles and LA County plan check. The clearance reader is written around that city's worksheet and letters, so another jurisdiction needs its own examples in the verification set. The one-week pilot, the twenty-to-fifty-document check and the thirty-emails-a-week example come from how we set agents up, not from surveys. monday's AI features and metering are as its site described them on October 7, 2026; we sell these builds, so weigh the costs section accordingly. If timesheets are the real problem, or you would rather set things up yourself, read the page on when not to hire a consultant first.