Guide · AI for admin
AI can take most of the writing, sorting and remembering out of a small firm's office work: client updates, meeting notes, inquiry replies, invoice reminders, receipts and the monthly close. It should not choose fees, make contract decisions or send anything a person has not read.
Start with prompts to learn what AI does well on your own material. Move a task to an automation once it happens every week, follows the same steps, and you are tired of copying and pasting. Keep every client-facing message behind a person who reads it before it goes out.
Write every task down, however small, with roughly how long it took. The Admin Cost Report does this with you and estimates what admin costs your firm (the result is an estimate built from your answers, not a measurement).
Choose the one that is frequent, repetitive and annoying, not the one that is most important.
Use the prompt builder or the Prompt Pack. You will learn where the AI gets it wrong on your own material.
Put the information the task needs in one place, with one row or one file per item, named the same way every time.
Have the system write drafts or lists for a person to check. Keep it running for a month before you add the next task.
Free tool
Free: estimate what admin costs your firm before you decide what to automate.
For a firm of 2 to 25 people, AI is useful for admin in three ways, from least to most set-up:
This page covers admin only: the office work around projects. It does not cover AI for design, drawings, renderings or code compliance.
Admin is everything a firm does to win, run and get paid for projects that is not the design itself. In a small firm it usually falls on the principal, an office manager, or whoever is least busy that week. The tasks below are the ones this guide and its sibling guides cover.
| Stage | Admin tasks |
|---|---|
| New work | Answering inquiries, intake questions, setting up the project folder, first-draft fee proposals, proposal follow-up |
| Running projects | Meeting notes and action items, weekly client updates, RFI and submittal logs, decision logs, chasing late information, scope questions |
| Getting paid | Invoicing, late-invoice reminders, progress billing, accounts receivable aging |
| Closing the books | Receipts to projects, the month-end close, the pack for your accountant, cost against budget |
| The firm itself | Written procedures, onboarding a new team member, supplier quotes, an inbox nobody wants to open |
A prompt is the instruction you give an AI chat tool. The prompts guide explains the five parts of a good admin prompt and has a builder that writes one for you; the Admin Prompt Pack has 15 free prompts in English and Spanish.
They only know what you paste. You still have to remember to do the task: a prompt does not run itself on the first of the month.
Both main office suites now include AI in email: Google's Gmail help describes Gemini summarizing a thread and drafting with "Help me write", and Microsoft's support pages describe Copilot in Outlook doing the same. Features depend on your plan and edition. These help when the information you need is already in that mailbox.
Several practice-management tools sold to architecture firms now include AI features; Deltek lists an assistant called Ask Dela for Ajera, its product for small architecture and engineering firms. The practice-management guide compares Monograph, BQE CORE, Deltek Ajera and Deltek Vantagepoint from each vendor's own pages.
Built-in assistants see one system. Most admin crosses systems: an inquiry arrives by email, becomes a project in the practice-management tool, a folder on the shared drive, a row in the invoice log and a thread in WhatsApp. The gaps between those systems are where things get forgotten, and no single vendor's assistant covers them.
An automation is a small system that does a task on a trigger (a new email, a form submitted, the first day of the month) or on a schedule. It reads from one place, does something, and writes to another. AI is one step inside it, usually the writing step.
For example, each morning an automation reads the invoice log, finds invoices that are due or late and not marked Paid or Disputed, and writes the matching reminder as a draft in the mailbox of the person who owns that client. A person reads each draft, changes it if needed, and sends it. The full schedule and templates are in the late-invoice guide; the invoice chaser runs the same logic on your own invoice list.
Draft, do not send. An automation that writes drafts and lists can be wrong without harm, because a person sees the output first. An automation that sends, pays or deletes needs a level of testing most small firms do not have time for. Every system on this site is built to stop at the draft.
This table is our recommendation, based on how often each task happens and how much it varies. Your firm may differ.
| Task | Start with | Move to an automation when | Guide or tool |
|---|---|---|---|
| Weekly client update | A prompt on your notes | You send updates on several projects every week | Prompt Pack |
| Meeting minutes and tasks | A prompt on the notes or transcript | Meetings are recorded and the tasks need to land in a tracker | Prompt Pack |
| Client follow-up emails | Templates | You track open requests in a log | Follow-up guide |
| Late-invoice reminders | Templates and a schedule | You have more than a handful of open invoices a month | Late-invoice guide, invoice chaser |
| Receipts and month-end close | A checklist and a prompt on the bank export | Receipts arrive from several people every week | Month-end guide |
| Scope questions | A prompt that drafts the reply | Rarely: each one needs judgment | Prompt Pack |
| Fee proposal | A prompt for structure and wording only | Rarely: the fee is always your decision | Prompt Pack |
| Written procedures | A prompt on a spoken explanation | Not needed: it is a one-off task | Prompt Pack |
Some admin work looks mechanical but carries judgment. Keep these with a person, whatever tools you use:
| Risk | What it looks like | How to manage it |
|---|---|---|
| Details not in the source | A draft mentions a meeting date, an amount or a name that was not in your notes | Tell the AI to mark unknowns as [CHECK] instead of guessing; read every number against the source |
| Client data in the wrong place | Client names, addresses or financial details pasted into a personal AI account that may use chats for training | Use a business plan or change the training setting; remove details the task does not need; see the privacy guide |
| A message sent too early | An automated email goes to a client who already paid, or who is in a dispute | Automations write drafts only; status columns such as Paid and Disputed stop the drafts |
| Automating a broken process | The automation runs, but on a log that is never up to date | Run the readiness check first and fix the inputs |
| One person's system | Only the person who built it knows it exists or how to fix it | Write a one-page note: what it does, where it reads, where it writes, who to ask |
What happens to text you paste into an AI chat depends on the vendor, the plan and the settings. Whichever plan you use, never paste passwords, bank or card details, or government ID numbers, and remove names and addresses when the task does not need them. The privacy guide quotes each vendor's current wording and includes a tool that replaces client details with placeholders, entirely in your browser.
Most automations fail for ordinary reasons: the information is in three places, nobody owns the task, or the steps change every time. The readiness guide checks one task against ten questions and gives you a result and a to-do list.
We do not publish time-saved figures, because they depend on your volume and your current process. Time the task for two weeks before you change anything and two weeks after, including the time spent checking drafts. If the checking takes as long as the task did, the inputs need tidying or the task was not a good fit.
| Term | What it means for admin |
|---|---|
| Large language model (LLM) | The AI behind ChatGPT, Claude and Gemini. It predicts text, which is why it writes well and can still state wrong facts confidently. |
| Prompt | The instruction and material you give the AI. |
| Practice-management software | Software that connects projects, time, budgets and billing for a design firm. |
| Integration | A built-in connection between two tools, such as invoices passing from a practice-management tool to accounting software. |
| API (application programming interface) | The way one piece of software reads or writes data in another. Automations use APIs. |
| Webhook | A message one tool sends another the moment something happens, such as a form being submitted. |
| Automation | A system that runs a task on a trigger or schedule without someone starting it. |
| AI agent | An AI that can take several steps by itself, such as reading a mailbox and filing messages. Useful, and the reason the draft-not-send rule matters more. |
Keep this note with each automation, so it is not one person's system. Replace everything in [square brackets].
Download the templates as a plain text file, no signup needed.
When: As soon as an automation runs.
Subject: [Automation name]: what it does
What it does: [one sentence] Where it reads: [the log, mailbox or folder it reads from] Where it writes: [where the drafts or lists land] Who to ask: [owner] Rule: it writes drafts and lists only. A person reads and sends.
It is the use of AI chat tools, AI features inside the software a firm already uses, and simple automations to handle the office work around projects: intake, client updates, meeting notes, invoices, receipts and the monthly close. It does not cover design work.
Pick a task that happens every week, follows the same steps each time, and has its inputs in one digital place. Common candidates are late-invoice reminders, meeting notes to a task list, inquiry replies and sorting receipts by project. The readiness check on this site scores a task in two minutes.
It can, but we recommend it does not. Have the AI write drafts into your own mailbox and let a person read and send each one. Sending is where a wrong amount, a wrong name or a wrong tone does damage.
Usually not. Most of the tasks on this page run on a chat tool plus what the firm already has: email, a shared drive, a spreadsheet and the accounting or practice-management system. Some practice-management tools now include their own AI features.
Archon & Partners