AI can handle first replies to enquiries, member FAQs, tour scheduling and meeting-room availability questions today, provided it answers from your own fact sheet and the booking system confirms every slot. It should not handle price exceptions, contracts, complaints, access and security, billing disputes or anything touching safety. Those stay with staff.
Two practical points decide how well this works. The assistant is only as good as the one-page fact sheet behind it, and the booking system, not the assistant, is the source of truth for what is free. Both are the operator's decisions, not the software's. Searches for "cowork" and AI also turn up Anthropic's former Cowork feature, which is now simply part of Claude and has nothing to do with running a workspace.
A 60-desk space, one week of work, sorted into three lanes
Take an illustrative space with 60 desks, three private offices, two meeting rooms, a community manager full time and an administrator three days a week. In a typical week it gets 14 enquiries, runs five tours, handles about 25 room requests and fields 20 member questions, plus a handful of faults, invoice queries and the odd complaint. I have sorted the 19 recurring jobs into three lanes: the AI handles it and staff read a log, the AI drafts and staff send, or staff do it alone. The minutes are what each job costs the team today.
First lane, the AI handles it end to end from the fact sheet and the booking tools, and staff read the log once a week.
| Job | Per week | Minutes now | The condition |
|---|---|---|---|
| Factual first reply (prices, inclusions, hours) | 14 | 84 | Quotes only what is on the fact sheet; anything else becomes a task for staff |
| Offer tour slots and book | 5 | 25 | Shows only the slots the scheduling tool offers |
| Tour confirmation and reminder | 5 | 15 | Fixed template with directions and what to bring |
| No-show or reschedule message | 2 | 8 | Offers the rebooking link and nothing else |
| Member FAQ (printing, mail, guests, kitchen) | 20 | 60 | Answers from the member handbook; no codes, no passwords |
| Meeting-room availability question | 15 | 45 | Only if it reads the live booking calendar; otherwise it says it will check |
| Standard member room booking | 18 | 72 | Made in the booking system's own self-service; the AI explains and links |
| Fault report logging | 6 | 30 | Captures place, fault and photo into a ticket; staff do the fixing |
Second lane, the AI drafts and a person reads, edits and sends.
| Job | Per week | Minutes now | The condition |
|---|---|---|---|
| Quote for a team or private office | 3 | 75 | Rates from the rate card; a person confirms availability |
| After-tour follow-up with a proposal | 4 | 48 | Built from the tour host's three-line notes; staff add anything promised |
| Room request from an outsider with extras | 7 | 70 | Extras, catering and set-up are priced by a person |
| Invoice question | 4 | 40 | Needs account data, so staff check the invoice first |
| Event announcement and newsletter | 1 | 45 | The community manager edits for tone and checks the dates |
Third lane, staff only, with the reason each one stays human.
| Job | Per week | Minutes now | Why it stays with a person |
|---|---|---|---|
| The tour itself | 5 | 150 | The sale is made in the room, by someone who can read it |
| Discount or custom terms | 1 | 20 | A concession is a commercial decision |
| Noise or conduct complaint | 1 | 30 | Needs judgement, discretion and usually two sides of the story |
| Cancellation, refund or deposit exception | 1 | 25 | Policy exceptions and money |
| Door codes, fobs, lost keys | 2 | 10 | Security: the AI neither shares nor resets access |
| Member introductions | 1 | 10 | Needs the other member's consent |
Add it up and the week holds 862 minutes, about 14 hours. The first lane accounts for 339 of them, the second for 278 and the third for 245. Suppose staff spend an hour a week reading the first lane's logs, so that lane falls to 60 minutes. Reviewing a draft takes far less time than writing one, so say the second lane drops to 40% of its time, 111 minutes. The third lane does not move. The week falls from 862 to about 416 minutes, which hands back around seven and a half hours, and nearly everything left is work that needs a person. That is the shape to aim for: the staff-only lane stays the same size and becomes most of the job. The hospitality delegation rules tutorial sorts guest messages the same way for hotels and restaurants, if you want a second view.
Enquiry replies: the fact sheet that keeps them safe
Everything in the first lane depends on one page. It lists what you sell, at what price, with what rules, and it ends with a line saying what is not on the page and always goes to staff. The assistant is told to answer only from it. Here is an illustrative version.
FACT SHEET (answer only from this; if it is not here, say you will ask the team)
Spaces: hot desk $199/month; dedicated desk $349/month; private office for 4 from $1,450/month; day pass $25
Meeting rooms: Room A (6 seats) $30/hour members, $40/hour non-members; Room B (12 seats, screen) $48/hour members, $62/hour non-members; dedicated-desk members get 5 free hours a month
Staffed desk: Mon-Fri 08:30-17:30. Members have 24-hour access with their card
Included: Wi-Fi, tea and coffee, mail handling, 2 phone booths. Printing $0.10 a page
Guests: up to 2 per member per day, signed in at the desk
Pets: assistance animals only
Tours: weekdays 10:00, 12:30 and 16:00, 30 minutes, booked at [link]
Terms: monthly rolling, one month's notice; longer terms are agreed in writing by the community manager
ALWAYS PASS TO STAFF: discounts, contract changes, refunds, complaints, access codes, availability for a specific dateNow watch what a first draft does with it. The enquiry arrives at 9 pm and the draft is ready in seconds.
ENQUIRY: "Hi, we're a team of four looking for a private room from January. What's the price and can we visit this week? [first name], [company]"
FIRST DRAFT FROM THE ASSISTANT:
Hi [first name], thanks for getting in touch. A private office for four starts from $1,450 a month on our rolling monthly terms. We're happy to offer a 10% discount on a 12-month commitment, and the room is available from January. Tours run on weekdays at 10:00, 12:30 and 16:00: you can pick a slot here [link]. To prepare, could you tell us your start date, whether you need 24-hour access and how long you expect to stay?Two sentences have to go. The 10% discount is not on the sheet; the assistant invented a concession, which is the most common way these drafts go wrong, and it falls squarely under "always pass to staff". "Available from January" is an availability promise that only a person with the diary can make, so it becomes "I'll confirm which offices are free in January when we speak". Everything else is useful: the price, the tour link and a single message of qualifying questions instead of five. Replying in minutes at 9 pm, rather than at 9 am, is where I would look for the benefit first, and the five-minute replies tutorial covers the general method. The point here is that in a coworking space the discount line and the availability line are where the damage happens, so build them into the fact sheet as banned topics from day one.
Which enquiries a person should see first
Not every enquiry deserves the same speed from staff. Ask the assistant to tag each one when it drafts the reply: team size, desk type, start date and whether a tour is booked. Then apply a plain rule for what lands on the community manager's list first. Three enquiries from one week show it working.
| Enquiry | Tag | Why | Who acts |
|---|---|---|---|
| Team of four, private office, wants to start on 1 February, tour booked | Hot | Start within eight weeks and a tour is in the diary | The manager prepares the proposal before the tour |
| Freelancer, hot desk, "maybe in spring" | Nurture | No date and low commitment | The AI sends the handbook link and the day-pass offer, with a check-in in three weeks |
| Company asking about 20 desks from next year, wants a call | Large | More than 10 desks, so a custom conversation | The manager calls within one working day; the AI sends only a holding reply |
The 10-desk line is an example, so set yours where a custom conversation starts at your space. What matters is that the assistant never quotes for a large team or a bespoke layout. Its holding reply says the community manager will call, names a time window and asks for a phone number, and the tag puts the enquiry at the top of the manager's morning list.
Tours: from first message to follow-up
The best tour booking is a link, not a conversation. The first reply carries the scheduling link, the confirmation carries directions, how to get in, the first name of the person they will meet and what to bring, and a reminder goes out 24 hours and two hours before. A scheduling tool does the booking: Calendly's pricing page lists paid seats at $10 a month on its Standard plan and $16 on Teams, shown with its discount for yearly billing, so paying month to month costs a little more. Many coworking platforms have a tour feature built in too; either way, the AI offers the link and never holds the calendar itself. The scheduling basics are in the appointment scheduling tutorial: booking engine first, rules second, AI at the front door last.
A no-show gets a kind message the same day with the rebooking link. A tour that happened gets a follow-up within two hours, drafted from the three lines of notes the tour host types afterwards. That is the one lane-two job where speed matters more than polish.
HOST'S NOTES: team of 4, wants private office from 1 Feb, worried about meeting-room availability, asked about phone booths.
DRAFT FOLLOW-UP:
Hi [first name], thanks for visiting today. As discussed, a private office for four is from $1,450 a month on rolling monthly terms. On meeting rooms, Room A and Room B are bookable through the member app, and dedicated-desk members get five free hours a month; I will check what that means for a team on a private office and come back to you by tomorrow lunchtime. The two phone booths are free to use. Would you like me to hold a second visit for the rest of your team? [your name]The draft is honest because it does not answer the meeting-room question it cannot answer; it promises a person will. The community manager should check the "by tomorrow lunchtime" promise before sending, since the AI has just made a commitment on their behalf. Added up over a month, the numbers to watch are simple: enquiries, tours booked, tours attended, proposals sent and sign-ups. If 14 enquiries a week become five tours and two sign-ups, then lifting attendance by one tour a week through reminders can be worth more than any amount of copy polishing.
Meeting-room requests: the booking system decides
Members book rooms themselves in the booking tool, so the AI's job there is to explain how and to answer questions about capacity, equipment and rates. The harder case is the emailed request from an outsider, vague about dates and loose about extras. Here the assistant is a form-filler: it turns the email into a structured request for staff, marks what is unconfirmed, and drafts a holding reply.
EMAIL: "Hi, could we use your large room next Thursday afternoon, about 2-5 pm, for a team planning session with 10 people? We'd need a screen and maybe sandwiches."
STRUCTURED REQUEST:
Room: Room B (12 seats, screen) - suggested from headcount 10
Date: Thursday [full date to confirm: "next Thursday" is ambiguous]
Time: 14:00-17:00 (3 hours)
Headcount: 10
Needs: screen (included); catering "maybe sandwiches" (unconfirmed; staff to price)
Requester: external, non-member rate
Status: request only. Availability NOT checked.
HOLDING REPLY:
Thanks for your request. Room B seats 12 and includes a screen; the non-member rate is $62 an hour, so $186 for three hours, before any catering. I'm checking availability for Thursday [full date] and will confirm today.The sum is 3 hours at $62, which is $186, and the reply says "before catering" because catering is priced by a person. Notice that the reply restates the full date. Relative dates such as "next Thursday" are read differently depending on the day the email arrives, so the rule is to put the weekday and calendar date in every reply and let the sender correct it.
Now the failure that costs a space its reputation. Picture an assistant that tells a visitor "Room B is free on Thursday" because someone pasted a calendar export into the chat two days earlier. A member booked it yesterday, and two groups turn up. The fix is a rule, not a better prompt: availability comes from the live calendar or from a person, never from a pasted copy. Where room bookings live in your coworking platform, a general assistant sees them only through an integration the platform provides; where they live in a shared Google or Outlook calendar, a connected assistant may be able to edit events as well as read them. Check what yours can do, keep it read-only wherever there is a choice, and let the booking tool make the booking. The village hall booking tutorial works through the same pattern for a hall.
Member FAQs: one handbook, written from real questions
Most members ask the same 25 things. Build the list from evidence rather than from guesses: take three months of sent emails and community-chat messages, remove names, and ask an assistant to cluster the questions and count them. The top of the list becomes the handbook, written in short plain answers. The method for turning a list into answers an AI can use is in the FAQ-building tutorial.
Two exclusions matter more than any inclusion. The Wi-Fi password and door codes do not go into the sheet or into any bot: the answer is "it's in your welcome email, or ask at the desk". And nothing about other members goes in, ever, including who is in the building. Treat the assistant like a new temp: hand it the handbook, not the keys.
On tools, start with staff-facing drafting, which costs nothing beyond an assistant plan, and add a member-facing chat tool only when the handbook has survived a month. Options range widely. As listed in October 2026, Botpress has a free plan with 100 conversations a month, Chatbase has no free plan and starts at $32 a month billed yearly, and Intercom's Fin charges $0.99 per outcome, a count that can include a completed workflow or a hand-off as well as a resolved question. Check what each vendor bills for and what happens when you pass a limit. Whatever you pick, tell members at the start that they are talking to an automated assistant and show the way to a person.
What a coworking assistant must never be given
A coworking space holds a lot of other people's business and personal data. Keep these out of every prompt and every fact sheet:
- Door codes, alarm codes and Wi-Fi passwords.
- Member lists, and anything that shows who is in the building today.
- Invoices with bank or card details; give the assistant the invoice lines only.
- Incident reports that name people, and CCTV stills.
- Signed contracts, which carry signatures and personal details.
Choose a plan that does not train on your content by default (business plans such as ChatGPT Business and Claude Team work that way), or on an individual plan find the model-training switch in the privacy settings and check it rather than assume. If a member asks what data you hold about them, that request goes to a person and probably to a data-protection adviser; the AI should only hand it over.
The staff-only lane and the hand-over line
The third lane is not what is left after the AI has taken its share. It is what the business sells. A member who is thinking of leaving, a complaint about a neighbour's phone calls, a team that wants to bring a client for a pitch: these need someone who can listen, decide and be accountable. The aim of the first two lanes is to give that person more time for them.
Make the boundary mechanical so it cannot drift. Give the assistant a list of triggers that end its involvement at once: refund, cancel, complaint, noise, harassment, injured, fire, alarm, locked in, lost card, contract, invoice dispute, data request, press, police. When a trigger appears it sends a hand-over message and creates the task in the same moment.
HAND-OVER MESSAGE (sent by the assistant, with a task created for staff)
I can't help with that, so I've asked [community manager's first name] to look at it. The desk is staffed Monday to Friday, 08:30 to 17:30, and you'll hear back by [time]. If anyone is in danger or there is a fire or medical emergency, call the emergency services first.Three details make that message work. It names a person, so it does not sound like a dead end. It gives a time based on staffed hours, not a hope. And it puts the emergency line last and in plain words, because the assistant must never try to triage anything involving safety. The message is also the only reassurance it offers; "don't worry, this happens a lot" is not in its vocabulary.
Friday night to Monday morning, worked through
The clearest test of the three lanes is a weekend. Six enquiries arrive between Friday at 6 pm and Monday at 8 am. The assistant replies to all six within minutes. Three are factual questions answered from the sheet with a tour link, and two of those book Monday slots before the team has had breakfast. A freelancer asks about the day pass and gets the $25 price and the handbook link. One visitor asks for a discount on a private office, which is a trigger, so the assistant hands it over with a task. And one member emails about noise on Friday evening, which is also a trigger, so the same thing happens with the emergency line in the message.
At 8.30 on Monday the community manager opens a list of two hand-overs and two tours to prepare for, instead of a mixed inbox of six messages with no order. Four of the six never needed a person at all. Without the assistant, all six would have waited until 8.30 and joined the week's other work, and the two visitors who booked a tour over the weekend might have booked with a rival that answered sooner.
What the weekly log of AI conversations should show
Five measures are enough, and each takes a minute to read from your inbox, booking tool and ticket list. The figures below are illustrative.
| Measure | Before | Week 4 |
|---|---|---|
| Median time to first reply, out of hours | 14 hours | 4 minutes |
| Tours attended per week | 5 | 6 |
| Staff minutes a week on the 19 jobs | 862 | about 420 |
| Wrong answers found in the weekly log | Not counted | 1 (a stale price) |
| Hand-overs sent to staff | Not counted | 11 |
The wrong-answer log is the one to keep honestly. Every week, read the first lane's conversations, note anything the assistant said that was not true or not on the sheet, and fix the sheet before the prompt. A stale price is the usual culprit: put a review date at the top of the fact sheet and a calendar reminder to match. The number of hand-overs matters too. Too few suggests the triggers are not firing, and too many suggests the sheet has gaps. In both cases the log tells you which.
Questions coworking operators ask about AI
Do I need a website chatbot to start?
No. Start with AI-drafted replies to enquiries and a one-page fact sheet, because that proves the answers are right before any customer sees them unsupervised. A chatbot is the same fact sheet with a front end. Add one for members or visitors once your wrong-answer log has stayed empty for a few weeks, and keep a visible route to a person on every screen.
Can AI check meeting-room availability on its own?
Only if it can read your live booking calendar, and even then the booking system should make the booking, not the assistant. If it can only see a pasted export, it should say it will check and confirm, because an export goes stale within hours. Read-only access to the calendar is the safest set-up where your software offers a choice.
What should the AI say when it cannot answer?
Something short and honest: it cannot help with that, a named person will, and roughly when. It should also create a task or message for staff at the same moment, so the promise is real. Avoid 'I'll pass this on' unless a ticket exists. For anything involving safety, access or a complaint, it should hand over immediately and not try to reassure.
How much does this cost a small space?
Drafting with an assistant you may already pay for costs nothing extra, and individual plans list at about $20 a month. Scheduling tools start around $10 a seat a month. Member-facing chat tools run from free tiers with monthly conversation caps to paid plans from about $32 a month, and some charge per conversation or outcome instead. The larger cost is staff time spent keeping the fact sheet current, which is worth planning for.
Further reads
- Can AI Answer Gym Enquiries and Book Trial Sessions? — Another membership business: what AI may answer and how it books a trial.
- How to Build a Company Knowledge Base AI Can Answer From — Write the pages an AI answers from, with a 25-question test.
- Best AI Chatbots for Small Business Websites — Chatbot options priced and matched to real small businesses.
- One AI Inbox for WhatsApp, Instagram and Facebook Enquiries — One inbox for the messaging channels enquiries arrive on.
- What Is the Cheapest Way to Handle Enquiries Out of Hours? — Five ways to cover evening enquiries, priced at three volumes.
- AI Chatbots for Nursery Enquiries and Visit Bookings — Visit booking and hand-over rules from another tour-led business.
- AI Call Summaries: Log Every Phone Enquiry in Your CRM — Log every phone enquiry in your CRM with a callback date.
- AI for Church Administration: Rotas, Notice Sheets and Room Bookings — A hall reply built from your own terms, worked through.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Calendly, Botpress, Chatbase and Intercom pricing pages (checked October 2026); ChatGPT Business and Claude plan pages; the tutorial's weekly figures are illustrative.