ARCFLOWS
Book a Call

Projects

Real builds, shown plainly.

One of these runs every week at a large church. One is in development. The others are prototypes and demonstration builds. Each one is labeled.

How to read these

  • In real use

    Runs on its own for real people, week after week.

  • Prototype

    Built and tested by us. Not yet used by a customer.

  • Demonstration build

    Finished and switched on. Shown with test data only.

  • In development

    Being built now. Not finished and not in use yet.

In the diagrams

  • Where it starts
  • An AI step
  • A person steps in
  • What it produces

Please note

These are our own builds and community work, not client case studies. Client projects will be added here, with permission, as they go live. Approximate figures, as estimated by ARCFLOWS.

In real use 01 / 05 since July 2026 · at a large church

Automated scheduling and events planning

An automated system for scheduling and events planning, in use at a large church. Six connected workflows run the volunteer team from its group chat: they collect who is away, draft a balanced monthly schedule, wait for the team lead’s approval, post it, remind people the evening before, and find volunteers for special events.

The workflow
Monthlyrequest PERSONVolunteersreply AIAI readsthe dates Schedule draftedby rules Shared calendar PERSONTeam leadapproves Posted andpinned Eveningreminders Event onthe calendar Call forvolunteers PERSONVolunteerstap to claim Nudge if spotsare open PERSONTeam leadtold MONTHLY SCHEDULE SPECIAL EVENTS Monthly request PERSONVolunteers reply AIAI reads the dates Shared calendar Schedule drafted by rules PERSONTeam lead approves Posted and pinned Evening reminders Event on the calendar Call for volunteers PERSONVolunteers tap to claim Nudge if spots are open PERSONTeam lead told MONTHLY SCHEDULE SPECIAL EVENTS
The bottleneck

Coordinating a large volunteer team by hand: asking who is away, working out a fair schedule for two roles, posting it, reminding people the night before, and finding cover for weddings and other special events, all inside a busy group chat.

What we built
  1. Once a month the assistant asks the group which dates people can’t serve.
  2. Volunteers answer in their own words. An AI model turns each message into exact dates and the assistant repeats them back, so the volunteer can correct it.
  3. It reads the month’s services from the church’s online calendar and drafts the schedule with fixed rules, not AI: two roles per service, nobody on twice in one day, the work shared evenly. Gaps are flagged.
  4. The draft goes privately to the team lead. Nothing reaches the group until the lead approves it.
  5. The approved schedule is posted and pinned. The evening before each service, the people on duty are tagged in a reminder.
  6. For special events it posts a call for volunteers with tap-to-claim buttons, nudges the group if spots are still open, and tells the team lead the day before.
  7. It watches the calendar a week ahead: a wedding gets its call for volunteers automatically, and anything unfamiliar is put to the team lead first.
Tools connected
  • Telegram (group chat, private messages, buttons)
  • The church’s online calendar
  • Claude AI model (Anthropic)
  • Six connected n8n workflows, with built-in tables for the team list, away dates, schedule and special events
Where a person steps in

The team lead approves every monthly schedule before the team sees it, and can ask for a fresh draft instead. Changes to the team list are shown as a before-and-after and applied only after a confirming tap. If a post fails or the calendar can’t be read, the team lead is told. AI is used in two places only, both for reading plain-language messages; the schedule itself is made by fixed rules.

Outcome

~95% less time scheduling. Compared with scheduling by hand: for over two years the schedule took about 4 to 6 hours a month to build, plus 2 to 3 hours a month re-scheduling absences, finding volunteers for events and sorting other issues. In use since July 2026. From the system’s own activity log, checked on October 1, 2026: three monthly schedules approved by the team lead and published, 29 evening reminders posted, and one call for volunteers for a special event filled.

Prototype 02 / 05 tested with two contractors

First-pass plumbing estimate from a plan set

Upload a plumbing plan set. AI agents read the drawings and work out quantities, labor hours and prices, fixed rules add up the totals, and a first-pass Excel estimate is emailed for an estimator to check.

The workflow
Plan setuploaded AIAI readsthe drawings AIEight AIagents draftthe estimate Totals byfixed rules Review flagsset Excel estimateemailed PERSONEstimatorchecks it Plan set uploaded AIAI reads the drawings AIEight AI agents draft the estimate Totals by fixed rules Review flags set Excel estimate emailed PERSONEstimator checks it
The bottleneck

Before a plumbing bid can go out, an estimator reads the plan set page by page, counts fixtures, measures pipe, works out labor hours and prices every line.

What we built
  1. A web form takes the project details and the plan set as a PDF.
  2. AI vision models pick out the plumbing sheets, read the schedules, count fixtures and estimate pipe runs from the drawings.
  3. Eight AI agents then work in sequence, each with one job: intake, drawing analysis, scope, quantity takeoff, labor hours, material pricing, a deliberately critical quality check, and a senior-estimator review.
  4. Fixed rules, not AI, do the arithmetic: line costs, markup, overhead, contingency and totals.
  5. One Excel workbook is built and emailed. Every line shows where its number came from and how confident the system is, and estimates with open issues are marked for review.
  6. The job is saved to a database, and if a run fails the owner gets an alert.
Tools connected
  • Web form with file upload
  • Claude AI models (Anthropic)
  • Excel
  • Gmail
  • Supabase database and file storage
  • n8n
Where a person steps in

The output is a first pass for an estimator to check, never a finished bid. The current version has no approval pause: every estimate arrives with the basis, source and confidence of each line, and anything doubtful is marked REVIEW. Earlier versions of this system, in their standard mode, stopped after counting and emailed the counts to a person, and nothing was priced until that person approved.

Outcome

~96% less time spent estimating. Compared with estimating by hand. Tested with two plumbing contractors, and in our own testing in September 2026, where runs that included the drawing-reading step took about 25 to 40 minutes per plan set. Unit prices and labor hours in the output are AI estimates to be verified, not supplier quotes.

Demonstration build 03 / 05 our own internal tools

Inquiry intake and meeting follow-up

Two small tools for the start and the end of a sales conversation. One logs each website inquiry and alerts the owner. The other turns a call transcript into structured notes and a follow-up email, saved as a draft that only a person can send.

The workflow
Websiteform Detailschecked Logged inone place Owneralerted PERSONOwnerreplies Transcriptpasted in AIAI writesthe debrief Saved andlogged Follow-updrafted PERSONPerson reviews,sends INQUIRY INTAKE MEETING DEBRIEF Website form Details checked Logged in one place Owner alerted PERSONOwner replies Transcript pasted in AIAI writes the debrief Saved and logged Follow-up drafted PERSONPerson reviews, sends INQUIRY INTAKE MEETING DEBRIEF
The bottleneck

Inquiries sit in an inbox until someone notices them. After every call, someone has to re-read the transcript, write up notes, pull out what each side promised, and compose the follow-up.

What we built
  1. Inquiry intake: a website form submission is checked, stored as one row in a log, and emailed to the owner with the reply address set to the person who wrote in.
  2. Meeting debrief: after a call, a person pastes the transcript into a form.
  3. An AI model turns it into a structured record: summary, key facts, pain points in the speaker’s own words, objections, the agreed next step, and action items for each side. It is instructed to quote exactly and to write “unclear” instead of guessing.
  4. The notes and the transcript are saved to Google Drive, the meeting is logged, and the notes are emailed to the owner.
  5. If a contact email was given, a follow-up email is saved as a Gmail draft.
Tools connected
  • Web forms
  • Gmail
  • Google Drive
  • Claude AI model (Anthropic)
  • n8n, with built-in tables for the inquiry log and the meeting log
Where a person steps in

The system never emails a contact itself. The owner replies to each inquiry by hand, and the follow-up after a call is only ever a draft: a person opens it, edits it and sends it.

Outcome

~42% more leads captured. Result from test runs, not live use. Both tools are switched on and have been run with test data only.

Prototype 04 / 05 test data only, not in live use

Guardrails for an AI phone receptionist

A prototype of the rule-checking layer that sits between an AI phone assistant and a clinic’s calendar and patient records. It limits the assistant to an approved list of actions, ties each call to one caller, and gives it honest ways to pass things to staff.

The workflow
Voice assistantasks Access keychecked Action onapproved list? Calleridentified? Appointmentsand holds Clinic questionsanswered PERSONPassedto staff Logged,reply sent ONE OF THREE Voice assistant asks Access key checked Action on approved list? Caller identified? Appointments and holds Clinic questions answered PERSONPassed to staff Logged, reply sent ONE OF THREE
The bottleneck

Front-desk staff answer the same routine calls all day: opening hours, insurance questions, booking, moving and canceling appointments. An AI phone assistant can only take those calls safely if something controls what it is allowed to see and do.

What we built
  1. Every action the voice assistant wants to take arrives at one entry point, which checks the clinic’s access key and checks the action against an approved list. Anything else is refused, with set wording for the assistant to say.
  2. The caller is looked up by phone number and the call is tied to that one record. Requests to see or change someone else’s appointment are refused.
  3. Open appointment times come from the clinic’s opening hours and calendar, and a chosen time is held for two minutes while the booking is made.
  4. Everyday questions (hours, address, insurance, services) are answered only from the clinic’s stored details, with no AI guessing. If nothing matches, it offers to take a message.
  5. When the assistant asks for a person, the request is recorded and staff are emailed or a call-back task is saved. It never claims a transfer it cannot make.
  6. Each request is written to a log with a fixed list of personal fields masked.
Tools connected
  • Google Calendar
  • Gmail
  • An entry point for an outside AI voice assistant
  • n8n, with built-in tables for clinic settings, appointments, holds, staff tasks and the activity log

There is no AI model inside this build: the voice assistant is a separate service, and this is the layer it has to go through.

Where a person steps in

Anything the assistant cannot do is passed to staff as an email or a call-back task, with a record of what was asked. This prototype has no approval step on bookings.

Outcome

No measured results yet. This is a prototype, tested with test data over about a week in August and September 2026, and it is not in live use. Availability checks, time holds, cancellations and refusals passed their tests. Booking and moving an appointment end to end are built but not yet proven.

In development 05 / 05 for a large church’s tech crew

Automated audio technician

A system in development for the tech crew at a large church. Cameras follow what is happening on stage, and an AI agent uses what they detect to work the digital mixer on its own.

The workflow
Cameras onthe stage Detectionand tracking AIAI agentdecides Digital mixermoves Cameras on the stage Detection and tracking AIAI agent decides Digital mixer moves
The bottleneck

Live sound needs someone at the mixer for the whole service, following who is on stage and adjusting as things change.

What we are building
  1. Cameras watch the stage.
  2. Detection and tracking pick out what the cameras see.
  3. An AI agent takes what is detected and decides what the mix needs.
  4. The agent moves the digital mixer on its own.
Tools connected
  • Cameras
  • Detection and tracking
  • An AI agent
  • A digital mixer
Where a person steps in

It is being built for, and with, the church’s tech crew. How the crew oversees the agent and takes over the mixer is part of the design still in progress.

Outcome

In development. It is not finished and not in use yet, so there are no results to report.

A 15-minute discovery call

What could your business do with less manual work?

Let’s look at the repetitive tasks slowing your business down and identify where automation could make the biggest difference.

Book a Discovery Call

Back to home

Book a Discovery Call