Every call answered.
Hailey answers every call, books the job and chases your on-call tech, then you, then us. FlintCrew checks that the messages she takes really were saved.
After Hours $49.99/mo · Full Front Desk $199.99/mo · No contract · No setup fee
Here are our answers, and why each one matters to your shop.
| Ask them | FlintCrew | Why it matters to you |
|---|---|---|
| Can your AI make up a price or a promise? | No. Hailey speaks only set, written lines. | You’re on the hook for what your AI tells a customer. |
| What if my on-call tech doesn’t answer? | We keep going: your tech, then you, then one of our people. | A burst pipe at 2 a.m. can’t wait for voicemail. |
| How do I know it worked? | A plain report every Monday, good week or bad. | Proof without listening to recordings. |
| Do I need a sales call and a year’s contract? | No. $49.99 or $199.99 a month, priced right here. No contract, no setup fee. | Try it for a month, not a year. |
Research shows AI agents often say a task is done when it isn’t. FlintCrew checks the record, not the AI’s word: today, for messages taken on calls that go through FlintCrew. And if neither your tech nor you answers an emergency, a real FlintCrew person is paged.
of AI agent failures on customer-service tasks were “false success”: the agent said done, the system said otherwise. Advani, 2026
In a 2024 benchmark of AI agents doing booking-style tasks, the best model got fewer than half right. Yao et al., τ-bench, 2024
of contractors have high or complete trust in their AI tools. Accuracy and reliability is their top concern. ServiceTitan, 2026 State of AI in the Trades
of people strongly prefer a human over an AI agent. That’s why a real FlintCrew person is the last rung of our emergency ladder. SurveyMonkey, 2025
And you answer for what your AI tells a customer: in Moffatt v. Air Canada (2024), a tribunal held the airline responsible for its chatbot’s wrong answer. Summary
She speaks set, written lines, filled in with your business details and the answers you give her. Asked a price you haven’t given her, she says she can’t answer that one and that someone will get back to them. She doesn’t guess. She tells a caller they’re booked only once the booking is saved.
Hailey hides callers’ names, addresses and phone numbers from the AI that runs the conversation. Detection is best-effort, and the phone line’s speech-to-text still hears the call audio. How it works
Why it matters: more than 75% of consumers won’t buy from an organization they don’t trust with their data (Cisco, 2024), and 55% of confirmed small-business breaches involved a third party (Verizon 2026 DBIR).
Hailey knows HVAC, plumbing and electrical out of the box. Another trade? Tell us and we’ll set you up by hand.
Hailey answers the calls that come to her, day or night, instead of voicemail.
What broke, where, how urgent: a real emergency or a Tuesday appointment.
Your on-call tech gets a text with the address, the issue and the callback number. We ring them too. Hailey asks the customer whether they want a text: if they say yes and left a number, they are texted when the tech confirms, or told we are still trying if nobody has. The whole chain is written down.
Still no reply from your tech? We text you, then call our own team: a real FlintCrew person is paged the moment the ladder stalls. Nothing sits in a queue waiting to be noticed. Hailey does not transfer live calls.
A homeowner's AC goes down at 11pm in July. Here's what happens.
11:02pm — Customer calls. Voicemail. They hang up.
11:04pm — They call your competitor. Reach a live voice.
11:05pm — Job booked. Deposit taken.
7:00am — You wake up. Missed call. No message. No job.
11:02pm — Hailey answers. Calm, professional.
11:03pm — Name, address and phone swapped for placeholders before the AI that runs the call reads them.
11:04pm — Emergency confirmed. Tech texted. Customer told help is on the way.
7:00am — You wake up. Job's done. Invoice pending.
What missed calls cost you. What you're paying now to handle them badly.
You are choosing between us and a person, because a person is what we replace.
$3,750–$4,600/mo
Salary, payroll tax and benefits, all-in. Business hours only.
$179–$649/mo
Live operators answer your calls.
$49.99–$199.99/mo
Answers, qualifies, books (on the $199.99 plan), dispatches, escalates when nobody replies, and pages a real person if the ladder stalls. Calls written down turn by turn as they happen.
Plain English. It arrives whether the week was good or bad.
How many calls Hailey answered, what she did with them, and how many needed a human.
How many emergencies went out to your on-call tech, and how many nobody confirmed after every call-out was tried. The report gives counts, not names.
Redactions and blocked requests. We do not publish a number we have not measured.
FlintCrew takes its name from Flint, Michigan, where showing up and doing the work is just what you do.
We’re based in North Carolina and small on purpose. Every customer gets a direct line to the people running the system. No support tickets.
“Show up. Do the work. Help people.”
Tell us about your shop and we'll set up a live demo of the real product.
We'll reach out within one business day to schedule your live demo. In the meantime, press Call Me and Hailey will call you so you can hear her right now.
FlintCrew replaces names, addresses and phone numbers with placeholders before the AI that runs the conversation reads them — and gives you a plain-English weekly report of what happened on your calls. The report does not yet count what was redacted or blocked.
An AI receptionist that sounds right is not the same as one that did the job. These are the failures we built FlintCrew to catch.
You should not run an AI receptionist without this.
What FlintCrew checks today: messages, on calls that go through FlintCrew. The same check is switched on for cancellations and address changes, but has not yet been seen on a real call; we will say so here when it has. Bookings are not checked against your calendar while FlintCrew’s calendar connection is being renewed; we will say so here when it is back. It does not yet check emergency texts, customer confirmation texts or text-message consent; we will add each one here when it does. Detection of customer details is best-effort: a detail we do not recognise is not replaced. A sealed chain cannot by itself show entries cut off the end; keep a seal you saw earlier and the checker catches that too.
FlintCrew is our AI governance gateway. When Hailey books a job, cancels one, changes an address or takes a message, the request goes through FlintCrew. It comes with both plans.
Names, phone numbers and addresses are swapped for placeholders before the AI that runs the call reads them. The real details stay on our side, where Hailey uses them to read them back and book the job. When a request is screened for manipulation, the screening AI reads it as the caller gave it, names and addresses included, except on the route through FlintCrew in its default privacy setting.
Bookings and non-emergency messages are screened for manipulation before they run, and every action is checked against the list of actions Hailey is allowed to take. If it fails, it is refused.
After Hailey takes a message, FlintCrew looks in the real record itself to see it was saved. A check that fails comes back ✗, not ✓. The same check is switched on for cancellations and address changes, but has not yet been seen on a real call.
What was asked, what was protected and what was refused are recorded for each action, and every check’s result is sealed into a tamper-proof record, so there is something to show when someone asks what your AI did.
Today FlintCrew governs Hailey. It does not yet cover other AI tools in your business, such as ChatGPT or Copilot.
Built by engineers who've seen what happens when AI goes wrong.
Before the AI that runs the conversation reads anything, the transcript runs through Microsoft Presidio — an industry-standard NER (name recognition) engine. Names, addresses and phone numbers it finds are replaced with placeholders, so that AI works from "{{PERSON}} at {{ADDRESS}}", not the real details. If that step can't run, Hailey skips the AI for that turn rather than send it anything. Detection is best-effort: a detail the engine does not recognise is not replaced. The real details are kept on our side so Hailey can read them back and book the job, and they are stored securely. When a request is screened for manipulation (L3 and L4), the screening models read it as the caller gave it, names and addresses included. Only on the route through FlintCrew, in its default privacy setting, do they read placeholder words for the names and addresses it detected. A lookup by phone number alone is not screened for manipulation.
When Hailey is about to act on something a caller said — a booking, a non-emergency message — those words are compared against a database of known AI attack phrasings that we curate. If the input matches a known prompt injection technique or jailbreak attempt, the action is refused before it runs. Emergency dispatches are never held up by this check. The database covers the attacks on our list, not new ones.
When Hailey is about to act on something a caller said — booking an appointment, taking a non-emergency message — the caller's words are screened by an AI safety model for prompt injection, jailbreak, and data-exfiltration attempts before the action runs. A flagged request is refused and recorded instead of carried out. If the screen itself can't run, the request is rejected rather than waved through. An emergency dispatch, or the emergency page to you, is not held up by this screen.
A second AI model scores the same request on a 0–1 scale for manipulation intent — separate from the injection screen. Where that screen asks "is this an attack on the AI?" this layer asks "is this caller trying to manipulate the AI?" A score of 0.7 or higher blocks the action. If this model can't run, its score counts as zero and it blocks nothing; the L3 screen above still has to pass, and that one refuses when it can't run.
Every action Hailey takes — booking an appointment, dispatching a tech, taking a message — is checked against the fixed list of actions she is allowed to perform before it runs. An action outside that list is refused and recorded. This check does not compare the action with what the caller asked for.
Hailey's lines are fixed wording filled in with the caller's own details, not text an AI writes freely. The last step before a line is spoken checks it for a leftover placeholder: if one is there, the line is not spoken, so nobody hears a code word. Hailey says “Sorry — could you say that once more?” instead, or, if it was her last line of the call, “Thanks for calling — goodbye.” This guard checks for placeholders only. It does not review a line for other content.
Each call is recorded turn by turn as it happens: what the caller said, what Hailey said, and the actions she took, with names, addresses and phone numbers replaced by placeholders. Actions that go through FlintCrew get a second record there, including the ones it refused. The recorder sends each turn on its own, after Hailey has decided what to say, which means a turn can be missed if the recorder fails. Taking customer details out of each recorded turn runs on the same machine as the live calls, so it can add a short pause. It is a working record, not a guarantee of every line. Every Monday you get a proof report by email: what Hailey did that week, in plain numbers. It does not count redactions or blocked requests. No jargon. No tech degree required. If you ever need to show a customer, insurer, attorney, or regulator what happened on a call, this record is what we pull.
Your proof report tells you what Hailey did with the week's calls. It arrives by email every Monday. The figures below are an example, not a customer's.
A line with nothing to report is left out, and a week with no calls gets one sentence, not a grid of zeros. If we could not produce a figure, the report says so instead of showing a zero. It is a short email with no attachment, and it does not report on redaction or blocked requests.
So we are precise about who sees it. The AI that runs the conversation works from placeholders — the names, addresses and phone numbers Hailey detects are replaced before it reads a word, and if that step can’t run, Hailey does not call that AI on that turn, so it is sent nothing. We do see it, and so does your tech: that is how somebody gets to the door. So do the services that carry the work: the voice service that turns the call into text, the texting service that sends your tech the address, and your Google Calendar when a job is booked. When a booking or a non-emergency message is screened for manipulation, the screening models read it as the caller gave it, address included. The difference is that it is scoped to your business alone.
Your calls, your bookings, and the people waiting on you — from your phone, whenever you want it.
Your calls, newest first, with the time, the caller's number, what the call was about and how it ended. Open one to read the transcript, with names, addresses and phone numbers shown as placeholders.
Today's booked jobs on your home screen, and a count of the calls on your call log that left a booking.
Emergencies Hailey sent out, people waiting on a call back, and calls that got nowhere — at the top of your home screen, so you can respond fast when it matters.
A proof report by email every Monday: what Hailey did that week, in plain numbers.
We'll show you the real product on a live demo — your questions, your shop, no slides.
Two plans. No contracts. No setup fee. Protection built into both.
No contracts. Cancel anytime. Setup is free — most shops are live before their next busy weekend.
Hailey answers calls from day one. Text messages start once the phone carriers approve your number, which can take several weeks.