Enquiries to bookings
We build AI agents for WhatsApp, web chat and email that reply, collect the details, qualify and book into your real calendar or CRM. Rules check the facts, and anything unusual goes to your team with the conversation intact.
The problem
Messages arrive on WhatsApp, the website and email while you're on a job or in a meeting. Each one needs reading, a few questions and a booking before any money moves, and often it's you who does it.
An answering service or a chatbot takes the message, and then it waits again: nobody books it, writes it to the CRM or follows up. In Invoca's 2026 home-services benchmark, 55% of businesses didn't ask callers to book the job (vendor data, averaged across Invoca's customers).
What you get
Timeline
Pick one enquiry type, what a good booking looks like and the case where a wrong answer would hurt.
Collect real examples and set the thresholds the agent must meet.
Link the agent to your calendar and CRM through a queue people can approve or correct.
Go live for one enquiry type, then track corrections before widening it.
Where our agents work
In depth
An answering service puts a person on your phone. An AI receptionist answers calls or messages with software. Both take the enquiry, but neither books the job, writes it to the CRM or follows it up unless something behind them does. An AI agent that books does that part: it replies on WhatsApp, web chat or email, asks for the missing details and puts the booking into your real calendar or CRM, with a person on anything unusual.
Swipe or scroll to compare the columns.
| Option | What it does | What happens after the enquiry |
|---|---|---|
| Answering service | A person takes the call and a message | Someone on your team reads the message, calls back and books |
| AI receptionist | Software answers and takes details, often around the clock | The same, unless it is connected to your calendar and CRM |
| Missed-call text-back | Sends an automatic message when nobody picks up | The customer's reply waits for whoever is free |
| AI agent that books | Replies, collects details, qualifies and books within limits you set | The booking and the conversation are in your systems; hard cases go to a person |
If your team books everything promptly once a message is taken, a receptionist may be all you need. Many firms keep the receptionist they have and add the booking step behind it. Our guide Answering service, AI receptionist or missed-call text-back compares published UK prices and billing units.
The agent handles the routine part of an enquiry. It answers questions from approved information, asks for the details a booking or order needs, checks them against your rules and writes the result to the CRM or calendar. It can hold many conversations at once. Its limits are written down before launch: which questions it may answer, which actions it may take, and which situations end its turn.
Staff take over when:
The handoff carries the whole conversation and the details collected so far, so nobody asks the customer to start again. The WhatsApp order agent we built for a wall-art business works this way, with a custom CRM behind it. WhatsApp AI agents for orders and bookings covers the limits, the handoff and what an agent costs to build and run.
The full guide, section by section
A reply on its own doesn't move any money. For each enquiry type we agree what "done" means: a booking in the calendar, a CRM record with the details the job needs, and a follow-up if the customer goes quiet. Rules handle the parts that must be exact, such as whether the slot is free, whether the postcode is in your area and whether the required fields are filled. The agent's log records each reply, booking and handoff.
That log is also how you check the agent is working. Count this week's enquiries that waited more than a working day for a reply, or never got a booking. Count again a few weeks after go-live. The second count takes minutes.
AI is not the source of truth for price, contract, eligibility, identity, order status or access. It reads the message, drafts the reply and proposes the next step. Rules check the facts. A person approves anything with money or a commitment attached, and the agent never invents missing information or acts outside the authority it was given.
Swipe or scroll to compare the columns.
| Activity | AI agent | Rules and people |
|---|---|---|
| Interpreting input | Proposes fields, category or summary with its basis | Rules check format; an operator corrects uncertain results |
| Knowledge | Retrieves from approved sources and flags gaps | The owner approves sources, permissions and procedure version |
| Action | Prepares a proposal or a draft message | Rules control the state change; a person approves commercial or sensitive effects |
| Exception | Signals ambiguity and stops | The operator decides, records the reason and may resume the case |
Skip the agent when a booking page or form already handles the enquiry well, when nobody can say what a good reply looks like, or when a wrong answer would be expensive and couldn't be caught in time. It is also the wrong buy when a tool you already pay for does the job: many booking, helpdesk and shop platforms now include an assistant that answers from their own data. And if there is no clear enquiry type, no real examples and no owner, the problem is the process, not the model.
Some enquiries arrive as emails and attachments rather than chats. The same approach applies. The agent reads approved document types or free-text requests and turns them into fields, a category or a summary for the next person. Rules check the format and required fields, a confidence threshold decides what goes straight to the queue, and an operator corrects uncertain results with the source text beside the proposal. Request and document triage is a blueprint of this pattern, not a client result.
Complete the matrix before connecting a model to a real workflow. If a cell cannot be completed, the task is still in clarification.
Swipe or scroll to compare the columns.
| Task | Permitted data | Accepted result | Threshold / evidence | Permitted action | Approval required |
|---|---|---|---|---|---|
| Example: field extraction | Documents of accepted types | Field + source excerpt | Required fields present | Propose to queue | Yes, for missing or conflicting data |
The matrix describes a design boundary. It is not a model certification or an accuracy promise; re-evaluate it when sources, data, model or business rules change.
WhatsApp order processing and custom CRM is an anonymised delivered engagement. An AI agent handles routine order conversations within approved information and actions, a custom CRM keeps order status live, and staff take over ambiguous, sensitive or out-of-permission requests with the conversation intact. No response, conversion or volume metric is published.
AngajatAI agentic AI platform is a published case study of an agent workspace with company knowledge, tasks, scheduled jobs, integrations and connected applications, with permissions and human approval kept explicit. It publishes no performance metrics.
What Claude Code can do on a VPS shows a source-bounded operating model for an AI agent with phone access, monitoring, support drafts, scheduled tasks and human approval. Its scenarios are illustrative, not an AutomateFlow production result.
In the UK, a Pilot that puts an agent for one enquiry type into production, connected to up to three systems, costs £15,000 excluding VAT and typically takes four to six weeks. It includes exception paths, an audit trail, documentation and a 30-day defect period. Running the agent adds model usage, hosting and channel fees such as Meta's WhatsApp charges, paid to those providers, plus Run at £2,000 a month if you want us to monitor and fix it. We estimate running costs on your volumes before you commit. See the engagement models.
Fill in the matrix for one enquiry type: what arrives, what a good booking looks like, and where the agent must stop. Compare it with the public AI boundaries and how we work, or take the AI readiness check.
For a real week of enquiries, book a call about your enquiries. Bring where enquiries arrive, where they stall, and the case where a wrong answer would hurt. Send a process-level description, not credentials, customer data, documents or internal prompts.
"AI agent" means software that uses a language model to pursue a defined task with tools and permissions. "Bounded AI" means an interpretation step with a defined purpose, sources, output and authority. "Human approval" means a visible decision, attributed to an owner, before an action with impact. Generative AI is the model family that produces text; this page uses it only inside those limits. Nothing here is a safety, compliance or accuracy certification.
FAQ
Yes, within limits you set. The agent replies, asks for missing details and books into your calendar or CRM. Anything outside its permissions, or any customer who asks for a person, goes to your team with the conversation intact. Our WhatsApp agent for a wall-art business does this for orders, with a custom CRM behind it.
An answering service or AI receptionist takes the call or message. That's often enough if your team books everything afterwards. You need an agent, or an integration behind the receptionist, when enquiries still wait to be booked, written to the CRM and followed up. Many firms keep the receptionist and add the booking step.
Building one agent for one enquiry type is a Pilot: £15,000 excl. VAT, typically four to six weeks. Running it adds model usage, hosting and channel fees, such as Meta's WhatsApp charges, paid to those providers, plus Run at £2,000 a month if you want us to monitor and fix it. We estimate running costs on your volumes first.
Only within limits you approve. Replying with approved information and asking questions are low risk. Changing an order, refunding or cancelling need explicit permission, thresholds and usually a person's approval. Each permission is written down and can be withdrawn.
The agent answers only from approved sources, shows where an answer came from, and is tested on your real examples before launch. When it lacks the information or two sources disagree, it stops and hands the conversation to a person rather than guessing.
Count this week's enquiries that waited more than a working day for a reply, or never got a booking. Count the same thing a few weeks after go-live. The agent's log shows every reply it sent and every handoff, so the count takes minutes, and you don't have to take our word for it.
