What a finished AI system looks like in practice.

A practice running several sites shares one reception team. Calls arrive in bursts around opening, lunch and the end of the working day, and during those bursts a large share of callers hear an engaged tone or a voicemail box they do not leave a message in. Anything arriving after hours waits until morning, by which time some callers have already booked elsewhere. Staff are also interrupted mid-task by calls that are simply asking for opening hours, directions or a prescription status.

A voice agent answers every inbound line on the first ring, identifies the caller and the reason for the call, and handles the routine categories end to end: booking, rescheduling, cancellation, directions, hours and preparation instructions. It books directly into the existing practice management calendar rather than a parallel diary, so there is one source of truth. Clinical questions, complaints and anything ambiguous are transferred to a person, or captured as a structured callback request with a written summary attached.

Support volume in a growing store scales with orders, and the queue is dominated by a handful of questions: where the parcel is, whether an item fits, how to change an address before dispatch, and how returns work. Agents answer these by opening three tabs — the helpdesk, the store admin and the carrier tracking page — and copying details between them. Replies are slow at exactly the moments volume peaks, and policy answers vary depending on which agent replies.

A conversational agent on site, email and WhatsApp, grounded in two sources: the approved policy and product documentation, and live order data pulled per conversation after the customer is verified. It resolves status, delivery, sizing and policy questions on its own, and it can perform a bounded set of actions such as resending a confirmation or updating a delivery address before dispatch. Refunds, goodwill credits and anything outside policy are prepared as a drafted action and held for human approval.

Enquiries arrive from portals, the website and paid campaigns at all hours, and the agency that replies first usually gets the conversation. In practice replies wait until an agent is between viewings, and evening enquiries wait until the next morning. Agents then spend their own evenings qualifying enquiries by hand, most of which are outside the budget, area or timeline for the property they asked about. Viewing times are agreed in long message threads that never reach the CRM.

Every enquiry receives a reply within a minute on the channel it came from, referencing the specific property. A qualification conversation establishes budget range, area, timeline, financing position and whether there is a property to sell first, then matches against live inventory and offers real viewing slots from the right agent’s calendar. Qualified enquiries are handed to an agent with a written brief; the rest enter a long-horizon nurture sequence that reactivates when their stated timeline arrives.

A finance function receives invoices, statements and remittance advices as PDFs, scans and email attachments in dozens of layouts. Someone reads each one and types the values into the accounting system, then reconciles them against bank lines at month end. The work is slow, it is concentrated into the days when everyone is busiest, and transcription errors are usually found weeks later during reconciliation, when the original context has been lost.

Documents are captured from a monitored mailbox and shared folder, classified by type, and read with OCR and extraction into a structured record. Validation runs before anything is written anywhere: totals must reconcile against line items, dates must be plausible, tax must compute, and the supplier reference must exist in the ledger. Records that pass every rule with high confidence post automatically. Everything else goes to a review queue where a person sees the extracted field alongside the highlighted region of the source document and corrects it in place.

Quotes come back from carriers and vendors as free-text email, PDF attachments and spreadsheets, each with its own structure, currency, incoterms and surcharge treatment. An operations coordinator reads them, retypes the comparable parts into a spreadsheet, chases the vendors who have not replied, and makes a decision under time pressure with an incomplete picture. Comparisons are inconsistent because different coordinators normalise the numbers differently.

Requests go out to the vendor list automatically from the shipment or order record, with a defined response window and scheduled follow-ups for anyone who has not replied. Replies are parsed regardless of format into a common structure: base rate, surcharges, currency, transit time, validity window and incoterms. Everything is converted to a comparable basis and presented as one ranked table with the assumptions shown. A person still makes the award decision; the system removes the retyping and the chasing.

Between a signed proposal and the first piece of real work sit a dozen small coordination tasks: engagement letter, compliance and identity checks, document collection, folder and workspace setup, access provisioning, billing configuration, internal team assignment and scheduling the kickoff. Each is quick on its own, none is owned end to end, and they are performed slightly differently every time. New clients experience the delay as a quiet gap immediately after they committed.

Signature is the trigger. From that event the pipeline runs the entire sequence: a structured intake form is issued, required documents are requested and chased until received, compliance checks are recorded, the client folder and project workspace are created from the correct template, the CRM and billing records are completed, the delivery team is assigned by workload and skill, and kickoff is scheduled against real calendars. A person is involved at two points only — reviewing compliance evidence and approving the delivery team assignment.