Experiment 04Process walkthrough
How AI agents work — a process walkthrough
What actually happens between an enquiry arriving and a person replying to it?
Explore the process
Explanation only — no model runs and no request is sent
Nothing here runs a model, calls an API or generates text. The stages are written content you step through. We build the systems this describes, but this page is an explanation of them, not one of them running.
How to use it
Step through the four stages in any order. Each one lists what is automated, what stays with a person, and how it fails safely.
Step through it
Process walkthrough · nothing here runs a model
Four steps, explained. This page does not call an API or generate any text — the examples below are illustrations of how the systems we build behave, written in advance.
Step 01 of 04
Capture
Everything lands in one place, tagged with where it came from.
Forms, phone calls, email and message threads arrive at a single intake surface rather than three inboxes and a notepad. Each one carries its source, the page it came from, and any campaign parameters, because "where did this come from" is unanswerable later if it is not recorded now.
Automated
- Deduplication against existing contacts
- Source and page attribution
- Spam and bot filtering
Stays human
- Nothing — this step runs unattended
Illustration
A form submission at 23:40 on a Sunday is recorded, deduplicated against an enquiry from the same address six weeks earlier, and linked to it.
When it goes wrong
If the intake surface is unreachable the form falls back to email delivery, so an outage delays an enquiry rather than losing it.
Step 02 of 04
Qualify
An agent asks what you would have asked, immediately.
A language model works through your qualifying questions — scope, timeline, budget range, location, fit — and writes the answers into the record as structured fields, not as a transcript someone has to read. The questions are yours; the agent is only the thing that asks them at 23:40.
Automated
- Structured extraction into fields
- Scoring against your fit criteria
- Flagging anything ambiguous for a person
Stays human
- Setting the criteria
- Reviewing anything the agent marks as uncertain
Illustration
The same enquiry is scored against service fit and location, and the reply asks two follow-up questions rather than the usual nine, because seven were already answered in the message.
When it goes wrong
Low confidence does not become a guess. The record is flagged unqualified and routed to a person with the full thread attached.
Step 03 of 04
Route
The right person gets it, with the context already attached.
Assignment runs on service area, language and availability, and the enquiry arrives in whatever system your team already works in. The routing rules are written down and inspectable, so a mis-route is a rule to fix rather than a mystery.
Automated
- Assignment by service, language and availability
- Acknowledgement to the enquirer
- Write-through to your CRM
Stays human
- Owning the enquiry once assigned
- Overriding an assignment when the rule was wrong
Illustration
A Maltese-language enquiry about automation routes to the person who handles that, not to whoever is first alphabetically, and the acknowledgement goes out within the minute.
When it goes wrong
If nobody matches the rules the enquiry escalates to a default owner rather than sitting in an unassigned queue where no one is responsible for it.
Step 04 of 04
Follow-up
Nothing goes quiet by accident.
Sequences are built around what the enquiry actually is, not a single generic drip. Timing respects working hours and time zones, personalisation is limited to facts already in the record, and every sequence has an exit — a reply, a booking, or an explicit stop.
Automated
- Sequenced sends with working-hour windows
- Reminders before booked calls
- Stall detection on quiet threads
Stays human
- Writing the sequences
- Deciding when to stop chasing
Illustration
A thread with no reply after four working days triggers one short nudge referencing the specific service asked about, then stops. It does not enter a nine-email sequence.
When it goes wrong
Any reply cancels the remaining sequence immediately. The common failure of automated follow-up is emailing someone who already answered.
What this is, and what it is not
This is an explanation. Advance through the four steps above and you get a description of how the systems we build behave, written in advance. There is no model behind it, no API call, and no generated text — the examples are illustrations, not output.
We built it this way on purpose. A convincing agent demo on the Lab page of an agency that sells agent systems is a claim a prospective client could act on, and you cannot caveat your way out of that. If you want to see one of these running, it will be running on your own enquiries, with your own questions, and we will show you the logs.
The part that matters most is not the model
The expensive failure in an enquiry pipeline is almost never a bad reply. It is no reply — an enquiry that arrived while everyone was busy and quietly aged out. Fixing that needs no AI at all: one place things land, and a written rule about who owns them.
We do the unglamorous half first for that reason. Capture and routing usually recover more revenue than qualification does, and they are the parts that keep working when a model is unavailable.
Where the model earns its place
Qualification, at 23:40 on a Sunday. A language model working through your questions and writing structured answers into the record means the first human conversation starts at the useful part instead of collecting basics.
Two rules make that safe rather than clever:
- Low confidence escalates, it does not guess. Anything the agent cannot classify goes to a person with the full thread attached, and the failure is logged so the rule can be fixed.
- Personalisation uses only facts already in the record. No inference about someone’s business from their domain name, no invented context. The most common way automated follow-up embarrasses a company is confidently getting a detail wrong.
What we will not build
An unsupervised agent talking to your clients on day one. Autonomous replies are reasonable for narrow, low-risk cases — confirming an appointment, acknowledging receipt — and that scope should be earned gradually rather than assumed.
We also will not automate a process nobody can describe. If the manual version of your intake is undocumented, automating it makes the confusion faster and harder to see.
Where this gets used
Automation
Automation here means the path an enquiry takes after it arrives: captured in one place, qualified by an agent asking what you would ask, routed to the right person with context attached, and followed up on a schedule rather than from memory.
More from the lab
Want something like this built?
Tell us what you are building and we will say whether this belongs in it.
Start a conversation