4fleet uses AI (Anthropic's Claude) as a command layer over your live business data. It reads inbound emails, drafts replies, helps you quote, and - through the chat agent - can actually carry out actions on your behalf. This article explains what each AI feature does, where it is genuinely AI, and how usage is metered so you do not get surprised.
The AI chat agent (it takes actions)
The chat agent is the centrepiece. It has your live business data injected into its context and can both answer questions and execute real actions against your account. Ask it to assign a driver, change a booking status, generate or send a quote, create a task, generate an invoice, or update a client, and it does it - not just describes it. There are around two dozen action types it can perform.
Two safeguards matter here. First, delete actions require an explicit confirmation. Second, you can control which actions the agent is allowed to auto-perform from the AI settings panel. Send actions (emailing a quote or invoice to a customer) reach the customer in a single turn, so review your auto-action settings before you let it run unattended.
AI email parsing
When a booking-request email arrives, AI extracts the structured details - names, pickup and dropoff, date and time, passengers, luggage, flight, extras, multi-trip legs - and turns the email into an enquiry. It also disambiguates the booker from the passenger, detects modification requests in replies, and pulls out the customer's explicit questions so you can be sure you have answered them.
AI quoting - where it is and is not AI
This one trips people up. On the enquiry screen, the main quote buttons use the deterministic pricing engine - your configured rates, applied the same way every time - even though some are labelled "AI". A genuinely AI-composed quote (Claude writes the price and email) exists only through the chat agent's quote action. For predictable, defensible numbers, rely on the deterministic engine; use the AI-composed quote when you want a narrative draft you will review.
There is also a grounded, human-in-the-loop email composer: it drafts a reply answering a customer's questions using only verified facts from your database (the live quote, enquiry, and vehicle data), so it does not invent details. It is never auto-sent - you always review and send.
AI memory
The agent has a per-tenant memory: facts it remembers across conversations. You can add, edit, and delete facts manually, and the system also auto-extracts facts from completed conversations. Only approved facts are used in the agent's prompt, and you can require approval before any auto-extracted fact takes effect. This keeps the agent on-brand and consistent with how you actually operate.
Other AI helpers
- Smart replies, draft reply, sentiment, and categorisation on inbound emails.
- Dashboard insights - a few prioritised business insights refreshed twice a day.
- Grounded funnel narration - a plain-English read of your lead-to-revenue funnel, with a verifier that rejects any number not present in your real data.
- Reply classification - inbound replies to follow-ups are classified (accepted, declined, question) so sequences branch or stop automatically.
A few screens are branded "AI" but are actually rule-based templates (the enquiry assistant and the dashboard dynamic-pricing widget). They are useful, but they are not doing live analysis.
How AI usage is metered
AI work is charged against a per-company AI operations quota and gated by your subscription tier:
- Every metered AI call reserves an AI op before it runs and rolls back if it fails, so failed calls do not burn your quota.
- If you hit your limit, AI actions return a clear "limit reached" response rather than failing silently.
- Your tier also affects which model runs your requests: lower tiers route more work to the fast, cheaper Haiku model; higher tiers use the more capable Sonnet model for reasoning tasks, and the top tier can opt into a premium model for the chat agent.
- An AI usage and cost dashboard (available on the Business tier) breaks your spend down by operation type and model with a daily trend, so you can see exactly what your AI features cost.
The bottom line
Let the agent handle the repetitive work, keep pricing on the deterministic engine, review anything that goes to a customer, and watch your AI-ops quota via the usage dashboard. The AI is built to be grounded in your real data - it is designed not to invent numbers or details it cannot verify.