Automating customer enquiries with human approval

The fear: "it will say something we can't take back"

Enquiry automation rarely stalls because the drafts are bad. It stalls because a message sent under your business name is irreversible: a wrong price you now have to honour, a promise about delivery you can't keep, a tone that doesn't sound like you. One bad send costs more trust than fifty good ones earn. So teams do the maths and keep typing every reply by hand — and the enquiries that arrive at 9pm wait until morning.

The mistake is treating this as a choice between two extremes: a human writes everything, or software sends everything. There is a third design, and it is the one Olano is built around: let the agent think freely, and gate the actions. Human-in-the-loop is not a feature bolted on afterwards — it is the architecture.

The design answer: gate actions, not thinking

Split every enquiry into two halves. The first half is reading and preparing: understanding the thread, checking live systems, pulling the right answer from a knowledge base built on your own documents — price lists, policies, product specs, SOPs — and writing a grounded draft. All of that is read-only or internal, so it runs without gates, at any hour, in any of your channels.

The second half is acting: sending the reply, updating a record, touching money. On Olano, anything outbound waits for approval by default. The agent does the minutes of work; a human spends seconds deciding whether it leaves the building.

That queue is not a permanent tax. How much waits in it is set by a trust level — and trust is a dial you turn, not a switch you flip.

The trust ladder: levels 0–4

Every Olano agent runs at an explicit trust level. Here is what each one means for enquiry handling.

Level 0 — Observer

The agent reads and reports, nothing more. Point it at your WhatsApp number or shared inbox and it tells you what customers asked this week, what went unanswered, and which questions repeat. It sends nothing and changes nothing. This is how you learn what an agent would do before it does anything at all.

Level 1 — Assistant

The agent drafts. Every enquiry gets a proposed reply, grounded in your documents, waiting in the queue. Your team stops writing from scratch and starts reviewing — which is faster, and safer, because a draft that never sends can never embarrass you. Most businesses start one workflow here.

Level 2 — Collaborator

The agent works in multi-step jobs: it reads the enquiry, checks stock or the calendar, drafts the customer reply and the CRM update together, and presents the whole package. Each outbound action still stops at the gate — you approve outcomes rather than keystrokes.

Level 3 — Autonomous

Routine, in-scope replies send without per-message sign-off: opening hours, live stock answers, delivery status, booking confirmations. This is not hands-off automation — the per-category gates below still hold, anything unusual still queues, and every send is logged. Autonomy here is scoped, earned from review history, and reversible.

Level 4 — Developer

The widest setting, meant for technical workspaces where an agent works on files and code inside its own private, isolated workspace. Even here, the gates on file and shell writes and on secrets still apply, and everything is audit-logged. Enquiry handling never needs level 4 — it is on the ladder so the ceiling is explicit, not a workaround.

Per-category approval gates

Trust levels set the default posture; category gates cut across it. Five kinds of action carry their own gate, whatever level the agent runs at:

• Financial actions. Refunds, charges, anything that moves money. A level-3 agent that answers stock questions freely still queues a refund for a human.
• External sends. Any message leaving under your business name — replies, follow-ups, notifications.
• File and shell writes. Changes to files or systems, as opposed to reading them.
• Secrets. Access to credentials and API keys, which are stored encrypted.
• Delegation. Handing work to another agent — even inside a team of specialists, hand-offs follow the same rules.

This is what makes widening autonomy safe: promoting an agent to level 3 for routine replies does not open the door to anything risky, because risk is gated by category, not by optimism.

The audit trail

Every action — what the agent read, what it drafted, who approved it, what was sent and when — lands in an immutable audit trail. In practice that record earns its keep three ways. When a customer disputes what they were told, you look it up instead of reconstructing it from memory. When drafts keep going wrong in the same way, the trail shows you the gap in your documents. And when you are deciding whether to raise a trust level, you decide from a record of hundreds of reviewed actions, not a feeling.

A rollout playbook

Start with one workflow at low trust. Pick the single channel where enquiries hurt most — for many SMEs that is WhatsApp customer service; for others it is the shared inbox, where an AI inbox assistant triages and drafts. Run it at level 1. A managed workflow like this is typically live within 1–2 business days of onboarding — live in days, not months.

Review the queue daily. In the first weeks, approving is teaching. Every decline tells you something specific: usually that a price list is stale or a policy was never written down. Fix the document, not the draft, and the next hundred drafts improve.

Widen scope as confidence grows. When the queue turns into approve, approve, approve for a category — say, stock and hours questions — promote that category towards level 3 and keep your attention for the messages that deserve it. Financial actions and other gated categories stay queued regardless. Bigger rollouts across more channels and workflows take 1–3 weeks, and every step is quoted and approved before we build.

Where the approvals live

Approvals only work if reviewing them is effortless, so they come to you. Approval requests arrive as ordinary messages in the channels your team already works in — Slack, Telegram, WhatsApp, email, or another of the 15+ channels Olano supports. Reviewing the queue is answering a message, not babysitting a new dashboard. The result, from the customer's side, is a private AI front desk that responds quickly around the clock; from your side, a queue that supports your team instead of replacing your judgement.