Your team already works in WhatsApp. So should your AI.

Most software wants you to come to it. You learn the dashboard, you remember to check it, you train your team to live inside it. AI agents have mostly inherited that habit: a new tab, a new chat window, a new pane of glass that someone has to remember exists. The trouble is that the people running a mid-market business already have a command surface, and it is not your dashboard. It is the email thread with the supplier, the WhatsApp group with the site team, the Teams channel where the finance questions land.

Usermode's agents work where the operators already work. Sarah chasing an overdue invoice, Henry confirming a gas safety re-inspection, Jacob asking whether to submit a tender — none of them asks a person to open a new app and check a queue. They reach the right human on the channel that human already lives in, ask a precise question, and act on the reply. The channel is not a notification layer bolted on at the end. It is where the work happens.

Why The Channel Is The Product

An agent that does real work has to do two things a chatbot never has to: reach a specific person at the right moment, and accept an instruction back. Both of those are channel problems before they are intelligence problems.

If the only way to interact with an agent is a portal, every interaction carries a tax. The operator has to remember the portal exists, navigate to it, find the relevant item, and respond — all while the actual decision is a five-second "yes, send it." Multiply that by a credit controller working dozens of overdue accounts, a compliance manager tracking hundreds of safety certificates, a bid writer juggling live tenders, and the tax compounds into the thing every ops tool eventually becomes: another place nobody checks.

The best interface for an agent is the one your team didn't have to be told to open.

Email, WhatsApp, Teams and SMS are not a concession to legacy habits. They are the surface where decisions already get made — where a director clears a payment plan, where a site manager confirms an asset is offline, where a finance lead says "hold that one." Putting agents on those channels means the agent meets the work instead of asking the work to relocate.

In Band Approvals: Ask, Reply, Act, Log

The most important pattern this unlocks is the in-band approval. An agent reaches a decision point it isn't allowed to pass alone — a spend, a sensitive send, a judgement call — and instead of stalling or guessing, it asks a human, in the channel, in plain language.

Concretely, that looks like this. Sarah has worked an overdue ledger and wants to offer a tenant a payment plan that exceeds her standing authority. She doesn't open a ticket. She messages the right person on WhatsApp: "Unit 5 North House is 32 days overdue on a hypothetical £2,140. I'd like to offer three monthly instalments and pause the late-fee escalation. OK to send?" The director replies "yes." Sarah sends the agreed wording to the tenant, records the instruction, and the whole exchange — request, human approval, resulting action — lands in the audit ledger.

Three things make this safe rather than reckless:

• •The agent cannot act first and apologise later. External sends pass through a fail-closed tool policy. Every outbound message needs a signed authorisation — HMAC-SHA256, time-limited, bound to the specific recipient — so a "yes" approves that send, not a general licence to email.
• •The human reply is the gate, not a formality. Spend and sensitive actions sit behind explicit approval. No reply, no send. The agent escalates rather than quietly proceeding.
• •Everything is written down. The approval and the action it authorised are captured in a tamper-evident, append-only record. "Who said yes to this?" always has an answer.

The human experience is a single message and a one-word reply. The machinery behind it — signing, recipient binding, logging — is invisible to the person, which is exactly the point. Governance you have to visit is governance you skip. Governance that rides along inside a WhatsApp reply is governance that actually happens.

Attachments Are Where The Real Work Lives

Operational decisions almost never happen on text alone. They happen on a document — the bid pack, the rent statement, the inspection certificate, the reconciliation. An agent that can talk but can't carry a file is stuck describing things it should simply be handing over.

So the channels move documents both ways. Jacob assembles a branded, priced tender pack and sends it as a real attachment to the person who has to approve submission — not a link to yet another system, the actual PDF, in the email or Teams message they're already reading. Aaron can attach the month-end reconciliation. Henry can forward the certificate that proves a safety check is back in date. And it runs inbound too: a human can reply to an agent with a scanned document attached, and the agent reads it — including scanned and image-only PDFs, via an OCR pipeline — rather than throwing its hands up at a non-text file.

This matters because the attachment is usually the decision. "Approve this" means nothing without the thing being approved sitting right there. By keeping the document in the channel where the conversation is, the agent collapses the gap between "here's what I propose" and "yes, do it" into one continuous exchange.

Untrusted inbound content — anything a human or outside party sends in — is sandboxed against prompt injection before the agent reasons over it, so an attachment can't smuggle instructions past the governance gates.

Time Critical Nudges And The Right Channel For The Moment

Not every message deserves the same channel, and choosing well is part of doing the job properly. A routine update can sit in email. A decision that's blocking work belongs in WhatsApp or Teams, where people actually look. And something genuinely time-critical — an approval window about to close, a deadline a few hours out — warrants an SMS nudge, because SMS is the channel people still check when everything else is muted.

This is where Usermode's operating model shows its hand. Agents run on standing daily and weekly schedules and act when something needs doing — they don't wait to be prompted. So the agent that notices a tender deadline tightening, or a compliance date slipping, isn't reacting to a human poking it. It got there first, on its own schedule, and now it needs a person for the one part it can't do alone: the decision. The nudge is the agent escalating with urgency proportionate to the moment.

It ties back to a rule that sits underneath all of this: the delivery contract. A run cannot end without a real, logged outbound action. An agent is not allowed to go quiet — to "finish" by doing nothing and saying nothing. If it can't complete the work, it escalates to a human, on a channel that human will see, rather than leaving a silent gap where someone assumes it was handled. The channel choice is how that escalation reaches the right person in time to matter.

What This Replaces

Put the pieces together and you get an operations layer with no dashboard to adopt. The agent reaches the operator on email, WhatsApp, Teams or SMS; asks a clear question when it needs a human; carries the document the decision depends on; and writes down what was approved and what it did. The human side is a message and a reply. The governance — signed authorisations, approval gates, the audit ledger, per-tenant Azure isolation, prompt-injection sandboxing — runs underneath without asking anyone to manage it.

This is proven live, not theoretical: at Legacie in property development and block management, and at WH Scott Group across industrial lifting and inspection. Named agents — a credit controller, a bid writer, a compliance manager, an executive assistant — doing the work through the channels their colleagues already use.

The test for any operational AI is simple. Did it make someone learn a new place to check, or did it meet them where they already were? If your team works in WhatsApp, that's where your AI should be working too.

If you'd like to see an agent ask, get a reply, and act on it in your own channels, you can book a walkthrough at /demo.