Meet hermes: the AI teammate that runs our back office
The teammate that works before anyone's had coffee
We call it hermes — an AI agent that runs quietly in the background of our business. Before most of the team had their first coffee this morning, hermes had already scanned the inbox for supplier invoices and filed them into the books, checked every active client project for anything slipping, drafted the messages that needed a nudge, and sent a one-line status to each project owner. It works on a schedule, and it never forgets a step.
That's the shift worth paying attention to. Not a chatbot you talk to — a teammate that does the work. Here's a look under the hood: the research on why it matters, four jobs hermes actually runs today, and how a business like yours could have one.
Why your best people spend so little time on their real job
Before AI, the maths of a knowledge-work week were already sobering. McKinsey's landmark study of the "social economy" found the average interaction worker spends about 28% of the week just managing email, and nearly 20% more looking for internal information or tracking down the colleague who has the answer. That's close to half the week gone before anyone touches the work you actually hired them for. (McKinsey Global Institute)
What changed recently is that this exact category of work — reading, sorting, drafting, looking things up — is now the most automatable. In its 2023 analysis, McKinsey estimated that generative AI and related tools could automate activities that today absorb 60–70% of employees' time, precisely because these models finally understand natural language well enough to handle the messy, text-heavy tasks that used to require a person. (McKinsey, 2023)
Read those two findings together and the opportunity is obvious: the work that eats your team's week is the same work an agent can now quietly take off their plate. hermes is our attempt to actually do that — not in a demo, but in the day-to-day running of a real company.
A chatbot answers. An agent acts.
The word "AI" now covers two very different things, and the difference is everything for a business:
| A chatbot | An AI agent | |
|---|---|---|
| What it does | Answers when asked | Watches, decides, acts |
| Where it lives | A chat window | Inside your tools & data |
| Who starts it | You, every time | An event, or a schedule |
| What you get | A helpful reply | A task actually done |
Four jobs hermes actually runs today
None of this is hypothetical. Here are four real workflows hermes owns for us right now — each one a place where, in most companies, hours quietly disappear.
1. Accounting — from inbox to books, on its own
Supplier invoices arrive as email attachments all week. hermes watches the inbox, recognises an invoice, extracts the supplier, amount, VAT and date, matches it to the right account, and files it — attachment safely stored, ledger entry drafted. Anything ambiguous (a new supplier, an odd amount) it flags for a human instead of guessing. The month closes with the paperwork already done, not waiting in a shoebox. Use case: an SME that spent a full day a month on expense entry now spends minutes reviewing.
2. Project management — a morning briefing, and nothing forgotten
Every morning hermes reviews every active client project — deadlines, milestones, what shipped, what's overdue — and sends a single WhatsApp digest: overdue items first, meetings today, calls from yesterday, client emails awaiting a reply. Across dozens of projects it's the difference between "I think we're on track" and a live, honest picture. When a project needs a client update, hermes drafts it in the client's language and waits for a yes. Use case: cutting a 46-project portfolio's status-chasing from hours of Monday meetings to a 60-second read.
3. SEO — a weekly audit no one has to remember
hermes crawls a client's site on a schedule, scores it across on-page factors, and surfaces only what changed or broke — a missing H1, a slow page, a structured-data error — with the fix. It runs whether or not anyone remembered to check, and pings a human only when something genuinely regressed. Use case: turning SEO from an occasional, forgettable chore into a standing weekly health check.
4. Content — a trilingual blog on autopilot
hermes also runs a content workflow: it picks a topic, writes it in French, English and Arabic, generates a unique themed cover, and sends the whole thing to a human as a preview to approve — then publishes on a yes. The loop is the same one that files the invoices: prepare, preview, get a human's approval, act. Use case: a consistent trilingual blog without a standing content team.
A week in hermes's life
Rolled up, a normal week of an agent embedded in a real business looks like this:
The one rule: it asks first
An agent with access to your business needs guardrails, and this is where a good one earns its trust. hermes follows one rule above all others: nothing goes out into the world without a human "yes." It prepares everything — the invoice, the client email, a blog post — as a preview, sends it for approval, and only acts once a person signs off. Every action it might take can be dry-run first, with zero side effects, so you can see exactly what would happen before it happens.
Where hermes draws the line
An agent you trust is defined as much by what it won't do as by what it will. hermes doesn't decide who to hire, doesn't move money, and doesn't send a single email, invoice or post to anyone without a person approving it first. By default it has read-only access to sensitive data; write access exists only through narrow, logged, reversible steps. It's a teammate, not an autopilot — the judgement calls that carry real consequences stay firmly with a human. That boundary isn't a limitation we're apologising for; it's the thing that makes handing over the next job feel comfortable instead of nerve-wracking. An agent that can't be trusted with a small task will never be trusted with a big one, and trust is built one approved action at a time.
Under the hood
hermes isn't a model we trained — it runs on the open-source Hermes Agent from Nous Research, extended with our own skills (accounting, project management, SEO, content) wired into the real tools. Each skill is documented, testable and approval-gated. That modularity is the point: you add one skill at a time, prove it, and only then hand over the next.
The open-source agent framework hermes runs on — "the agent that grows with you". It reads events and documents from your own tools, does the routine judgement work, and escalates only the decisions that need a human; we layer Azinove's own skills on top.
What it means for your business
You don't start with "an AI that runs everything." You start with the one job nobody enjoys and everybody does — filing invoices, chasing updates, first-draft replies — and you give an agent exactly that, with a human keeping the final say. If the research is right that most of a knowledge-worker's week is now automatable, the real question isn't whether an agent can help, but which repetitive corner of your business to hand it first.
In practice, the first month is small on purpose. Pick the one workflow. Wire the agent to the tools it needs. Run everything in preview for a week, so you watch its judgement before it ever acts. Then let it act — with a human approving each step — and measure the hours that come back. One workflow proven that way tends to make the next three obvious.
hermes proved itself on one workflow, then earned the next. If you're curious where an agent like it would give your team back the most hours, that's exactly the conversation we love to have.