When someone books with Altimood, they’re not just booking an activity or a trek. They’re reaching out. Directly.
With Alice, we created Altimood for exactly that: to remain accessible guides, truly reachable. Not a platform, not an impersonal agency. Just two guides who respond, organize and accompany, from the first contact to the return home.
Except some of our offerings started doing really well. Too well, even. The flow of requests exploded, and the risk was getting sucked in: only doing what sells easily, dropping the projects that require more energy but that we’re passionate about.
As freelancers, we want to keep the joy of doing everything. Not become a volume-processing machine. But behind every email, there are real people waiting for a response. And each response takes time. In the evening, after a day outdoors, you still have to open the laptop.
And every freelancer knows this: administrative work tends to eat up all available space. It invites itself to dinner, creeps into family evenings and ends up stealing that precious disconnection time needed after a day in the field. It becomes an invisible chore that sometimes overshadows the very heart of our work.
To avoid this fate, I simply used the other side of my skill set. My second hat is web development and acquisition consulting (piedweb.com). It’s not just a separate contractor activity — it’s the lever that lets me run Altimood intelligently: where the guide was starting to saturate, the developer built a solution.
Not a chatbot. Not an auto-reply. Not a robot speaking “on behalf” of a human. I built an AI assistant that drafts my replies for me. I review them, add a sentence, a joke or a detail about the current snow conditions… and hit send.
AI doesn’t replace the relationship. It just gives me back the time to truly be present in it.
Here’s how I designed this tool.
The solution: a secretary who prepares, a guide who validates
The principle is simple: every 15 minutes, a program scans my new messages, sorts them and prepares drafts directly in my inbox.
When I open my laptop in the evening, the heavy lifting is already done. I review, adjust a detail if needed, click Send.
The AI never responds on my behalf. The client receives an email signed by Robin, reviewed by Robin, with Robin’s tone.
The difference is that it takes me 10 minutes instead of an hour, and the information is sometimes much better organized.
How does the assistant know what to reply?
AI is not a magic wand: its relevance depends entirely on the solidity of the context you provide.
If you throw it a raw email without explanation, it will fill in the gaps by inventing details. This is what’s called hallucination.
With two years of experience designing AI-powered processes for my clients at PiedWeb, I can assure you that the quality of the response is directly linked to the quality of the input data.
It’s the famous Garbage In, Garbage Out.
For my assistant to be truly useful and accurate, I give it access to three concrete information sources:
- My calendar (CalDAV/Nextcloud): it knows if I’m available or already engaged with a group, without ever revealing other clients' names.
- My response templates: templates I’ve written for each recurring case (trip confirmation, fully booked, referral…). They contain my vocabulary, my safety advice and my usual tone.
- My product catalog and website: up-to-date prices, required levels, technical specs. The assistant can draw raw data from where it lives and insert the right links in responses.
Note: Point #3 pushed me to keep my website constantly updated. At first, it was just small corrections. But very quickly, I realized that the clearer and more complete my site was, the better my assistant could help me respond quickly and well. Today, updating it has become almost a game: I prepare the content, the assistant suggests improvements, I correct a few details, and everything is ready for clients. It’s a true virtuous cycle: the more precise the site, the more human and personalized my responses remain. I’ll tell this story in another post, because it led me on a journey I never would have imagined at the start.
If a client asks: “Is there still room for the next igloo workshop on Saturday February 21st for 2 adults and 2 children?”, the AI checks my schedule, confirms availability, finds the exact price and prepares an email with the link to the technical sheet. It even goes so far as to ask the question I would have asked myself: “How old are the children?” (to adapt equipment or messaging).
Strict rule: never invent a price, a date or a link. If information is missing, the AI places a [TODO] that I spot immediately during review.
Automatic sorting
Not all emails deserve a reply. The assistant categorizes each message:
- Draft reply for real client requests.
- Skip for newsletters, invoices, mailing lists, spam (including phishing, like that fake ANTAI fine detected automatically).
- Newsletter signup: when someone subscribes via my site, the contact is added directly to my database.
Result: I only see in my drafts the emails that require my attention.
The field bonus: voicemail messages
When I’m in the mountains, I sometimes receive voicemail messages. My system retrieves them, transcribes them to text (locally on my machine) and sends me a summary by SMS with a draft reply.
Handy when I’m coming down from a hike with just mobile coverage, to know if it’s urgent or a simple question.
Under the hood (for the curious)
The stack is intentionally simple: a PHP program launched by cron, orchestrating everything via command line.
Cron (every 15 min)
→ IMAP retrieval of unread emails and filtering (whitelist outside process)
→ Extraction to my contact database
→ Calendar download (CalDAV)
→ Send to Claude with system prompt + templates + catalog (website)
→ Deposit drafts via IMAP
I currently use Claude Opus 4.6 (Anthropic). The model excels at following strict instructions without inventing information. But the architecture isn’t locked to a provider: the script sends a system prompt and context files. Any LLM capable of following structured instructions could take over.
I’ve also tested open-source models locally. The quality isn’t there yet for this use case (respecting a strict output format, handling multiple emails in batch, cross-referencing calendar and catalog). But it’s progressing fast, and the day a local model does the job, the transition will be trivial.
Personal data anonymization
This is an essential point for the ethics of our profession, but it’s also a strict obligation in Europe under GDPR (General Data Protection Regulation). You can’t send your clients' contact details to a third-party provider’s servers without solid guarantees.
Before each AI call, an anonymization layer replaces names, emails and phone numbers with pseudonyms (Contact-001, contact-001anon.local ). Claude works exclusively with these neutral codes, and my system restores the real values when creating the draft.
Result: no personal data leaves my infrastructure in clear text. It’s security for my clients, but also legal peace of mind for me. With a future move to a local model, this step would become superfluous since nothing would leave my machine. It’s one more argument for keeping a flexible architecture decoupled from any single provider.
What it actually changes
Since the assistant has been running:
- Efficiency: I never start from a blank page anymore. Pure typing time and date research have disappeared. I go from 1 hour of laborious processing to 15-20 minutes of active review and personalization. Part of the mental load tied to admin work has evaporated.
- Responsiveness: drafts are ready 15 minutes after receipt. Even if in my case it’s not always exact, in our line of work, the one who responds quickly and well is often the one who converts the contact into a commitment.
- Consistent quality: the AI never forgets to mention the equipment list or cancellation conditions, even when I’m tired.
- Nothing falls through the cracks: newsletter signups captured, invoices sorted, spam eliminated.
And above all: direct contact remains intact. That pillar on which we built Altimood. Every email sent passes through my eyes. If a client has a special request (a disability, a custom project), I take back the keyboard. The AI handles repetition, leaving me time to handle the human side.
Tips if you want to do the same
- Structure your info. AI can’t guess your prices. If your website is clear and your templates well-written, the AI will be effective. This is the real productivity lever, even before plugging in an AI.
- Provide context, not just the email. Calendar, catalog, history… the more context the AI has, the less you edit. Put yourself in your assistant’s shoes — what do you need to respond?
- Always keep the last click. The draft is the right compromise: you keep control without losing the time savings. It’s your signature and your reputation at stake.
- Protect the data. Anonymize what you can, especially with a cloud service. Or aim for a local model when quality catches up.
- Filter upstream. Newsletters and spam shouldn’t even reach the AI. Simple filters by sender or subject reduce noise and cost.
Possible improvements / Going further
The current system does the job, but several paths tempt me:
- Move to a local model. It’s the natural evolution. The day an open-source model reaches sufficient quality for this use case, no data leaves my machine at all. Anonymization becomes unnecessary, API costs disappear, and I gain complete autonomy.
- Handle multiple channels. Today it’s email and voicemail. Tomorrow, it could include WhatsApp messages and SMS — the principle remains the same: retrieve the message, enrich with context, propose a draft.
- Learn from my corrections. Every time I modify a draft before sending, it’s a signal. Collecting these gaps would allow refining templates and instructions over time, without touching the model itself.
- Connect my ticket/quote/invoice system. For now, the assistant prepares the reply but doesn’t handle the administrative side. Automatically generating a quote or triggering an invoice from an email conversation, always validated by me before sending, would save yet another layer of management time.
Automation often scares people, especially in trades driven by passion and human connection like ours. We fear becoming cold, distant, standardized.
My experience with this “AI secretary” proves the opposite: it’s by entrusting the robotic work to the machine that you keep your energy for what’s human. This system isn’t about working less — it’s about working better. It lets me come down from the mountain with a light mind, knowing I can respond to my clients with the attention they deserve, without sacrificing my evenings.
If you’re a guide, craftsperson or freelancer, don’t see AI as a competitor that will distort your profession. See it as a tireless and rigorous third hand you’ve always dreamed of having.
Your turn now.
AI, SEO and Web strategy session
Is this automation worth building?
Validate the expected gain, architecture, and guardrails before connecting your data and business tools.
€350 excl. VAT, up to 90 minutes
Also read: