Time given back to your teams
No more copy-pasting from one tool to another. Your teams focus on the work that really needs a person.
Automation
n8n automation, with or without AI: less manual re-keying
A manual task that kept your teams busy disappears for good, in two weeks. We only bring in AI at the steps where a rule is not enough, and you know at all times what the automation costs you.
n8n, what is it?
n8n is automation software whose code is public, published by a Berlin-based company. You draw a process as a diagram: each box is a step (an email received, a record created in the CRM, a message sent), and the links say in which order they follow one another.
It connects to hundreds of off-the-shelf tools, from the CRM to invoicing, messaging apps and spreadsheets, and to the API of those without a ready-made connector. It runs on your server or ours: your data does not go through a third-party service, and the bill does not climb with the number of operations.
AI is one box among others. We place it at the step that needs it, and only there: the rest of the diagram stays made of rules, faster, cheaper and always the same.
- Email received , then Sorted by Jev
- Sorted by Jev (quote), then Quote in the CRM
- Sorted by Jev (complaint), then Draft reply
- Sorted by Jev (unsure), then A person decides
- Quote in the CRM , then Sales rep alerted
- Draft reply , then Reviewed, then sent
A request arrives by email. Jev files it, a rule creates the quote in the CRM, a language model drafts the reply to a complaint, and a person decides as soon as the model hesitates.
- Site form , then Record in the CRM
- Record in the CRM , then Ideal customer?
- Ideal customer? (priority), then Sales rep alerted
- Ideal customer? (unsure), then A person decides
- Ideal customer? (to nurture), then Email sequence
A contact fills in the site form. A rule creates their record, Jev compares them with your ideal customer, and the sales rep is only called for priority contacts; the others get an email sequence.
- Invoice received , then PDF read
- PDF read , then Matching
- Matching (matches), then Accounting entry
- Matching (mismatch), then Mismatch reviewed
A supplier invoice arrives. A language model reads it, the code compares it with the purchase order, and only a mismatch needs a person to look.
- Order confirmed , then Invoice created
- Invoice created , then Payment received?
- Payment received? (yes), then Shipping started
- Payment received? (no), then Customer reminded
An order is confirmed. Invoice, payment check, shipping or reminder: everything follows rules, with no call to a model. Many processes look like this one.
- Free-form request , then AI agent
- AI agent , then Safe action?
- Safe action? (safe), then Action carried out
- Safe action? (risky), then Human approval
- AI agent : Model
- AI agent : Memory
- AI agent : Tools
A request written in free form. The agent searches over several rounds, with its model, its memory and its tools, but every action goes through Jev before it leaves, and anything risky waits for a person.
Swipe to follow the diagram
- Rule, no AI
- Decision model
- Language model or agent
- Person
It connects to your tools
A few of the tools n8n links, among hundreds of connectors. For the others, it goes through their API.
Twenty Odoo Dolibarr HubSpot Shopify Stripe Gmail Outlook Google Sheets Excel Notion Airtable
Slack Microsoft Teams Telegram WhatsApp PostgreSQL Supabase Jev Clef Mistral Claude Qwen Webhooks et API
What you gain
Fewer data-entry errors
Information entered once reaches everywhere it is needed, with no typos or omissions.
No silent failures
If something gets stuck, you are alerted and the work resumes where it stopped. Nothing is lost along the way.
A known, controlled cost
You see what each run and each month costs. No per-operation bill that climbs with your activity.
Your data stays with you
The automation runs on your server or ours, not with a third-party service that bills by usage.
A risk-free transition
The old way of working remains available until the new one has proven itself.
Who it is for
- Your team copies orders, requests or invoices from one tool into another.
- You want a digest every morning without anyone spending time preparing it.
- Your CRM, your invoicing and your messaging do not talk to each other.
- Your teams sort emails, requests or files by hand every day.
- You want AI in your processes, but only where it brings in more than it costs.
What you receive
- Your process automated end to end, in production.
- Alerts if something gets stuck, and a log of every run.
- A meter showing the cost per run and per month.
- For each step, the written choice between a rule, an AI model or a person, and what it costs.
- Clear documentation and your team trained.
- A checkpoint at 30 days.
How it goes
-
Scoping
We watch the task with the person who does it today, and the whole process around it: the tools in place, what connects to them and what does not, what the team is ready to change. For each step, we decide whether a rule is enough or whether AI brings in more than it costs.
-
Build
We build the automation and test it on your data.
-
Go-live
It takes over gradually, with the old way kept as a fallback.
Follow-up
Alerts in place and a checkpoint at 30 days.
With or without AI: six example workflows
Without AI: the quote request filed and assigned
The form on your site creates the record in the CRM, assigns it to the sales rep for the area based on the postcode and alerts management above a certain budget. The fields are already structured: a rule is enough, with no paid call to a model and the same result every time.
Without AI: the order followed end to end
A confirmed order creates the invoice, updates the stock and the customer record, then reminds the customer if the payment does not arrive. Nothing to interpret, so nothing to hand to AI: a fixed sequence, cheaper and safer.
With Jev: the inbox sorted
Each incoming email is filed (quote, complaint, invoice, urgent) by a decision model such as Jev, from TypeSafe AI. It writes nothing: it picks from a closed list and says how sure it is. Above the threshold, the workflow passes it to the right team; below it, a person decides.
With Jev: the prospect compared with your ideal customer
A new contact comes in through the site or a trade show: the decision model compares it with your ideal customer, spots buying intent and sets its priority. Your sales reps start their day with the contacts that matter, and keep the first conversation and the negotiation.
With a language model: the supplier invoice read and matched
The PDF received is read, its references and amounts are extracted, then the workflow compares them with the purchase order. The arithmetic stays with the code, which does not get sums wrong: a mismatch goes to review instead of being paid.
With a person: the decision that commits you
An off-grid discount, a refund, anything that affects an employee or a candidate: the workflow prepares the file, and a person approves in one click from their messaging app. AI can sort and prepare, never decide alone about a person.
Frequently asked questions
Why n8n rather than Make or Zapier?
n8n runs on your side: no per-operation bill that explodes with volume, and your data remains yours. It is our default choice, not an exclusivity.
What happens if the automation goes down?
You are alerted, and it resumes where it stopped. An automation that fails silently is worse than a manual task, so we do not deliver one that way.
What if we have five processes to automate?
We start with the one that costs you most, get it running, then move on to the next. When everything seems worth automating, the Audit Sprint first reviews your processes, your tools and the people who use them: it tells you in which order to proceed, what the team needs to learn, and what is not worth automating.
Do you need AI to automate?
Often not. Most processes follow simple rules: this field filled in, that action. AI comes in where a rule is no longer enough, such as reading an invoice or sorting a request, and we combine several models, each on the task it does well at the lowest cost, with a cache so that the same answer is never paid for twice. On Palmora Property, switching to an AI model divided the translation cost by 60, and a cache serves 95% of segments without calling the model again.
What does a decision model such as Jev bring?
Many decisions in a process need no sentence at all: filing an email, judging whether an action is risky, choosing the team that will handle a request. A decision model such as Jev, from TypeSafe AI, or Clef, from Cloudflare, answers them with a choice from a closed list and a probability, faster and far more cheaply than a large model used for sorting. We first test it on your own cases, it then works alongside the team without acting, and the threshold that lets it act alone is set on what it has shown. Jev is a young service hosted in the United States: the decision therefore goes through your own entry point, and switching to Clef, whose open weights install on your side, remains a configuration change.
What about an AI agent in the automation?
Only when the path cannot be planned in advance, for example a free-form request that must be understood before it is filed in the right place. The agent then runs in a loop: it acts, checks its result and starts again until it reaches the goal. In n8n, we cap its number of rounds, limit its permissions to what it needs, have a decision model judge each sensitive action before it goes out, and have a person approve anything that commits the company. The rest of the process remains a fixed sequence, safer and cheaper.
20-minute diagnostic
Twenty minutes on a call, two or three concrete leads, no strings attached.
We look at your situation, identify what can be automated or improved, and tell you plainly whether we can help.