Embedded AI team · Paris
Most SMBs aren't behind on AI. They're behind on themselves.
Havoop is an AI team that works from inside your company. We start by writing down how it runs day to day, pick three to five workflows where an agent changes something, and build it with a person keeping the final say. Hosted in Europe.
Where do you stand?
Three questions, two minutes. They are the ones we ask at the start of every engagement. The result is a first reading on Turing Post's six-level maturity ladder, the one we use in our diagnostics.
What usually blocks it
You have tried ChatGPT, maybe Copilot. Someone on the team cobbled together an assistant that half works. And nothing changes in operations.
The blocker is almost always the same: the company cannot describe its own work in a form a machine can follow. The rules live in two or three people's heads. The process changes with each client. Every team has its own vocabulary.
An example we tell often. A bookkeeping firm serving restaurants wanted to automate supplier invoice entry from PDFs and photos. OCR was the easy part. The rest took six weeks: some suppliers filed fuel costs under sundries, others put bottle deposits there. Some reissued an invoice under a new number, others reused the old one. All of it had to be written down before an agent could use it. Once the rules were written, several supplier "exceptions" disappeared on their own.
That foundation is what we lay first. The agents come after.
55% of French small and mid-sized businesses use generative AI. 58% of their leaders call it a matter of survival. 32% have actually deployed it. Bpifrance Le Lab, 2025.
Case reported by Turing Post, "The Unsexy Truth of AI Adoption", April 2026.
Where to start
There are several doors in. The small ones exist to see whether we work well together before committing to a diagnostic.
- Frontier workshop
- Half a day, 5 to 15 people
- €2,500 excl. VAT
- Each person maps their tasks into three zones: where AI does most of the work, where it drafts and you correct, where it does not enter. Everyone leaves with a one-page rule of their own.
- Cold Start sprint
- 1 to 2 weeks
- €2,500 excl. VAT
- Six to twelve interviews and a map of the organisation as it actually runs. Also useful when taking over a company or a function.
- Diagnostic
- 2 to 4 weeks
- From €4,000 excl. VAT
- Map, maturity level, three to five costed automations, six-month roadmap. 50% deducted if we deploy afterwards.
- Agent deployment
- 4 to 12 weeks
- Quoted
- The agent built, integrated with your tools, in production with a person in the loop.
- Production sprint
- 2 weeks
- From €5,000 excl. VAT
- Audit trail, human in the loop, AI Act, GDPR. A file you can put in front of the board and the DPO.
- Sovereign deployment
- Monthly
- €1,499 excl. VAT / month per agent
- For professions bound by professional secrecy: the agent runs on infrastructure you control, audit log included.
Built for French reality
The tools come from the United States. Your constraints come from here. We know the difference between a delivery note and a pro forma invoice, and we build agents around:
- URSSAF filings and the monthly DSN
- VAT at 5.5%, 10% and 20%, and its categorisation rules
- your sector's collective agreements
- the handover with your chartered accountant
- the annual closing cycle
- mandatory invoice mentions and retention rules
No Havoop deployment is wired to a single model provider. In June 2026, Anthropic's most advanced models were cut off worldwide for eighteen days over US export controls. Production models were untouched, and a Havoop deployment would not have noticed: every workflow can switch providers, and every task runs on the cheapest model that holds the required quality.
For law firms, accounting firms and healthcare practices, we can go further: open models on infrastructure you control, data that never leaves your environment, a log of every action. We only propose it when an obligation justifies it.
It starts with a conversation
Thirty minutes on a video call. You describe your business, what eats your time, what you have already tried. At the end, we tell you whether we can help. If not, we tell you who to talk to.
How we work
Write down what you do.
One day on site, then six to twelve thirty-minute interviews. The format comes from Andrew Bosworth, who uses it at Meta when taking over a new team: twenty-five minutes of "tell me everything I should know", three minutes of "what's broken", two minutes of "who else should I talk to". After six interviews you have a map of the company that rarely matches the org chart. That document is what justifies the diagnostic.
Decompose each workflow.
What an agent does at each step comes down to seven moves: wait for a trigger, check, classify, enrich, produce, execute, ask. We note which ones chain together, and where a person decides. Then a simple test. Same input, same path every time? That is fixed-flow automation: two to six weeks. The path changes case by case? That is a reasoning agent: two to four months. Six or seven ideas out of ten fall in the first bucket, and that is where the return comes fastest.
Build one agent, with a person who decides.
Every agent starts in consultant mode: it proposes, you approve. After thirty days above 95% accuracy, it moves to approver mode on low-risk actions. Some actions will never be autonomous, and that is written into the contract.
We build one agent per workflow. On tasks nothing verifies automatically, every added agent is one more place where an error slips in silently. An analysis published by AlphaSignal in May 2026 measured a 17x error amplification between multi-agent and single-agent systems on the same task.
Hold it in production, and watch where the time goes.
An agent that works in a demo and an agent still working six months later are two different objects. The second has recovery points, memory separated from working context, a written list of what it may do, failure handling, and an isolated environment with limited access to your tools. We do not ship below those five points.
The time saved has to go somewhere too. We write down with you what the freed hours are reassigned to. Left alone, they dissolve into email and the return stays at zero.
For the first two or three months, a deployment that works and one that fails look the same from the outside: extra costs, no visible effect on results. So we set leading indicators from the start. If they have not moved by week eight, we stop and rework.
The whole method, with its sources