What we're testing
Havoop is a studio, and a test bench. Every engagement tests a hypothesis about what an SMB needs to buy. What holds up becomes an offer. What comes up often enough becomes a product. Here is where we stand.
Three hypotheses in progress
Cold Start interviews run asynchronously by an AI produce an organisation map as accurate as a consultant's interviews, at a tenth of the cost.
- Why it might be true
- Without a consultant across the table, people speak more freely. Capture is word for word, cross-referencing between interviews is faster, and the 25/3/2 format is mechanical enough to automate.
- How we test it
- On the same engagement, a consultant runs the classic interviews and the AI runs the asynchronous version. We compare the two maps.
Documenting your workflows for AI improves those workflows by 15 to 20%, even if no AI is ever deployed.
- Why it might be true
- Writing forces clarity. Explicit rules get corrected. Dependence on a single person drops. The two cases published by Turing Post (an accounting firm, a construction company) show exactly this side effect.
- How we test it
- Productivity measured on documented workflows, before any deployment, across six engagements. Numbers published afterwards.
A difficulty met in three independent engagements predicts a tool other SMBs would buy.
- Why it might be true
- SMBs in the same segment share more constraints than they think: same tools, same obligations, same bottlenecks.
- How we test it
- We keep an anonymised register of signals. Reviewed every six months. Whatever comes up three times gets prototyped. We publish the aggregated patterns, never the cases.
Open questions
Making an agent hold inside a heterogeneous stack without migrating anything
The typical French SMB runs Qonto, Pennylane, HubSpot, Gmail, Notion and a trade ERP. Consumer solutions break at scale, enterprise solutions are oversized.
Measuring the return at 90 days without observation bias
Starting point: the chain task, time, usable capacity, reassignment, result, with leading indicators set before deployment.
Calibrating the human-in-the-loop without killing the return
Starting point: the thirty-days-at-95% rule. The open question is self-calibration of the threshold.
Comparing open and proprietary models on French SMB workflows
Test benches for accounting, client relations, administration. Results published.
Commitments
- Prices displayed when the offer is standardised.
- An explicit privacy policy: sub-processors named, hosting specified.
- We publish the aggregated patterns of what we learn, never the client cases.
Models and hosting
We pick the model per task, not on principle. Claude when reasoning matters, a smaller model when it is enough, an open model on your infrastructure when professional secrecy requires it. No workflow is wired to one provider. Hosted in Europe. No client data is used to train a model.
The full method, with its sources, is laid out on the home page, section How we work.