IAkhathon. How to ship 4 AI agents in 48 hours (and what must hold for 100 days)

Let us be clear. 95% of generative AI projects in companies produce no measurable ROI (MIT Media Lab, NANDA, August 2025). Not 50%. Not 70%. Ninety-five. Meanwhile, strategy committees keep slicing AI adoption into 18-month quarterly roadmaps. This article dissects the Arkange IAkhathon format. 48 hours, four business agents shipped, the structural decisions that hold, the recurring traps that do not, and the 100-day review with no corporate speak.

What is an Arkange IAkhathon and why does this format exist?

An Arkange IAkhathon is an intensive 1 to 2-day format, Qualiopi certified, that turns business frictions listed in the morning into AI agents tested in production by the next evening. The format exists because the “eternal pilot” has become the dominant anti-pattern. According to the McKinsey State of AI 2025 report (1,491 respondents, 101 countries), 71% of companies use generative AI, but only 1% reach operational maturity.

The problem is no longer access to technology. The problem is translating it into real usage. An IAkhathon crushes the decision cycle. No steering committee, no 40-page specification, no POC sleeping six months on an Azure server. You audit, you build, you test, you document. In 48 hours.

The typical context before the intervention

The IAkhathon finds its relevance in organizations that share the same pattern. Dozens or even hundreds of ChatGPT or Copilot licenses distributed without a usage framework. One or two POCs launched with a vendor, stopped after a few months for lack of a clear business case. A few half-days of “AI awareness” by a generalist consulting firm, with no operational deliverable. An Arkange AI maturity matrix score often sitting at N1 Curious. Widespread shadow AI, zero governance, zero measured business usage.

The mandate Arkange typically receives is simple, formulated by a leader. “I do not want slides anymore. I want my teams to save time on Monday morning.” According to the CSOEC 2025 barometer, 67% of French accounting firms declare using generative AI, but only 9% have deployed an industrialized business use case. The finding generalizes to most French SMBs and mid-caps.

Decision 1. Frame 4 use cases, not 12

The first structural decision happens at the framing workshop on day -7. Select exactly four use cases, not one more. This discipline is counter-intuitive because every participant wants their agent. A typical friction list in a mid-sized organization surfaces 15 to 25 candidates. Four are kept, chosen on three criteria from Arkange’s 10/30/100 method. Weekly usage frequency above 5, estimated time saving above 30 minutes per occurrence, and input data already available in an exploitable format.

Why four and not ten? Because according to internal Arkange data consolidated across the IAkhathons run since 2023, beyond five simultaneous agents the share of agents actually used at day 30 drops drastically. Cognitive saturation of the field teams is the first adoption failure factor, ahead of the technical quality of the agents.

The agent families that most often emerge in an IAkhathon are stable. A note or synthesis writing assistant, an incoming email pre-qualification agent, a meeting minutes agent built from transcripts, and a consistency review agent on structured documents (trial balances, quotes, files). These four patterns cover most of the needs observed in SMBs and mid-caps.

Decision 2. Bring executive sponsors in as co-prompters, not spectators

This is where the classic political trap sits. Many organizations are tempted to reserve the IAkhathon for junior staff and engagement managers, because they are the ones who will use the agents daily. Apparent logic, strategic trap. The partners or decision makers who were not brought into the design see tools they never validated land on their screens on Monday morning.

The natural reflex of a sponsor who was not brought in is to refuse to sign off the output of an agent whose logic they do not understand. The most technically advanced agent can be deactivated by its own sponsor in week 2 on that ground alone. The rule now applied to every IAkhathon. No executive sponsor signs the production go-live of an agent they have not themselves tested and modified at least once in the system prompt.

Concretely, this means a 2 to 3-hour workshop with the sponsors during the IAkhathon, where they write the tolerance thresholds and the business rules themselves. It is not comfortable. It is the condition for the agent to survive beyond day 14.

Decision 3. Multi-LLM agnostic architecture from sprint 1

The third decision is technical but heavy with consequences. The natural temptation in an organization already subscribed to ChatGPT Team is to build everything in custom GPTs. That temptation must be refused. According to the a16z Enterprise CIO Survey 2025 (sample of 100 CIOs), 37% of companies already run 5 or more different AI models in production. Lock-in to a single vendor has become a major operational risk.

On the Arkange platform, every agent is built in agnostic mode. Claude for synthesis note writing, GPT for email sorting, Mistral for meeting minutes when data residency requires it, fine-tuned local models for sensitive reviews. The immediate benefit. When a vendor ships an update that degrades quality, you switch models in under two hours, without touching the rest of the architecture, without a single user noticing.

“You do not buy an AI vendor. You buy the ability to change vendors without pain.” That is the sentence that comes back most often in IAkhathon debriefs, and it is the right reflex.

Decision 4. Frame confidentiality from sprint 1, not after the incident

The most expensive mistake in an IAkhathon is not technical. It is deploying an email pre-qualification agent on a standard mailbox with no upstream filter on messages containing sensitive data. A non-negligible share of an organization’s incoming emails contains SIREN numbers, confidential financial amounts, or attachments covered by professional secrecy (medical, accounting, legal, HR).

The typical warning signal. A reply draft explicitly mentioning the context of another file, generated by classic context hallucination linked to accumulated contexts. The draft is never sent in well-framed deployments, but the mere fact that it can be generated is a red flag.

The systematic countermeasure applied in an Arkange IAkhathon. Adding an upstream pre-filtering agent that classifies sensitivity, refactoring the prompt to forbid any cross-reference between files, and setting up a mandatory human review on 100% of replies for 30 days before returning to sampling. Operational cost. 2 to 3 extra person-days per affected agent. Cost avoided. A potential CNIL report and a professional warning. AI Act compliance is not recovered post-incident.

Decision 5. Measure usage from day 1, not day 90

The fifth decision is to connect a usage dashboard from the first day of production, measuring invocation count per agent, per employee, per day, estimated time saved (declared plus session time), completion rate of generated replies, and post-generation modification rate. Too many organizations wait for a steering committee at day 90 to discover that three of five agents have not been used since week 3.

On a consolidated base of Arkange IAkhathons at day 100, typical orders of magnitude on four well-framed agents run around several hundred invocations per agent, validation rates without major modification of 65 to 80% on the most mature agents, and an aggregated saving between 250 and 400 employee-hours over the first quarter. On the usual cost profiles, the IAkhathon (training + platform + 100-day follow-up) typically pays for itself in under 60 working days.

What we would do differently. Three method adjustments

The honest question to ask about any AI deployment is the counterfactual one. Three method adjustments have become standard through iteration.

First, bring every executive sponsor in from the day -7 framing workshop, not as spectators but as co-prompters on at least one agent. The political cost of their initial absence is always higher than the operational cost of their presence.

Second, frame confidentiality from sprint 1, with a data sensitivity matrix per use case. Any agent that touches emails, legal documents or HR data must be deployed in mandatory human draft mode from the start. The European AI Act, in full application since 2026, no longer forgives approximation on limited-risk uses in regulated professions.

Third, keep agents with professional liability stakes out of a 48-hour IAkhathon. An agent touching the core business (accounting trial balance review, aeronautical quality control, HR application scoring) requires a dedicated 3-week sprint in a full ADA cycle. Not everything is hackathon-able, and recognizing the limit of the format is what makes it credible on everything else.

Replicable patterns vs sector specificities

Not everything transfers from one IAkhathon to another, but the structural patterns replicate systematically. The 4 use cases maximum rule, the agnostic multi-LLM architecture, the usage dashboard from day 1, and the mandatory onboarding of the executive sponsor as co-prompter produce an agents-still-active-at-day-100 rate above 85% on the consolidated base of Arkange IAkhathons.

What varies is the grammar of confidentiality. In an accounting firm, it is the professional secrecy of article 226-13 of the French Penal Code. In an industrial or aeronautical environment, it is ITAR constraints and industrial secrecy. In the training and education sector, it is learner GDPR. In the public sector, it is compliance with the DINUM frameworks. The method stays, the risk matrix gets rewritten.

Extractable checklist. 7 rules before launching an IAkhathon

Here is the consolidated checklist any leader can apply before launching their own intensive AI deployment format.

• Cap at 4 use cases maximum, selected on frequency + time saving + available data
• Bring the executive sponsor in as co-prompter, not as spectator of the final demo
• Build on an agnostic multi-LLM architecture from sprint 1, never single-vendor
• Frame the confidentiality and data sensitivity matrix before the first line of prompt
• Connect a usage dashboard at day 1, not day 90. Measurement creates adoption
• Reserve agents with professional liability stakes for a full 3-week ADA cycle, outside the IAkhathon
• Measure ROI in employee hours saved and incidents avoided, not in the number of agents deployed

An IAkhathon is not a technology demonstration. It is a protocol to move an organization from N1 Curious to N3 Deployer on the Arkange AI maturity matrix, in 48 hours of focused work and 100 days of disciplined measurement. The rest, the slide about “accelerating digital transformation”, belongs to the previous cycle.