Arkange80% of AI projects never reach production according to Harvard Business Review (March 2026, “The Last Mile Problem Slowing AI Transformation”). The blocker has more to do with method than with budget. Everyone runs POCs, nobody operates them.
This article walks through the ADA cycle (Audit, Deployment, Adoption), a three-phase framework to turn a ChatGPT workshop into measured business AI agents. Three phases, dated deliverables, and an extractable checklist at the end.
Why do 80% of AI projects stop at the pilot?
80% of AI projects stop at the pilot because companies confuse “running an experiment” with “structuring a transformation”. The McKinsey State of AI 2025 study (1,491 respondents, 101 countries) quantifies the gap. 71% of companies use generative AI, but only 1% reach operational maturity.
The typical scenario looks like this. An executive committee decides that “AI has to happen”, an enthusiastic sponsor runs a ChatGPT workshop.
Three months later, a demo impresses the executive committee. Six months later, the project has vanished from the reporting. Nobody knows why, because nobody defined success criteria at the start.
The root cause holds in one number. 90% of French companies use AI without a governance framework (Usine Digitale, 2025). Without governance, the POC cannot move to production. The CIO blocks it for security reasons, the CHRO for adoption reasons, the CFO for ROI reasons.
What is the ADA cycle?
The ADA cycle is a three-phase AI transformation framework. Audit (diagnosing business frictions and data maturity), Deployment (short sprints building custom AI agents), Adoption (training, change management, usage measurement). Each phase has a target duration, named deliverables and a binary exit criterion.
The sequence counts as much as each phase. Skipping the Audit produces a technically perfect agent nobody uses. Skipping the Adoption produces an agent used for three weeks then abandoned.
The ADA cycle pairs with a 10/30/100 rhythm. 10 days for the first agent in test, 30 days for team adoption, 100 days for measurable ROI. ADA describes the what, 10/30/100 sets the when.
Phase 1. Audit. What to diagnose before writing a single line of code?
The Audit phase runs 2 to 4 weeks and produces three deliverables. The mapping of prioritized business frictions, the AI maturity score on a five-level grid (N1 Curious to N5 Scaler), and the shortlist of 3 to 5 high-ROI use cases. Not a single line of code is written before these three deliverables are signed off by the executive sponsor.
Map the frictions, not the use cases
The most frequent mistake is to start with “which AI use cases should we test?”. The right question is “which business processes cost our teams the most unvalued time?”. A business friction is a felt, measurable and costly problem. An AI use case is one possible answer among others.
A typical audit surfaces 10 to 15 frictions. Only three or four are good candidates for an AI agent. The rest belong to classic automation, process redesign or hiring. Without an audit, the AI budget goes to problems that never needed AI.
Position the company on an AI maturity grid
The AI maturity grid has five levels. N1 Curious (individual tests, shadow AI, zero governance), N2 Experimenter (validated POCs, a few business uses), N3 Deployer (agents in production on 1 to 3 processes), N4 Industrializer (agentic AI on critical processes, measured ROI), N5 Scaler (systemic AI, internal center of excellence).
The maturity level determines what comes next. An N1 organization needs an AI usage charter and baseline training before even thinking about agents, while an N3 organization must work on industrialization rather than multiply POCs.
37% of CIOs surveyed by a16z (Enterprise CIO Survey 2025) report running 5 or more AI models in parallel, a typical symptom of a failed N2 to N3 transition with no target architecture.
Phase 2. Deployment. How to build an AI agent in short sprints
The Deployment phase builds the AI agent in 1 to 2-week sprints, with real user testing starting from day 10, without a 60-page specification or a frozen spec. A trio holds the rhythm. One product owner, one business owner, one AI engineer, with short iterations and weekly demos to the sponsor.
Sprint zero. Architecture and guardrails
Sprint zero sets three things. The target LLM (and its fallback), the data architecture (RAG, fine-tuning or plain prompt engineering), and the business guardrails. The architecture gains from staying agnostic. The agent can switch from Mistral to Claude to GPT without a rebuild, which protects against price changes and vendor roadmap shifts.
Iterate with the users
Weekly iteration with 5 pilot users changes everything. A shared error log, the system prompt adjusted every 48 hours for 3 weeks.
Technical validation alone is not enough. An agent can give the right answer 95% of the time and still be rejected by the teams because it takes 12 seconds to answer when the operator expects 3.
The exit criterion of the Deployment phase
The agent moves to the Adoption phase when three criteria are validated. A correct-answer rate above 85% measured on a business test set, a response time compatible with the target usage, and three pilot users signing a written go. If one is missing, the sprint is extended. If all three are missing, it is back to the Audit, the initial framing being at cause.
Phase 3. Adoption. How to go from 5 pilot users to 200
The Adoption phase is where 80% of projects die. Three levers count.
- Training adapted to the user’s maturity level
- Weekly measurement of real usage (not declared usage)
- A structured feedback loop to the Deployment team for continuous adjustments
Train by level, not in one single session
A differentiated training plan changes the outcome. 2 hours for N1 users, 30 minutes for N3 users already familiar with ChatGPT, individual coaching for the 5% most reluctant.
A single training session would leave 40% of users by the roadside. The training can be funded through the OPCO when the provider is Qualiopi certified. On a typical mid-cap deployment, 40 to 60% of the training cost is covered when the plan is structured from the Audit onward.
Measure real usage, not declared usage
The gap between “I use the agent every day” declared in a survey and the actual usage log averages around 2.3. In other words, when a user declares using the agent 5 times a day, they actually use it 2 times. Without telemetry in place, the sponsor is flying blind.
Telemetry also enables scope decisions. If an agent built for 12 business use cases is actually used on 3 of them, the right move is to remove 9 cases from the scope, reinforce the 3 that are used, and redeploy the remaining budget on a second targeted agent. Without measurement, the agent would be abandoned as “moderately useful” when it is excellent on a reduced scope.
Day 100. ROI or pivot
Day 100 closes the first round of the cycle. Either the ROI is measurable and documented (time saved per user, lower error rate, higher client satisfaction), or the pivot decision is made. This checkpoint avoids the zombie-project drift that burns budget while producing nothing.
Extractable checklist. The ADA cycle in 12 points
A list to print and pin on the wall of the AI project meeting room.
Audit phase (weeks 1 to 4).
- Map 10 to 15 business frictions with a quantified annual cost
- Position the organization on an AI maturity grid (N1 to N5)
- Shortlist 3 to 5 use cases with ROI projected at 100 days
- Sign a named executive mandate for the sponsor and the business owner
Deployment phase (weeks 5 to 8).
- Choose a primary LLM and its fallback (agnostic architecture)
- Define the business guardrails and the data policy
- Ship a usable V1 at day 10 for 5 pilot users
- Iterate weekly with a shared error log
Adoption phase (weeks 9 to 16).
- A training plan differentiated by user maturity level
- Real-usage telemetry installed before general rollout
- A structured feedback loop to the Deployment team
- A 100-day ROI review with a written decision. Industrialize, adjust or pivot
FAQ. Common questions about the ADA cycle
Does the ADA cycle fit every company size?
The ADA cycle is calibrated for SMBs and mid-caps of 50 to 500 employees. For smaller structures, the Audit phase is compressed to one week. For large groups, ADA applies per business unit or per business function, not at corporate level.
How much does a full ADA cycle cost?
The cost varies with the scope. An indicative range for a first AI agent in an SMB or mid-cap runs between €30,000 and €80,000 over 100 days, training included and partially fundable through the OPCO.
Can you run ADA in-house without a firm?
Yes, provided you have three profiles. A product owner experienced in AI, an operational AI engineer, and a business owner available 30% of their time for 4 months. Most SMBs and mid-caps do not have all three profiles in-house, which explains the market for AI consulting.
What happens if ROI is not measurable at 100 days?
A written decision is mandatory. Industrialize the scope that works, adjust the scope that underperforms, or pivot to another use case. The absence of a decision at day 100 is the first symptom of the zombie project.
How many AI agents can you deploy in parallel?
Over the first year, 1 agent suits N1-N2 organizations, up to 3 agents in parallel for N3, and 5 to 10 agents for N4. Beyond that, the limiting factor is no longer technology but the business teams’ capacity to adopt.
The first step
A diagnostic that produces a maturity positioning, a shortlist of 3 priority use cases and a budget and timeline range gives the project its starting point. An AI transformation gains from starting with an honest audit rather than a POC.
