Many companies tested AI in 2024-2026 via ChatGPT or isolated tools, often yielding limited scalable gains and growing concerns about data and costs. This post outlines a pragmatic 30-60-90 day roadmap to transform experiments into operational, measured, and secure capabilities.
Many companies have already “tested AI” in 2024-2026, often via ChatGPT or isolated SaaS tools. The result is frequently the same: a few individual gains, few measurable gains at scale, and growing concern regarding data, compliance, costs, and quality.
An enterprise AI plan serves precisely to avoid this scenario. It is a short, results-oriented roadmap that transforms trials into operational capabilities (integrated into workflows, measured, secured). Here is a pragmatic version in 30-60-90 days adapted for SMEs, scale-ups, and growing companies.
An AI plan is not “a list of tools to deploy.” It is a mini-product program that aligns on:
The why (priority business objective, quantified)
The what (maximum 2 use cases to start, chosen for their frequency and measurability)
The how (data, integration, UX, governance, run)
The proof (before/after KPIs, guardrails, go/no-go decision)
The right AI plan has a key property: it produces a useful V1 in less than 90 days, without creating unmanageable technical or legal debt.
You can start without overhauling your IS, but you need a minimal foundation.
A sponsor (executive or head of) prioritizes, frees up time, and decides on trade-offs (quality vs speed, confidentiality vs SaaS, etc.). Without arbitration, pilots die in discussion loops.
Before any experimentation, formalize a color code (e.g., public, internal, sensitive) and what is authorized in each tool. For France, the CNIL publishes useful benchmarks on AI and data protection: CNIL references on artificial intelligence.
A pilot without a baseline proves nothing. Measure at a minimum: volume, time, quality, cost. Even if it is approximate, do it beforehand.
AI is not solely an IT subject. Each use case must have a business owner responsible for defining scenarios, validating quality, and driving adoption.
Your AI plan must iterate quickly (ideally weekly). The longer your loops, the more you “over-specify” and the less you learn.
The goal is not to “do a lot.” It is to do a little, but in controlled production, with a clear decision at Day 90.
Period | Objective | Concrete Deliverables | Expected Decision |
|---|---|---|---|
D1 to D30 | Frame and secure | 1-2 use cases, KPI baseline, data and integration mapping, guardrails, instrumented prototype | “What are we piloting, with what KPIs and limits?” |
D31 to D60 | Pilot in real conditions | MVP integrated into workflow, evaluation protocol, logging, point-of-use training, feedback loop | “Does it work on real cases, at acceptable cost and risk?” |
D61 to D90 | Move to controlled production | runbook, monitoring, cost control, incident management, adoption plan, V2 backlog | “Do we industrialize, iterate, or stop?” |
A good starting duo:
Foundation: frequent, transversal, low risk (e.g., internal knowledge assistant on procedures, offers, FAQs, operational documents)
Showcase: visible, rapid impact (e.g., support triage, lead qualification, data extraction from documents)
The most predictive criterion is not “the most impressive AI,” it is frequency. A gain of 30 seconds on 2,000 monthly occurrences beats a gain of 30 minutes on 5 occurrences.
A minimal set per use case:
North Star (impact): processing time, resolution rate, conversion rate, response time, etc.
Process KPIs: usage rate in the workflow, completion rate, volume processed
Guardrails (risks): human escalation rate, critical error rate, potential data leak, unit cost
At this stage, the goal is not perfection, but making the impact manageable.
The classic trap: building “a chatbot on the side” that has access to nothing and can do nothing.
At Day 30, you must document:
Sources of truth (CRM, helpdesk, drive, ERP, knowledge base)
Necessary integration points (read, write, actions)
Constraints (PII, client data, secrets, access rights)
In 2026, the right reflex is to apply a “risk-based” logic consistent with reference frameworks (e.g., NIST AI RMF and the European approach to AI). For the EU context, you can consult the European policy on artificial intelligence.
Concretely, at Day 30 you want a simple document:
What the assistant is allowed to do (and what it is not allowed to do)
When it must escalate to a human
Which data is forbidden as input
What is logged (for audit, debug, improvement)
The Day 30 prototype must already:
Run on real examples (scenarios)
Produce usable logs
Make costs visible (even if approximate)
A demo without instrumentation sets you back 30 days.
This is the phase where you move from “it seems to work” to “it works within our constraints.”
A good pilot is placed in the daily tool, not in yet another tab. Examples:
In a helpdesk (triage, assisted response, article suggestion)
In a CRM (call summaries, pre-filled fields, follow-ups)
In an intranet (document search with sources)
At Day 60, you must be able to answer with facts:
What proportion of responses is acceptable without editing?
What types of requests fail?
What is the “critical” error rate (legal, factual, security)?
Even a basic protocol on a stable sample (e.g., 100 cases) is already a huge advantage.
AI is managed like a product. The pilot must integrate:
A user feedback channel
A weekly review (30 minutes): top errors, top gains, decisions
A version discipline (prompts, rules, sources)
Training “everyone” in a room for a day rarely works.
What works better: training teams on their use case, with:
5 typical scenarios
5 anti-patterns (what not to do)
1 rule on data
From Day 60 onwards, the question is no longer “is it possible?” but “is it operable continuously?”.
At Day 90, you want simple operations:
Who is on-call (even if light)?
What do we do when quality drops?
What do we do when a connector fails?
What do we do when costs rise?
Many companies discover too late that the cost is not just “the model price,” but also: integration, monitoring, knowledge maintenance, internal support.
The right reflex at Day 90: track a cost per useful action (e.g., cost per ticket resolved, cost per document extracted, cost per qualified lead) and not just a global monthly cost.
You don’t need a monthly committee of 12 people. You need a short ritual, with a decision.
Example of pragmatic governance:
Ritual | Participants | Duration | Goal |
|---|---|---|---|
Weekly pilot review | business owner, tech lead, ops | 30 min | Correct, prioritize, decide |
Day 90 Gate | sponsor, owner, tech, compliance if needed | 45 min | Scale, iterate, or stop |
Here is a simple scorecard that avoids endless debates.
Dimension | Question | Decision Threshold (examples) |
|---|---|---|
Value | Is the gain measured vs baseline? | Yes, significant and stable |
Quality | Are critical errors rare and detected? | Yes, with escalation |
Integration | Is the workflow fluid, without friction? | Yes, natural usage |
Cost | Is the unit cost acceptable? | Yes, controlled |
Risk | Data, compliance, security under control? | Yes, documented |
The idea is not to have 100/100, but to avoid a “go” on an uncontrolled solution.
“The teams are using it” is not a business KPI. Measure time saved, quality, resolution rate, incremental revenue.
An “assistant for the whole company” right from the start often leads to a vague system that is impossible to evaluate. Start with a tight scope.
An AI that doesn’t read your reliable sources and cannot trigger actions remains a gadget. Integration is often the real barrier.
Data rules and human escalation must exist before production, not after a leak or an error.
What is the best first use case for an enterprise AI plan? The best first case is generally frequent, measurable, and low risk, for example, an internal knowledge assistant with verified sources, or support triage with human escalation.
Can you do a 30-60-90 day AI plan without a data team? Yes, if you start with use cases that rely on your existing tools and light integration. However, you will need a business owner, a technical lead, and KPI discipline.
Which KPIs should be chosen to prove ROI quickly? Take 1 North Star KPI (impact) and 2 to 4 support and risk KPIs. Example for customer support: first response time, resolution rate, escalation rate, critical error rate, cost per ticket.
Should we buy a tool or build custom? This depends on your integration, data, and governance constraints. A simple rule: if the AI needs to integrate finely with your processes and sources, custom (or a hybrid approach) often becomes more robust in the medium term.
How to remain compliant with AI in 2026? Adopt a risk-based approach: data classification, usage rules, traceability (logs), continuous evaluation, and guardrails. Align with your GDPR obligations and governance compatible with European requirements.
If you want an enterprise AI plan that leads to a measurable V1 in 90 days (and not a graveyard of POCs), Impulse Lab can assist you on all or part of the journey:
AI Opportunity Audit to prioritize 1 to 2 use cases with rapid ROI (see our guide on strategic AI audit).
Development and integration of custom web and AI solutions, connected to your tools.
Process automation and controlled production deployment.
Adoption training oriented towards real usage and data rules.
Discover the agency at Impulse Lab and let’s discuss your 30-60-90 day roadmap.
Notre équipe d'experts vous répond rapidement pour comprendre vos besoins et vous proposer la meilleure solution.
Vous avez des questions ? On a les réponses.
Leonard
Co-fondateur
Continuez votre lecture avec ces articles
Vous avez probablement déjà vécu la scène : quelqu’un poste un prompt “miracle” sur Slack, un autre teste un outil gratuit “juste pour voir”, puis, deux semaines plus tard, vous ne savez plus qui utilise quoi, avec quelles données, ni si les résultats sont fiables.
Un prototype d’agent IA peut impressionner en 48 heures, puis se révéler inutilisable dès qu’il touche des données réelles, des utilisateurs pressés, ou des outils métiers imparfaits. En PME, le passage à la production n’est pas une question de “meilleur modèle”, c’est une question de **cadrage, d’i...