AI Deployment in SMEs: The 90-Day Plan for Executives

Can AI be deployed in an SME within 90 days?

Yes, 90 days is enough for a pilot or a well-defined use case. It is less certain once the project involves complex integrations across several tools. One source indicates that with a well-defined scope, a first measurable result is realistic within 30 days and a production deployment within 90 days [27].

The announced timelines vary according to the nature of the project:

Type of project	Announced timeline
Simple AI agent, once the scope is defined	2 to 4 weeks [19]
Structured deployment, from diagnosis to skills transfer	4 to 6 weeks [37]
AI agent, implementation guide	6 weeks [36]
First structured deployment (audit, stack selection, training of champions, pilot, measurement, iteration)	approximately 8 weeks [18]
Integration across several tools (CRM, web forms, invoicing, dashboards)	often 6 to 12 weeks [33]

The scopes described do not match exactly, which makes the comparison approximate. The gap between the 4 to 6 weeks in [37] and the 8 weeks in [18] is one example.

Timelines to read with caution

Almost all of these sources are service providers. Their figures are practical recommendations drawn from their own activity, not results validated by an independent study. It is better to read them as orders of magnitude rather than delivery commitments.

What this means for a business leader

A 90-day plan is consistent if the ambition is a pilot or a targeted use case. If the project involves connecting several tools, the 6 to 12 weeks announced for this type of integration [33] may consume most of the time. In that case, it is better to start with a simpler use case. A simple agent is announced in 2 to 4 weeks, provided the scope is defined [19]. The integration work can then be planned separately.

What are the main phases of a 90-day AI plan?

A 90-day AI plan for an SME breaks down into three blocks of about 30 days each: frame, pilot, then measure and extend. One guide puts it this way: 30 days to frame (audit, use cases, rules), then the pilot, then the measurement phase [40]. A checklist follows the same rhythm: 6 prerequisites, 3 phases of 30 days, one checkbox per action and one deliverable per week [26].

Block	Period	Content
Frame	Days 1 to 30	Audit, choice of use cases, usage rules [40][26]
Pilot	Days 30 to 60	Test on a restricted scope [40][26]
Measure and extend	Days 60 to 90	Measurement of results, then expansion [40][26]

Detailed variant for a general-purpose tool

For adopting a general-purpose tool such as ChatGPT, one guide proposes four phases of two weeks to one month [24]:

• Days 0 to 15: audit and choice of tool.
• Days 15 to 30: pilot with 3 people.
• Days 30 to 60: expansion to 10 to 15 employees.
• Days 60 to 90: generalization across 2 to 3 key processes [24].

This variant reorganizes the three blocks: the pilot starts from day 15, and the expansion phase is inserted before generalization.

What the sources allow us to assert

The sources give few day-by-day details. The outline presented here is a reconstruction from their common points, not a validated schedule.

What recurs across sources is the 90-day framework itself. It appears as an action plan for SMEs and mid-sized companies [14], a concrete roadmap [15], or the "30-60-90" method [35]. A short roadmap, one to two pages, generally names a first pilot with its success indicators and sets a 90-day calendar [29].

These contents come mostly from service providers and consulting bodies. Their durations are practitioner recommendations, not measured results.

What should be done during the first 30 days to frame the project?

The first month serves to draft a short roadmap and to choose a single first pilot, before any tool purchase or mass training.

A roadmap of 1 to 2 pages

The roadmap is a one- to two-page document that brings together four elements [29]:

• Prioritized use cases, ranked by effort, estimated ROI and regulatory risk [29].
• A first pilot, explicitly designated [29].
• The pilot's success indicators [29].
• A 90-day calendar [29].

One consulting source also recommends identifying three quick wins and planning how to measure ROI, including without an IT department [15].

Audit, stack and indicators

One provider places the audit and the choice of technical stack at the head of its structured deployment, before training champions, the pilot, measurement and iteration [18]. Indicators should be set from the start, not after the pilot. Another commercial source cites four: time saved, error rate, user satisfaction and quality of outputs [34]. Without a measured starting point, the day-90 review cannot be verified.

Start small: France Num's approach

France Num recommends progressive adoption: first general consumer tools, then connecting AI to existing tools, and a dedicated project only if the need justifies it [4]. Framing therefore means choosing the level of ambition that fits, not aiming straight for the heaviest option. Oupi, which is deployed team by team with integrated training and follow-up, fits a first pilot limited to a single team (learn more).

A checklist to keep pace

One publisher offers a 90-day checklist with 6 prerequisites, 3 phases of 30 days, one checkbox per action and one deliverable per week [26]. It comes from a commercial player and targets the deployment of a specific tool (Claude) [26]. Its format is transferable: each week of the first month ends with a written deliverable, for example an audit, a list of use cases, a validated roadmap, then a framed pilot. This breakdown is not the source's own, which does not detail the content of the weeks.

How to choose and run the pilot (days 30 to 60)?

A good pilot targets a precise use case, involves a volunteer team and relies on indicators set before launch [34]. Between day 30 and day 60, the goal is to prove value on a reduced scope, not to deploy everywhere.

The four criteria of a successful pilot

One consulting source identifies the following criteria [34]:

• A precise case rather than a broad theme: "automating sales follow-up email drafting" rather than "AI in marketing."
• A team of 5 to 10 people, volunteers, with a referent.
• A bounded duration: the source mentions 3 months. In a 90-day plan, this requires a narrow scope. Another source estimates a first structured deployment, from stack selection to iteration including the pilot, at about 8 weeks [18].
• Indicators defined from the start: time saved, error rate, user satisfaction, quality of outputs [34].

Team size varies across methods. A ChatGPT adoption plan starts with a pilot of 3 people, then expands to 10 to 15 employees between days 30 and 60 [24].

Integrating AI into processes

The pilot must fit into real work. A training programme for small businesses structures this step in four stages [22]:

1. break down the tasks;
2. decide the level of human control;
3. test on real cases;
4. document the operating procedure.

This operating procedure then serves as the basis for generalization.

An example use case: chasing unpaid invoices

One provider cites chasing unpaid invoices with an AI agent. Each morning, the agent queries the invoicing software and identifies invoices overdue by 7 to 10 days. It analyzes the customer profile (history, amount, sector), then drafts a personalized reminder email [36]. The case is well bounded and measurable: payment delay, number of reminders handled, time spent.

The same provider presents it as "statistically the fastest-ROI workflow" for a services SME [36]. This is a commercial claim that the file does not allow us to verify: it should be validated against your own figures.

With Oupi, deployment happens team by team, with integrated training and follow-up, which matches the logic of a pilot limited to a group of volunteers (enterprise deployment).

How to measure results and generalize (days 60 to 90)?

The final phase consists of measuring the pilot's results, iterating, then gradually extending the system while setting the first indicators [20][18]. The sources do not agree on the calendar. One training content places this step between days 75 and 90 [20], and another includes "measurement and iteration" in a first structured deployment of about 8 weeks [18]. Both sources are commercial players. The logic stays the same: generalize only after verifying.

The three actions of the final phase

• Measure and adjust. Review the pilot's results and correct before extending. A book on financial automation puts this logic as "adjust the parameters before general deployment" [8].
• Move to production. This is the final step of the path from pilot to production [38].
• Extend in stages. The system expands progressively, not all at once [20].

Avoiding loss of momentum

The main risk is neglect. A book on AI in business says it plainly: "at 90 days, the topic has dropped off the agenda," and the budget is raised right afterwards [5]. Day 90 should therefore not be a finish line.

Setting milestones after day 90

Set follow-up meetings for what comes next now. One piloting example evaluates teams every 90 days against a grid [12]. A quarterly review of indicators, with a designated owner, builds on the same idea. For team autonomy, one source proposes field support of 3 to 6 months after a one-day business training [38].

Oupi is deployed team by team, with integrated training and follow-up, which corresponds to this logic of progressive extension. Its offers for companies detail how it works.

How to bring teams on board and manage change?

Adopting AI depends less on the tool than on the training and internal champions who keep it alive. The sources converge on this point, even though many come from service providers and training organizations, whose recommendations are not empirically validated.

Start with people, not technology

AI literacy does not start with a large technology plan [9]. The need is documented: a diagnostic of very small and small businesses in Wallonia identifies training needs in generative AI [11].

Internal champions and the framework of use

A 90-day plan relies on identified AI champions and prioritized use cases, according to a training organization [25]. Training these champions is one of the steps of a first structured deployment, which a provider estimates at about 8 weeks [18].

Training works best when based on real work situations. A public programme in Île-de-France devotes 7 hours to this, for example: task breakdown, decisions on human control, tests on real cases and documentation of the operating procedure [22]. This work sets the framework of use: what the team entrusts to AI and what it controls.

Supporting employees through the transition

The LaborIA guide states that training plans should support employees' professional transition [21]. This guide, published by the programme bringing together the Ministry of Labour and Inria, is aimed at managers of very small and small businesses [2].

Aiming for autonomy

One provider offers a one-day business training, followed by 3 to 6 months of field support, to get a team able to maintain and develop its solutions [38]. This duration goes beyond 90 days: the plan launches the momentum, and autonomy is built afterwards.

Where Oupi fits

Oupi is deployed team by team, with integrated training and follow-up. This logic matches the breakdown above: we move forward team by team, rather than opening the tool to the whole company at once. The training offers detail this aspect.

What budget to plan and should you seek support?

A pilot budget of €5,000 to €15,000, licences and support included, is put forward by a training and consulting provider [34]. This figure is an indicative benchmark, not an established market price. It mainly serves to calibrate the order of magnitude of a bounded pilot, not to price a full deployment.

Should you get support?

An SME can deploy AI solutions integrated into its tools without hiring, provided it relies on a partner who delivers the solution and trains the teams [28]. This claim comes from a player who sells precisely this kind of service: it describes a possible option, not a demonstrated necessity.

A book also mentions a budget for a serious first deployment in an SME or mid-sized company [5]. The available extract is truncated: the amount is not usable and we do not reproduce it.

Before signing, three questions help assess a partner's offer:

• Does the delivered solution remain your property [28]?
• Are the teams trained to run it without depending on the provider [28]?
• Are success indicators (time saved, error rate, user satisfaction) set from the start [34]?

Public resources to consult before buying

Two public documents help frame the need before any spending:

• The LaborIA guide: produced by the programme bringing together the Ministry of Labour and Inria, it is aimed at managers of very small and small businesses planning to integrate AI. It is available on France Num [2].
• The Bpifrance Conseil and Siparex methodology: it offers a structured deployment approach and a list of 20 real use cases [3].

Reading them first helps formulate a precise specification, and therefore compare quotes on a common basis.

Where Oupi fits

Oupi is deployed team by team, with integrated training and follow-up. This organization concerns companies that want to bring together tool, skills building and follow-up in a single approach: see the Enterprise page.

What pitfalls and risks should be avoided (compliance, scaling up)?

The main pitfall is remaining at the level of individual use: each employee uses a tool on their own, without the company gaining structured benefit. According to the Bpifrance barometer, cited by a training provider, 32% of French SMEs use generative AI, but only 11% have deployed it across the company [25].

Figures that depend on the definition

Adoption statistics are not easy to compare. Another commercial player claims that 48% of French SMEs use at least one AI tool [23]. A study on Belgium indicates that 16% of SMEs there have integrated AI [11]. Depending on whether one counts occasional use, a tool, or an integration, the gap is considerable. It is therefore better to set your own success criterion than to aim for a market average.

Compliance: GDPR and the AI Act

A French SME is subject to the GDPR and to the European AI Act [38]. This source speaks of application "soon," and the file contains no usable date. Check the calendar and obligations in the official texts. For generative AI, the CNIL publishes questions and answers on the use of generative AI systems, cited by [37].

Scaling up: what to avoid

Scaling up requires rigorous planning on the technical, organizational and human levels [30]. The most frequent pitfalls:

• Sticking to individual uses, without prioritized use cases or shared indicators.
• Handling compliance after the fact, when it should shape the choice of tools.
• Neglecting the human side: without training or referents, the tool stays underused.
• Taking a sales promise as proof. One provider announces for example a "guaranteed measurable ROI within 6 months" [35]. This is a sales argument, not a study. Demand the indicators, the measurement method and comparable cases.

To move beyond individual use, Oupi is deployed team by team, with integrated training and follow-up (see the enterprise offer).