Arkange71% of companies use AI, but only 1% reach operational maturity (McKinsey State of AI 2025, 1,491 respondents, 101 countries). In between sits a poorly built diagnostic. Most leaders order an AI audit the way they order an accounting audit, to get a report. A report does not deploy an agent.
This article gives the playbook on the company side (what to prepare) and on the consultant side (what a good diagnostic must deliver), with a five-level AI maturity grid as the positioning tool.
Why do 9 out of 10 AI diagnostics end up in a drawer?
An AI diagnostic ends up in a closet when it confuses technology mapping with operational framing. 90% of French companies already use AI without a governance framework (Usine Digitale, 2025), which means the field is already saturated with attempts. The diagnostic does not arrive on a blank page. It arrives after ChatGPT in shadow IT, after two aborted POCs, after an executive committee that “validated an AI roadmap” without knowing what it was signing.
The standard failure looks like this. A firm delivers an 80-page report with a 2x2 matrix, ten use cases ranked by “impact vs. feasibility”, and a governance recommendation in quarterly AI committee style.
Six months later, zero agents in production. HBR named this gap the “Last Mile Problem” (HBR, March 2026). The distance between the roadmap and the agent that runs on Monday morning at the management controller’s desk is not crossed with a PowerPoint.
The recurring methodological trap is to start a diagnostic on a scope declared by the executive committee, discover at sprint 2 that the real friction is elsewhere, and have to reframe the scope mid-course. A wrong initial framing costs more in time than in money. The framing work must therefore precede the first diagnostic deliverable.
What is a company AI diagnostic?
An AI diagnostic is a 4 to 8-week exercise that delivers three things. A positioning on a maturity grid, a prioritized list of business frictions solvable by AI, and a costed roadmap for the first 3 to 6 use cases.
The difference with a classic data audit lies in the angle. A data audit looks at what exists (data quality, architecture, governance). An AI diagnostic looks at the target usage (which business processes will be transformed, by which agents, with what ROI at 100 days). 57% of French companies have been deploying generative AI use cases for over a year (Usine Digitale, 2025), so the “discovering AI” angle is outdated. The 2026 diagnostic starts by mapping what already exists in shadow AI, not by explaining what GPT is.
Three non-negotiable deliverables of a well-made AI diagnostic.
- A maturity score positioned on a 5-level matrix (see next section)
- A short list of 5 to 10 use cases with estimated gain in FTE, deployment cost and technical complexity
- A 100-day roadmap with a first agent identified, scoped and budgeted down to the sprint
If the deliverable stops at a “strategic roadmap”, it remains an audit.
An AI maturity grid of 5 levels to locate yourself
An AI maturity grid ranks an organization on five levels, from N1 Curious to N5 Scaler. It gives the leader a common language to locate their company against its peers, and cross-checks the Stanford segmentation (Brynjolfsson et al., Stanford Digital Economy Lab, April 2026, 51 deployments analyzed).
N1 Curious
Individual tests, ChatGPT in the browser, zero governance, zero measurement. The CFO uses Claude to proofread their notes, marketing tests Midjourney, nobody at leadership follows up, a pure shadow AI. 30% of French SMBs are here in May 2026 (estimate based on the BPI base).
N2 Experimenter
Validated POCs, 2 to 3 business use cases identified, a budding AI committee. At least one ideation workshop has taken place. An AI budget exists but it funds licences, not agents. No agent in stable production. This is the level of most French mid-caps that launched experiments in 2024-2025.
N3 Deployer
AI agents in production on 1 to 3 business processes, active usage measurement, first documented ROI. This is typically the level of organizations that industrialized content or structured note production through AI. The N2 to N3 transition is the hardest jump, and it is exactly what the “Last Mile Problem” describes.
N4 Industrializer
Agentic AI on critical processes, formalized governance, ROI measured and reported to the executive committee. Several agents interact. This is the level of mature industrial organizations that structured a cross-functional AI program on targeted scopes.
N5 Scaler
Systemic AI, internal center of excellence, the company produces its own agents without external help. Fewer than 1% of French SMBs and mid-caps (consistent with McKinsey 2025). This is the long-term target, not the 12-month target.
A leader sitting at N2 gains from aiming for a solid N3 within 6 months. The diagnostic exists to set that realistic ambition.
How to prepare for an AI diagnostic on the company side?
Preparing an AI diagnostic means gathering five elements before kick-off that save the consultant two weeks and the deployment three months. Preparation represents 20% of the total time and conditions 80% of the deliverable’s quality. a16z (Enterprise CIO Survey 2025, 100 CIOs) shows that companies with structured preparation deploy their first agent on average 2.3x faster.
The documented friction list
A list of 15 to 30 business frictions surfaced by the teams, not the executive committee’s assumptions. Minimal format. Process concerned, time lost per week, the person who suffers. This is the diagnostic’s fuel. Without this list, the consultant invents, and a consultant who invents produces a generic roadmap.
The shadow AI mapping
A 5-question internal survey on the AI tools already used without validation. Guarantee anonymity, otherwise the answers are false. The goal is not to sanction, it is to know what is already running. On diagnostics run in professional services firms, a well-run shadow AI survey regularly reveals 8 to 12 tools used by employees, none of which appeared in the initial roadmap.
Access to the key people
5 to 8 one-hour interviews with operational staff, not only with the executive committee. The executive committee describes the ambition. Operations describe reality. A diagnostic that interviews only the executive committee produces a roadmap disconnected from the field.
The documented data scope
One A4 page per critical data source, with volume, freshness, owner, technical accessibility. Not a full audit, just the essentials. If HR data sits in an Excel file shared on OneDrive, the diagnostic must know it before proposing a profile-matching agent.
The deployment budget
The classic trap is paying €30k for a diagnostic without having provisioned the €80k to €150k of the first deployment. A diagnostic only has value if the deployment follows.
What must a well-run AI diagnostic deliver?
A well-run AI diagnostic delivers four artifacts usable on the Monday after the readout. A maturity positioning, a costed short list of use cases, a roadmap on 10/30/100 milestones and a minimal governance plan. If any of the four is missing, the diagnostic is incomplete. Stanford (April 2026, 51 deployments) shows that deployments preceded by a structured diagnostic reach their ROI 40% faster.
Positioning on the maturity grid
An N1 to N5 score argued by dimension (usage, governance, data, skills, measurement). The score breaks down into five sub-scores. An SMB can be N3 on usage and N1 on governance, and that is the useful information.
A costed short list of use cases
5 to 10 use cases ranked by gain/effort ratio. Each case includes.
- A one-sentence description
- The impacted process
- Estimated gain in hours/week or euros
- Technical complexity out of 5
- Data dependencies
The roadmap on 10/30/100 milestones
A 10/30/100 rhythm structures the first 100 days. 10 days for the first agent in test, 30 days for team adoption, 100 days for measurable ROI. The diagnostic must identify this first agent precisely and budget the bootstrapping sprint. On a diagnostic in a professional services firm, this first agent is typically a structured document pre-review assistant or a synthesis note writer.
The minimal governance plan
Three pages maximum. Who validates new use cases, who measures ROI, how AI incidents are handled. An operational governance that fits in a monthly meeting. The model applies as well to a local authority as to an industrial SMB or a professional services firm.
Checklist. Are you ready for an AI diagnostic?
This extractable checklist lets an SMB or mid-cap leader know in 5 minutes whether their organization is ready for an AI diagnostic or whether the ground must be prepared first.
On the ambition side
- The executive committee has a quantified AI objective at 12 months
- An operational sponsor is named
- The post-diagnostic deployment budget is provisioned
On the field side
- A list of 15+ business frictions is available or collectable in 2 weeks
- The current shadow AI has been mapped (or can be under an anonymity agreement)
- 5 to 8 operational staff are available for one-hour interviews
On the data side
- The 3 to 5 critical data sources are identifiable
- An internal data owner exists (CIO, DPO or data steward)
- GDPR/AI Act compliance has been raised at least once at the executive committee
On the method side
- The diagnostic’s success criterion is defined before kick-off
- A post-diagnostic decision horizon is set (4 to 6 weeks maximum)
- The readout will be made at the executive committee
If three or more boxes are empty, plan 4 weeks of preparation before the diagnostic. If everything is checked, the diagnostic can start next week.
The thesis. An AI diagnostic is only useful when a deployment follows
The conviction is simple. An AI diagnostic delivered without a deployment commitment within the following 90 days is an expensive intellectual exercise. On diagnostics followed by a deployment within the quarter, the first agent is in production before day 100. The cases where the diagnostic stayed isolated produced no agent.
A diagnostic designed as the primer of the ADA cycle (Audit-Deployment-Adoption) naturally leads to the first agent development sprint. Positioning the organization on a maturity grid then starting an agent within the month remains the safest sequence.
