In short
- Most roadmaps fail because they stack proofs of concept with no go-live criterion and no named owner.
- The right compass is value delivered, not technical execution speed. A quantified gain on one specific business flow triggers a purchase; an abstract promise triggers nothing.
- Twelve months are enough to deploy agents across the whole installed base, generate first revenue and consolidate the model at scale, provided you sequence it into four watertight quarters.
- The ICE framework lets you choose between ten use cases in a week, on three criteria a board can read.
- AI is monetised in credits, not in tokens. One credit represents one successful action, billed at around 10 % of the value saved at the end client.
- AI Act compliance is thought about from month 1, not retrofitted once the system is deployed.
- Outsource then internalise is the most common sequencing: first results in three months, autonomy built progressively.
Why most AI roadmaps fail before production
Ideas are not what is missing in this market. Sequencing is. Gartner predicted that at least 30 % of generative AI projects would be abandoned after proof of concept by the end of 20251. The cause is not technical: it lies in roadmaps that stack experiments with no go-live criterion, no dated milestone, no named owner.
The scenario is almost always the same. Feasibility is validated, the prototype stays on the shelf, the team moves to the next subject. The plan advances on paper, never in the product, and even less as far as the end client.
A roadmap that works does the opposite. Every stage produces a usable deliverable, names an owner, sets the date of the next stage and defines what “done” concretely means. For a software vendor, “done” means a client is using the system and you are able to quantify what they gain from it.
The real starting point: how fast value is delivered
The question to ask is not “which technology do we choose” but “how fast does this stage bring a measurable gain to an end user”. That is what separates a roadmap you can defend in front of a board from a catalogue of good intentions.
A quantified gain on one specific business flow, of the kind “invoice processing time goes from six minutes to forty seconds”, triggers a buying decision. An abstract promise, of the kind “we are integrating AI into the product”, triggers nothing, because there is nothing to compare it with.
The 12-month AI roadmap, phase by phase
Each phase produces a result that conditions the next. The first puts agents into production, the second generates first revenue, the third consolidates at scale, the fourth transfers autonomy.
Months 1-3: scope, then go live
The first quarter is for understanding the real work, not for choosing tools. You map use cases from the friction observed at users, you score them, you keep one or two, and you ship them. By the end of the third month, a system is running at pilot clients, with its decision log and its escalation threshold.
This is also when the regulatory classification is done, and when per-client measurement is put in place. Both are expensive to graft on later.
Months 4-6: sell, and build what differentiates
The first credits are sold. The business model produces its first euros of recurring revenue, and real usage lets you adjust the price against concrete data rather than meeting-room assumptions.
In parallel, the team builds the first two proprietary business features, the very high value ones, specific to what the software does. Those are what create durable differentiation: a competitor does not reproduce them by wiring a generic language model into their interface.
Months 7-9: industrialise and monitor
The scope widens, from pilot clients to the whole installed base. Monitoring becomes central and covers three dimensions: agent reliability, credit consumption, and results delivered to end clients. The vendor then holds a complete sales kit: pitch, documented use cases, real impact metrics.
Months 10-12: consolidate the model and build autonomy
Pricing is adjusted against real consumption data: tiers, per-user quotas, offers by segment. Internal teams are trained to maintain the agents and read the metrics. By the end of the twelfth month, the vendor has an active AI revenue line, agents deployed across the whole base, and the ability to evolve the system without total external dependency.
Prioritising use cases with the ICE framework
Mapping ten use cases is a good start, but not enough: you need a method to choose.
| Dimension | Question asked | Scoring |
|---|---|---|
| Impact | What business value if the use case succeeds? | 1 to 10 |
| Confidence | What probability of technical success and adoption? | 1 to 10 |
| Ease | What implementation effort? | 1 to 10 |
The final score is the average of the three. A high-impact use case that is hard to ship loses points in the ranking. That simplicity is an asset: prioritisation becomes readable for a board, with no technical skill needed to interpret the result.
In a week, every identified use case is scored. One or two winners emerge and become the commercial triggers of the roadmap. For the heaviest projects, the “Ease” score draws on five dimensions: data availability, integration capability, technical feasibility, tractability of the problem, and scaling.
The business model: turning AI into a revenue line
This is the subject most roadmaps leave out, and yet a deployment plan without a business model is a spending plan. The full reasoning is set out in what business model for AI inside software.
From cost centre to revenue centre
Badly billed, AI is a cost line that swells with usage. Properly modelled, it becomes a source of revenue: you buy capacity wholesale, you package it into units of value the client understands, you resell it with a margin.
Why bill in credits rather than tokens
The token is the technical unit of model consumption. It is a good internal cost indicator, and a very bad price indicator for the end client, for three reasons.
Instability. Model prices change regularly. Pricing tied to the token fluctuates with no understandable logic, which creates distrust.
Illegibility. No head of operations thinks in tokens. It is a unit with no meaning for the person signing the purchase order.
The perverse incentive. The better the AI gets, the fewer tokens it consumes for the same result. Billing per token means penalising yourself at every efficiency gain.
The credit fixes all three. One credit represents one successful action: a question handled, a file qualified, an import completed. Technical efficiency becomes a margin lever instead of reducing revenue.
Value-based pricing
The price of a credit is not calculated on production cost, but on the value saved at the end client. The basic rule: capture around 10 % of the value created. The client keeps 90 % of the gain. These values are starting assumptions; the final price is co-built with the sales teams in the first quarter, then validated against real usage in the second.
Start with a quantified pilot
The most effective way to convince an end client remains the pilot: a few weeks on one specific business flow, an allowance of credits provided, a quantified objective. At the end, you measure the real gain and convert it into a paid offer.
A good pilot client meets four criteria: a repetitive and measurable process, enough volume to see the effect quickly, a motivated internal contact, and value that is easy to express in time or money saved.
AI Act: building compliance in from the design stage
Compliance is not an end-of-project subject. For a software vendor, it is a design parameter to build in from the first weeks.
The European regulation on artificial intelligence distinguishes four levels of risk2. Systems applied to human resources, education or healthcare are classed as high risk, with documentation, transparency and human oversight obligations. The application deadline for those systems is expected during 2026; the exact date is still being adjusted at European level, but the principle does not change.
Concretely, building compliance in upfront means planning, from the specification onwards, the anonymisation of sensitive data, the traceability of the agent’s decisions and the human oversight mechanisms. These are expensive to graft onto a system already in production. The detail of what applies to business software is on AI Act.
Build, hire or outsource over 12 months
A roadmap is only credible if the execution model holds for its whole duration. Three options exist, and they complement each other more than they compete.
Hiring an experienced profile takes time and represents a high investment well before the first line of code reaches production. Outsourcing to a specialised partner lets you start immediately, with a go-live in the first quarter, then build autonomy progressively.
In practice, most vendors who start with a partner continue in a recurring collaboration once the first systems are deployed, while they build their own team. The two models follow one another along the roadmap.
Sources
- Gartner, “Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept by End of 2025”, July 2024.
- European Commission, “Regulatory framework on artificial intelligence”, digital-strategy.ec.europa.eu.