Insights
AI implementation roadmap
An AI implementation roadmap is the sequence of phases that takes you from a readiness score to a system actually running in production, with a decision gate at each step so you can stop before you waste money. Here is the whole template, free, and the honest note about what kills each phase.
The roadmap, phase by phase
Read this top to bottom. Every phase has a gate: a specific condition that has to be true before the next phase starts, and a permission to stop if it is not. The gates are the point. A roadmap without them is just a list of hopes with dates attached. The template prints cleanly from your browser if you want it on paper or as a PDF; we do not gate it behind a form.
| Phase | Typical duration | Owner | Go/no-go gate | What kills it |
|---|---|---|---|---|
| 0. Readiness assessment | 3–6 weeks | Executive sponsor, with outside eyes | A scored read across six dimensions and one named constraint. | Skipping it and guessing which problem to solve first. |
| 1. Fix the constraint | Weeks to a quarter | The owner of the weakest dimension | Your weakest dimension is out of the Foundations band. | Building on a foundation the assessment already flagged as weak. |
| 2. Scope the first use case | 1 to 2 weeks | Sponsor plus the process owner | One named process, a written definition of "working", a number. | "Do AI" with no named process and no success test. |
| 3. Time-boxed pilot | 6 to 12 weeks | Delivery lead | It hits the success definition with real users, not a demo. | A pilot with no kill criteria, quietly extended forever. |
| 4. Production integration | 8 to 16 weeks | Engineering, with the system owner | It runs inside your real systems, with guardrails and an owner. | No integration path, or nobody signed up to own year two. |
| 5. Operate and scale | Ongoing | A named operations team | Monitored, budgeted, with an escalation path when it is wrong. | Unowned usage spend and no process to catch mistakes. |
Why it starts at zero
Phase 0 is the assessment, and it is numbered zero on purpose: it is the work that decides whether the rest of the roadmap is even pointed at the right problem. It scores you across six dimensions, strategy, data, systems, people, governance, and operations, and hands back one thing that matters most: your weakest dimension. That constraint is where the roadmap actually begins, because it is where your first project will break if you ignore it. Score yourself free with the interactive readiness assessment or the readiness checklist, and see the how to assess AI readiness guide for the full scoring method.
Phase 1: fix the constraint before you build
If the assessment put a dimension in the Foundations band (a score of 4 or less), that is phase 1, and it comes before any AI build. There is no point piloting a model on data nobody can locate, or shipping an agent into a company with no rule about where data may be sent. This phase has no fixed length because it depends on what is broken: a missing owner can be fixed in a week, a genuinely messy data estate in a quarter. The gate is simple: the weak dimension is out of Foundations before phase 3 starts.
Phase 2: scope one use case, narrowly
Pick one process. Not "customer service", but "answer order-status questions in chat". Write down what "working" means as a number you can measure, and what you will not automate. The single most common failure on the whole roadmap is a use case defined as "do AI in support", which can never be finished because it was never defined. Narrow scope is not timidity, it is the only thing that makes a pilot judgeable.
Phase 3: pilot with a kill switch
Run it time-boxed, on real users, against the success definition from phase 2. The rule that keeps a pilot honest is deciding the kill criteria before you start: the numbers at which you stop, not just the ones at which you continue. A pilot with no way to fail becomes a permanent, unowned pilot, which is how a lot of AI budget disappears. If it hits the definition, it graduates. If it does not, you have learned something cheaply, which is a win the roadmap is designed to allow.
Phase 4: production is a different project
A pilot that works is not a system that works. Production means wiring it into your real systems, adding the guardrails and logging, handling the edge cases the pilot dodged, and naming who owns it. This is usually the longest phase and the one that surprises people, because the demo was the cheap part. The gate is that it runs inside your actual systems, with a defined process to catch it when it is wrong, and an owner for year two. If no one will own year two, do not ship it.
Phase 5: operate, then scale
Live systems need an owner, a budget with a ceiling on usage-based spend, and an escalation path for when the AI is wrong, because every AI system is wrong sometimes. Only once one use case is genuinely operating this way is it worth starting the next, which loops back to phase 2. Scaling is repeating the loop with the muscle you built, not launching ten pilots at once and hoping.
Where the effort really goes
One number worth holding onto: the widely cited 10-20-70 rule puts about 70% of the work of a successful AI effort in people, process, and change, and only 10% in the model itself. The roadmap reflects that. The model is a small part of phases 3 and 4; the People, Governance, and Operations dimensions run through every phase, and they are where projects actually succeed or stall. Plan for the 70%, not the 10%.
From template to a real, costed roadmap
This template is the shape of the plan. Turning it into your plan, with your constraint, your use cases, real durations, and a cost against each phase, is the output of our paid assessment: our senior team inside your systems for 3–6 weeks, at a fixed price of $20,000 to $80,000, ending in a prioritized, costed roadmap you own. Use this skeleton to run it yourself, or start with an assessment and we build the costed version with you.
Questions people ask
- What is the roadmap for AI implementation?
- At the level that survives contact with reality, six phases: assess where you stand, fix whatever the assessment flags as your weakest foundation, scope one specific use case with a defined success test, run a time-boxed pilot, integrate the winner into production with guardrails and an owner, then operate and scale. The order matters more than the labels. Most failed projects skipped straight to a pilot without the assessment or the scope, and found the missing foundation the expensive way.
- What is the 10-20-70 rule for AI?
- A guideline, attributed to BCG, that a successful AI effort spends about 10% of its energy on the algorithm or model, 20% on the technology and data around it, and 70% on the people, process, and change management. The lesson for a roadmap is that the model is the small part. Most of the work, and most of the failure, is in the People and Operations dimensions, which is why they are in the assessment and why phases 4 and 5 above are the ones that actually decide success.
- What is the 30% rule for AI?
- A rule of thumb that AI should handle roughly the routine 30% of a process while people keep the judgment calls. It is a decent guard against over-automating in a roadmap: scope your first use case around the repetitive core, keep a human on the exceptions and anything irreversible, and do not try to automate the whole process at once. The right split varies, but starting narrow is almost always correct.
- How long does AI implementation take?
- Published roadmaps commonly put an end-to-end enterprise effort at 6 to 18 months, and that matches what we see: a few weeks to assess, a quarter or so to fix a weak foundation and pilot, then a few months to reach production and operate it. Anyone promising a production system in two weeks is selling a demo. The durations in the template above are typical single-track ranges; running several use cases at once stretches the calendar and the people, not the phases.
- What comes before the roadmap?
- The assessment. A roadmap built without knowing your weakest dimension just sequences guesses, and the first phase will be the one you skipped. Score yourself first with the free readiness assessment or the checklist, find the constraint, and let the roadmap start there. The roadmap is the plan; the assessment is what tells you where the plan should begin.
Tell us about the work.
A few lines is enough. We read every enquiry ourselves and reply within one business day.