If your plan for next year involves "exploring AI across the organization," you don't have an ai roadmap for business, you have a mood. I've watched enough companies spend a full year "exploring" and end up with a ChatGPT subscription nobody's tracking and a slide deck full of use cases nobody built. The businesses that actually got value this year did the opposite: they picked one flagship project, gave it an owner and a number, and left everything else for later.
This is the point in December where every company writes a roadmap slide with six workstreams and a "TBD" next to each one. That slide will still say TBD in June. A real roadmap fits on one page because it only commits to what someone is actually going to build, not everything that theoretically could be built.
Here's the structure I'd put in front of any owner or CTO heading into January.
One flagship project, named and owned
Pick exactly one AI project that matters enough to survive a busy quarter. Not the most exciting one, the one with the clearest, measurable business outcome. Give it a named owner, someone whose job performance next year actually includes this project, not a committee.
Examples of flagship-sized, not roadmap-sized:
- A multifinance company automating first-pass document verification for loan applications, targeting a 40% reduction in manual review time.
- A retail chain building a demand forecasting workflow off two years of POS data, targeting a measurable drop in dead stock by Q3.
- A services firm building an internal chat-with-docs tool for staff, targeting a reduction in "where's the SOP for X" questions to a support channel.
Notice each has a number attached before the project starts. If you can't state the metric before you build it, you don't have a flagship project yet, you have an idea. Go back and find the number first.
Two quick wins, capped at two weeks each
Alongside the flagship, pick two small automations that any competent person can finish inside two weeks, using tools you likely already pay for. These exist to build organizational confidence and momentum while the flagship project takes its slower path through data access, integration, and testing.
Good quick-win candidates: an auto-categorization rule for inbound support tickets, a scheduled AI-generated summary of weekly sales sent to leadership, a draft-reply assistant for a high-volume, low-complexity customer question. None of these need a vendor contract or a build sprint. If you did the "quiet week" exercise before the new year started, some of these might already be done. See Quiet Week Projects for the kind of scope that fits here.
A training plan, not a training event
The single most common reason a flagship AI project stalls after launch isn't the technology, it's that the staff who have to use it daily were never actually trained on it, they were shown it once in a meeting. Put a real training plan on the roadmap: who gets trained, on what, by when, and how you'll know it worked.
I'd rather see "two 90-minute sessions with the collections team in February, followed by a week of shadowed usage" than "training TBD." The vague version is the one that never happens. For what actually derails this step, see Training Staff to Work With AI, Not Around It, the failure mode is almost always staff quietly reverting to the old manual process because the new tool wasn't explained in terms of their job, only in terms of the technology.
A build-vs-buy decision, made once, in writing
Before any of this gets built, decide explicitly whether the flagship project uses an off-the-shelf tool or a custom build, and write down why. This decision gets re-litigated constantly when it isn't documented, wasting a meeting every month re-explaining the same tradeoff. I go through the actual decision criteria in Off-the-Shelf AI vs Custom AI Workflows, but the short version: buy when the workflow is standard across your industry, build when the workflow is what makes your business actually different from competitors.
A review date, on the calendar now
Put a quarterly review on the calendar in January, not "sometime in Q2." Three checkpoints for the year, each with the same three questions: is the flagship project hitting its number, did the quick wins actually stick or quietly get abandoned, and does the training plan need to repeat for new hires. A roadmap without a scheduled review date is a wish, not a plan.
What the one-page version looks like
| Item | Owner | Metric | Deadline |
|---|---|---|---|
| Flagship: loan document auto-verification | Head of Ops | 40% cut in manual review time | Q2 review |
| Quick win 1: support ticket auto-tagging | Support Lead | 90% correctly tagged | End of January |
| Quick win 2: weekly sales summary automation | Finance | Delivered every Monday 7am | End of January |
| Training | Ops + HR | 2 sessions, shadowed week | February |
| Review dates | Owner | On track / off track | End of Q1, Q2, Q3 |
That's the whole roadmap. If it needs a second page, something on it isn't actually a commitment yet.
The takeaway
An ai roadmap for business earns its name by fitting on one page and naming a real owner for a real number. Pick one flagship project you can defend with a metric, two quick wins that ship in January, a training plan with actual dates, and a quarterly review already on the calendar. Everything else is a good idea for later, not a plan for now. If you want a second opinion on which project deserves the flagship slot, that's a conversation worth having before you commit a whole year to it, reach out at ervandra.com/partner.