The clock shapes the plan
AI roadmaps written for corporates assume unlimited time. A hold period doesn't have that luxury: value created late barely compounds before exit, and a buyer will discount unfinished work at sale.
The useful sequencing question is: what should exist by exit, and when must each piece start? The roadmap below works backwards from a four-year hold. Treat the years as phases, with a five-year hold stretching the middle and a three-year hold compressing it.
Year one: prove and pipeline
Year one has two jobs: bank a visible win fast, and build the pipeline of everything that follows.
The win comes from one of the standard quick-payback builds: automated lead response, document intake, or reporting automation. Choose it for speed, ship it to production within the first quarter of the programme and measure it against a baseline. The fastest project we have run paid back in twelve weeks, and that is the right ambition for the first build. The purpose is as much political as financial: management teams that have seen one system work stop debating whether AI is real.
The pipeline comes from an opportunity scan: every significant process ranked by volume, cost, error rate and revenue proximity. The output is a prioritised backlog with a number against each item. This becomes the AI section of the value creation plan, reviewed at board level like any other lever.
Year one should also fix the data plumbing required later. This does not require a grand data platform, only reliable pipelines out of the core systems that each subsequent build can use. The problems are rarely exotic, but they are real: we recently came across a wealth management business whose core platform feed effectively erased itself every 24 hours, with historical data recoverable only by emailing the vendor. Nothing intelligent can be built on top of that, so a simple retention pipeline came before any AI at all.
Year two: compound the margin
Year two is the volume year. With trust established and data flowing, work through the backlog's margin builds: finance automation, customer service triage, procurement analysis, operations exception-handling. Each build follows the same discipline as the first: baseline, production deployment, measured impact, handover.
Two things matter structurally in year two. First, adoption becomes the constraint rather than technology, so training and process change deserve as much attention as the builds themselves. The approach that survives is embedding AI in the surfaces people already use, such as a bot in the chat tool that updates the CRM after a meeting, rather than launching an internal app store of agents nobody asked for. Second, patterns start repeating: the second document-intake system costs a fraction of the first. This is where a portfolio-wide approach pays, because patterns proven in one portco transfer to the next.
Year three: reach for revenue
With the operational base solid, year three earns the right to the higher-risk, higher-reward work: pricing engines, sales intelligence, customer-facing AI where the brand risk is now manageable because the data and the team are mature.
Pricing deserves special mention because it is routinely the largest single prize. At Healf, the FT's #1 fastest-growing company in Europe, pricing was among the AI projects our founding team led that added over £10m of annualised revenue to a business running at £100m. Work like that is only safe on top of the data foundations laid earlier, which is exactly why it sits in year three and not year one.
Year three is also when to revisit the deliberate omissions of year one, including platform decisions and the ambitious ideas from the first workshop. Some will now be justified, although most still will not.
Year four: package for exit
The final year's job is to convert the programme into multiple. Buyers now probe AI in diligence, and the difference between a portco that answers well and one that doesn't shows up in price.
Answering well means: systems documented, impact quantified with baselines, maintenance institutionalised rather than resting on one person, and data assets clean and demonstrable. It means the management presentation can say "these eleven processes run without human touch, at these costs, measured monthly" instead of gesturing at innovation.
Nothing new and ambitious starts in year four. A half-finished AI project creates another diligence issue, so spend the available budget hardening and documenting what already exists.
The discipline that ties it together
Across all four years, every initiative should carry a number, a bar of the value bridge and an owner. Report it alongside pricing and procurement in the same table. This keeps AI under the same scrutiny as every other value creation lever throughout the hold.