Baseline, Sequencing, Gaps and Phasing

A roadmap for AI is a set of decisions about what your organisation will do with AI, in what order, with whose money and under whose authority. It is not a list of tools. Starting from a tool produces a plan that serves the tool; starting from the organisation's work, risks and capacity produces one that can survive reality. This lesson covers the baseline, sequencing, capability gaps and phasing.

Start with the baseline you already have

Almost no organisation starts from zero. Before anyone ranks new ideas, someone must write down the AI already in use. That includes tools that were bought on purpose, but also two categories that leaders routinely miss. The first is features that arrived by vendor update: the email platform that now summarises threads, the customer system that now drafts replies, the document editor with a built-in assistant. Nobody decided to adopt these, yet they touch real work and sometimes personal data. The second is unsanctioned use: staff who paste text into a personal account of a public tool because it saves them an hour. This is rarely malicious; it usually signals unmet demand.

Useful sources for the baseline are vendor release notes and administrator consoles, card statements and expense claims that show subscriptions, short and non-punitive conversations with team leads, and a review of what the contracts permit. Each entry needs a named owner, a description of the work it touches, the kind of data involved, and a decision: keep, restrict, replace or stop. An entry with no owner is itself a finding.

Sequence by value, risk and readiness together

Once you know what exists, you can rank what to add. Three questions must be asked together. How much value would this deliver, and how would we recognise it? How much harm could it do if it is wrong, and who would carry that harm? How ready are we in terms of data, people, process and ownership? Ranking on value alone favours exciting, high-stakes ideas the organisation is least able to run safely; ranking on readiness alone favours trivial work. Sensible early items are moderate in value, low to moderate in consequence and genuinely ready, because they build the skills and evidence that harder items need. A high-value item that is not ready is not rejected; it is sequenced later, with the gaps named and someone assigned to close them.

Name the capability gaps

Most stalled AI initiatives stall for the same reasons, and none of them is the model. The data may be scattered, out of date or not permitted for this use. The skills to specify, test and challenge the system may be thin. There may be no one who owns the outcome, so everyone assumes someone else does. And there may be too little review capacity: if a person must check every output and that person is already full, the initiative has a hidden bottleneck that will either slow the work or quietly turn review into a rubber stamp. A roadmap should list each gap beside the initiatives that depend on it, with an owner and a date.

Phase the work with stage gates

The usual pattern is three phases: a bounded test, then repeat use in a defined setting, then wider rollout. Each move to the next phase passes through a gate, which is a decision by a named person against criteria that were written before the test began. A bounded test answers a narrow question with a limited group, limited data and a clear end date. Repeat use asks whether the result holds up over weeks, with ordinary staff, in unusual and incomplete cases as well as routine ones. Wider rollout asks whether support, monitoring and ownership can stretch to the new volume. Skipping a phase is the commonest way a promising pilot becomes an incident, because the first time the system meets the full variety of real cases is the first time it is live.

A worked example

Fenwick Schools Trust is an invented group of primary schools. Its chief operating officer is asked for an AI roadmap by the trustees. She asks the schools' business managers and the IT lead to list the AI already present. The list turns up a pupil-information system whose latest update added a feature that drafts parent letters, a few teachers using a personal account of a public chatbot to write report comments, and an unused licence for a marking assistant bought a year earlier. None of the three had an owner.

She then scores three candidate ideas. Drafting routine parent newsletters is modest in value, low in consequence and ready. An assistant that suggests pupil support plans is high in value but touches sensitive data, has no clear owner and would need a specialist to review each output, and that specialist has no spare time. Automating attendance letters falls between the two. She proposes the newsletter work as a bounded test with a named headteacher as owner and an end date, assigns an owner to each baseline item, and lists the support-plan idea as later, with three named gaps: data permissions, a trained reviewer and an owner. The trustees receive a short roadmap that says what is being done, what is waiting and why.