MODULE 3 · LESSON 3
Free — no login requiredSign in to track progress, save quiz attempts and enrol in the full course.
Sign in to track progress / enrolBuilding the Capability, Not Collecting Pilots
Organisations that succeed with AI do not do one big thing. They run a sequence, and the order matters more than the speed.
A workable sequence
The counter-intuitive parts are the ordering of steps 1 and 4, and the breadth of step 3.
Why pilot before strategy? Because a strategy written by people who have never delivered an AI project is a list of guesses. After one real project you know how long the data work takes in your organisation, which systems are impossible to extract from, who is enthusiastic and who is obstructive. Those facts make a strategy worth writing.
Why train broadly? The binding constraint is rarely the number of engineers. It is the number of people who can look at their own function and recognise a well-shaped AI problem. A single trained operations manager who spots three good projects is worth more than a fourth data scientist with nothing sensible to work on. This is exactly why a course like this one exists for non-technical staff.
Central team or embedded specialists?
Both patterns work; they fail differently.
A central AI team — one group serving the whole organisation. Concentrates scarce expertise, avoids each department buying incompatible tools, and makes standards possible. Its failure mode is distance: a central team can end up building things the business does not want, prioritised by whoever lobbies hardest.
Embedded specialists — data people sitting inside marketing, operations, finance. Close to the problems, quick to iterate, trusted by their colleagues. Their failure mode is isolation: three people solving the same problem three ways, nobody sharing infrastructure, and career development that stalls because there is no professional community.
Most organisations of any size end up with a hub and spoke arrangement: a central group owning platform, standards and hard problems, with embedded people close to the business. Getting there deliberately is better than getting there through two failed reorganisations.
For a smaller organisation, the honest answer is usually simpler: one or two people, working closely with whoever owns the process being improved, and buying rather than building anything generic.
What a real AI strategy contains
Most documents titled "AI Strategy" are a list of technologies and an aspiration. A useful one answers five questions:
- Where could AI create advantage for this business specifically? Not "AI is transformative" — which processes, which decisions, and why here rather than anywhere.
- What data do we have that others do not? This is the durable advantage. Algorithms are broadly available; your operational history is not. If you have no distinctive data, be honest — your AI advantage will come from execution speed, not from a moat.
- What are we going to build, and what are we going to buy? Build where you are distinctive, buy where you are generic.
- What has to be true first? Usually data unification, and usually longer than anyone wants to hear.
- What will we not do? A strategy without exclusions is a wish list.
Question 2 deserves the most attention. The most valuable strategic move available to many companies is not a model at all — it is deciding to start recording something they currently throw away, so that in eighteen months they have a dataset nobody else can obtain.
The playbook above is written for organisations with departments. Most are smaller. A realistic version for a 40-person services business:
Months 1–3. Pick one repetitive, high-volume, low-risk task. Something like categorising inbound enquiries, extracting data from a recurring document, or drafting first-pass responses. Use an off-the-shelf tool. Do not hire anyone. The goal is a working thing and a real opinion about whether it helped.
Months 3–6. Measure it honestly against the baseline. Was time actually saved, or moved? Did quality hold? If yes, expand to one adjacent task. If no, say so plainly and stop — the discipline of stopping is what keeps credibility for the next attempt.
Months 6–12. Now the important, unglamorous move: fix the data problem the first project exposed. It always exposes one — a field nobody fills in, two systems that disagree, a process where the outcome is never recorded. Fixing it is worth more than a second model.
Throughout. Train broadly and cheaply. Have the people doing the work learn what these tools can and cannot do, so proposals come from the people closest to the problems.
What not to do. Do not hire a data scientist as your first move — with no data infrastructure and no defined problem, a capable person will spend a year frustrated and leave. Do not start with the hardest problem because it is the most valuable. Do not build what you can buy.
At this scale, AI capability is mostly organisational habit: noticing repetitive judgement work, being willing to measure honestly, and fixing data at the source. None of that requires a specialist headcount.
Why does a realistic sequence put a working pilot before writing an AI strategy?
The sequence
Click to flipPilot first, then build a small in-house team, then train broadly, then write the strategy, then communicate. The ordering matters more than the speed.
Click to flip backBuild capability in sequence: a pilot chosen to succeed, then a small internal team, then broad training, then a strategy grounded in what you learned, then honest communication. Train non-technical people widely, because spotting well-shaped problems is scarcer than building models. And make the strategy answer real questions — especially what data you hold that others cannot get, and what you are deliberately not going to do.