What "AI transformation" actually costs the second year.

The Year One AI business case is easy to write. Pilots are cheap on paper. The wins are visible. The costs are lumpy but bounded. The board signs off; the programme kicks off; the pilots produce their promised results; everyone feels good.

Year Two is the year the operating model has to work. And Year Two is where every programme we’ve walked into as second-opinion reviewers has missed at least three of the five cost lines below.

This piece is the version of the conversation we keep having with technology leaders in Q3 of the first year of an AI programme, when Year Two budget is being drafted.

Cost line 1 — Platform

Year One usually gets away with somebody’s team standing up a small platform that runs the pilots. Retrieval infrastructure, prompt management, eval tooling, observability, cost dashboards. It’s not strategic; it’s whatever the pilot team could ship on their own budget.

Year Two, three problems collide:

• Three other business units want to build on that platform now.
• The observability that was fine for one team is unreadable across five.
• The person who built it is bored and wants to move to something else.

Year Two cost isn’t the platform itself. Year Two cost is the funded, staffed, operated platform team whose job is to make the platform a reliable substrate for the rest of the organisation. That’s usually three to five people, cross-functional. The line item nobody had forecast.

What to do in Year One: name the platform team’s remit in the Year One strategy document, even if you’re not staffing it yet. Budget for it in the Year Two forecast, at real headcount, not as a project.

Cost line 2 — Data

The Year One pilots ran on data somebody had cleaned up specifically for the pilots. That data is not the data Year Two runs on.

Year Two costs, in every programme we’ve reviewed, split into two:

• Pipeline debt. The pipelines that fed Year One pilots were brittle. Year Two either invests in real pipelines or pays a repair tax that scales linearly with usage.
• Governance debt. Year One let people use data casually because there were only a handful of use cases. Year Two has a compliance officer asking about lineage, retention, and consent for every model call. That work is real.

Budget both. The pipeline number is usually the platform team’s third or fourth priority (see cost line 1). The governance number is a dedicated role, either inside your existing data-governance function or as a new capability inside the AI programme.

Cost line 3 — Capability

Year One pilots ran because you had a small number of people who knew what they were doing. Year Two needs those same skills spread across ten times as many people to sustain the wider adoption.

The temptation is to run a big training programme. Training programmes usually miss because they train the syllabus, not the actual skill.

The Year Two capability line, done well, has three parts:

• Cohort training on real projects. Not workshops in isolation — training embedded in delivery on live problems.
• Practice leadership. One or two senior people whose remit is the practice itself, not any specific project.
• Hiring. For the roles the Year One programme discovered are actually rare — usually senior AI engineers with production experience, occasionally AI-fluent product managers.

Budget capability at 15–25% of the AI programme envelope. It reliably gets under-budgeted because it doesn’t show up on the pilot cost lines.

Cost line 4 — Vendor consolidation

Year One let every team pick its own tools. That’s fine — Year One is supposed to sample. Year Two has to pay the bill.

Two Year-Two costs to prepare for:

• The migration off tools that didn’t stick. Some fraction of the Year One vendor choices will be wrong. Migrating off them costs real engineering time.
• The commitment discount on the tools that did. Vendors love an annual commit; annual commits mean less headache but more money spent before it’s needed. The negotiation position — and the fallback path if the vendor becomes unreliable — needs to be prepared in Year One.

Do the tool audit in Q4 of Year One, not Q1 of Year Two. Q1 is the month you’re signing the renewals.

Cost line 5 — Governance and risk

The Year One pilots each had somebody nervously reading the newest AI regulations while they were being built. By Year Two, that role is a real role, not a nervous evening.

Costs to plan for:

• A named AI risk owner. Not a committee; a person with a written remit, a reporting line into risk, and enough seniority to say no.
• Model and prompt inventory. What’s in production, running against what data, evaluated how, owned by whom. Boring, unavoidable, belongs in a system-of-record.
• Regulatory readiness. Depending on your jurisdiction, this ranges from “watching the space” to “delivering on a specific compliance regime”. Either way it’s not zero cost.

Composing the Year Two forecast

The Year One AI business case usually looks like:

• Pilot delivery cost
• Some model spend
• A little bit of consulting
• An enablement line

The Year Two AI operating cost, done honestly, looks like:

• Platform team (headcount)
• Data pipeline + governance (headcount + capex)
• Capability programme (headcount + training)
• Vendor + tool commits (opex)
• Governance and risk (headcount)
• Ongoing model spend (opex, now much larger)
• New project delivery cost (opex, similar to Year One)

Year Two is not more expensive because AI is more expensive. Year Two is more expensive because Year One under-forecast the organisational capability needed to sustain what the pilots proved was possible.

What to do about it

Three things, in Year One:

Write the Year Two operating budget in Q3 of Year One. Not the strategy — the operating budget. Line by line. Whether you get the numbers right isn’t the point. Getting the categories right is the point.

Name the operating roles that Year Two needs, in Year One. Platform lead. Data governance lead. Practice leader. AI risk owner. Even if you don’t fill them yet, the org chart Year Two needs shouldn’t be invented under time pressure.

Say the honest number out loud. In our experience, Year Two runs at 1.5–2× the Year One number, and most of the increment is the operating costs listed above. Leadership can absorb that number when it’s delivered in Year One with a story. They cannot absorb it when it arrives as a surprise line item in Year Two.

So what?

Every programme we’ve reviewed at the Year One / Year Two boundary has been on the same journey. Pilots worked. The operating model didn’t scale. The Year Two forecast under-shot by roughly the same set of line items every time.

If you’re mid-Year One, the moves that matter are cheap: name the categories, sketch the roles, tell leadership the honest number. Year Two is much easier when it’s expected.

Related reading

• The board deck slide we keep having to write — the specific slide that keeps carrying this argument.
• Evals for agents that use tools, not just tokens — the measurement discipline that makes Year Two operations sustainable.