Foundation first. Everything else depends on it.
Three stages, run in sequence. Foundation is the one you can buy today — four to eight weeks, fixed scope, defined outcome. The two stages after it describe where the work goes, not a package waiting to be sold.
Most organizations try to start at the second stage.
Eighty-eight percent of organizations use AI somewhere; 7 percent are fully scaled and 39 percent report enterprise-level EBIT impact (McKinsey Global Survey on the State of AI, November 2025). The gap is almost always sequencing. Work that skips the foundation does not fail loudly — it produces pilots that never leave the demo, and a second year that looks like the first. Only Foundation is available as a defined program today; the other two are described here so you can see where the work goes, not because they are for sale.
- 01
Foundation
4–8 weeks
Answers one question: are we ready to run an AI program responsibly today? Four layers get scored rather than debated — strategy and culture, data and governance, technology, people and skills. One domain is chosen and one accountable executive sponsor is named, not a steering committee. Governance is drafted against real risk tiers, tooling is stood up and configured, and people are trained on their own work. You end with a maturity assessment, a roadmap and a 90-day plan with an owner per item.
Build Your AI Foundation→ - 02
Acceleration
Forward stage
Two questions in order: what is actually possible in the chosen domain, and can we make one thing work end to end? Discovery with frontline operators produces a ranked shortlist of candidate use cases and an executive team able to challenge proposals on substance. Then one of them is rebuilt — not the existing workflow automated, the workflow redesigned for the outcome — and put into production against a metric the business already tracks. The discipline here is refusing to run five of these at once.
- 03
Business Transformation
Forward stage
The rebuilt workflow extends across the domain, and the foundation underneath it stops being per-project heroics: data architecture modernized, quality maintained continuously, evals and quality gates routine, funding and talent ownership settled. Eventually the test is not whether AI is used but where the decisions happen — new use cases get approved at functional leadership level, they largely self-fund from earlier returns, and board conversations are about capability rather than pilot counts.
Five things that hold true at every stage.
These are not stage-specific. They apply on week one of Foundation and they still apply two years in. Most of the expensive mistakes in an AI program are one of these five, ignored early.
Deep and narrow beats shallow and broad
Pick one domain and rebuild it end to end rather than seeding pilots across every function. Only 4 percent of companies take that focused approach, and those that do earn twice the ROI (2024 BCG survey of 1,000 CXOs, reported in Harvard Business Review, November 2025). Breadth is the reward for depth, not the route to it.
Redesign the workflow, not the task
Automating the steps people already perform makes an existing process marginally faster and usually more expensive to run. The value comes from redesigning for the outcome — which changes who does what, in what order, with which handoffs. That is an operating-model decision, so it needs an executive owner, not a tools budget.
Audit the foundation before chasing use cases
Strategy and culture, data and governance, technology, people and skills. Aim the first project at the weakest of those layers rather than the flashiest application. Self-assessment usually runs ahead of the evidence: organizations that rate themselves as optimizing often turn out to be experimenting.
Aim at growth, not only efficiency
Roughly 80 percent of firms set efficiency as the objective (McKinsey Global Survey on the State of AI, November 2025). Efficiency targets produce cost cases that are hard to defend once the novelty passes, because the savings are capped and the disruption is not. Growth and innovation objectives survive the second budget cycle.
Budget for the complementary assets
Training, change management, process redesign and the agreement of the people whose work changes are the bulk of the investment. The software is the visible fraction. A license line item is not a plan — budget for the part nobody demos, because that is where the money and the calendar actually go.
An illustration, not a client story.
No organization below is real, and nothing here is a case study or a result Provectia is claiming. It is the sequence the framework produces, written generically so you can hold your own organization against it and work out which stage you are actually at.
- Starting position
A mid-market organization with AI everywhere and nowhere. A dozen departmental subscriptions bought on expense cards, no policy anyone can locate, two pilots that impressed a committee and then stopped, and a board asking what the plan is. Internally this gets described as being well along. On the evidence it is stage zero.
- Stage 01 · Foundation
Four to eight weeks. The four layers get scored against evidence, every proposed use of AI is mapped to a risk tier, and one domain is chosen — which means several others are explicitly deferred, in writing. A single executive owns the outcome. Governance, configured tooling and trained people land together, because any one of the three without the other two decays.
- The decision point
Foundation ends with a roadmap and a 90-day plan, and that is a genuine stopping point. Many organizations run the next quarter themselves with what they own. Others want continuing executive-level AI leadership without hiring a full-time chief AI officer. That is a separate decision made afterwards, on evidence, not a condition of starting.
- Stage 02 · Acceleration
Whoever runs this stage — an internal team, a systems integrator, a fractional leader — starts with discovery inside the chosen domain, run with the people who actually do the work. That produces a ranked shortlist. One candidate is selected and its workflow rebuilt for the outcome, then shipped into production on a 90-day cadence with the outcome metric defined in week one. The common failure is optimizing individual productivity while the surrounding workflow still runs on manual handoffs.
- Stage 03 · Business Transformation
The rebuilt workflow extends across the domain. Data architecture and quality get modernized because the next five use cases cannot each fund their own rebuild. Governance, evals and funding mechanics become routine, so a new use case stops re-litigating the basics. Scaling here means depth in one place before breadth across the company. By this point the organization is running its own program; the framework describes that arc, it does not queue up engagements to sell against it.
- The board test
At any stage, an AI roadmap needs two columns: company impact — industry shift, customer expectations, revenue and pricing pressure — and operating model — workflow redesign, data access, evals, governance, talent, funding. If either column is blank, it is a pilot list, not a strategy, and it should go back. That is the test Provectia applies to every roadmap it reviews.
Start at stage one.
Foundation is the only stage available as a defined program today: four to eight weeks, scope and price agreed in writing before anything starts. It is also the only stage that makes the next two worth attempting.