Field notes · a build order for an AI practice
What an AI practice is actually made of
AccentureDeloitteMcKinsey / QuantumBlackBCGPwCEYKPMGCapgeminiInfosysTCS
Where to start, and why
Figure 1
The numbers in the small circles are the recommended build order below — start at 1, not wherever looks most exciting. Position on the plot is market research: how crowded each area is, and how solved.
← few firms competinghorizontal: how crowded · vertical: how solvedeveryone sells it →
Crowding and maturity usually get collapsed into one story. Split apart, most of the current noise sits in the lower right: heavily marketed, not yet operationally settled.
Least settled in the reading
Widely delivered already
# Recommended build order
Where the practice came from, where it seems to be going
Figure 2
Left of the rule is observed — what the ten firms published and sold. Right of it is forecast, and the confidence drops as it goes: the two-year column is mostly extrapolation, the ten-year column is a hypothesis I expect to be wrong about in detail.
Observed · 2019–2026
+2 years · 2028
+5 years · 2031
+10 years · 2036
2019Analytics and ML practices; data science pods inside digital units.
2023Generative AI arrives as a strategy offer. Use-case lists, maturity assessments, executive workshops.
2024Copilots and knowledge assistants reach production in service, finance and engineering.
2025Firm platforms and agent libraries launch; governance becomes a standalone practice; agents begin taking actions, not just drafting.
2026Agentic workflow redesign is the headline offer everywhere. Evidence of operated, evaluated agents remains thin.
Buyers ask for evidence, not demos: evaluation results and incident history enter the RFP.
Delegation and permission design become a named workstream with a named owner.
Pricing shifts from day rates toward outcome and per-resolution terms.
Consolidation: fewer, larger platform bets per firm.
Confidence: moderate
Managed AI operations outgrow implementation revenue for the firms that got there first.
Agent assurance looks like audit: recurring, evidenced, third-party.
Practice IP is licensed as much as delivered; consultants configure more than they build.
Human-agent operating models are a mainstream org-design discipline.
Confidence: low
“AI practice” stops being a separate unit — it is how every practice delivers.
The durable asset is accountability infrastructure, not models.
Firms compete on the quality of their evidence, the way auditors compete on trust.
Confidence: speculative
How to structure the practice
Figure 3
The hub owns the reusable parts. Everything client-facing runs through a spoke. This is the shape before you decide what to build first.
Hub
Central AI practice
Architecture · platform · evaluation · responsible AI · accelerators · standards · partnerships · talent
Spoke
Industry & functional teams
Client relationships · workflow expertise · business cases · regulatory context · value realisation
Spoke
Delivery & engineering
Data engineering · build · integration · cloud · testing · deployment · managed operations
Spoke
Ecosystem partners
Hyperscalers · model providers · workflow platforms · specialist vendors — integrated, not resold
Observed pattern: the AI practice supplies technology, method and controls, and the established practices supply workflow knowledge and the relationship. The reported failures cluster on that boundary.
Guiding principles
Placeholder · to be researched
Five working principles — this is also where research turns into sequencing. For each principle: what in the current market it responds to, which part of the forecast it bets on, and where it lands in the build order below.
PrincipleResponding to (observed)Betting on (forecast)Horizon
Recommended build order
Foundations and commodity wins first, guardrails before frontier work, managed services as the compounding end state. This is a judgment call, not a finding — disagree with it in the areas below.
Six shapes the work takes
Figure 4
EngagementDuration (weeks, to 104)Primary output
The areas, one post each
13 areas · 13 field tests
Browsing by theme? Start here. Browsing by what to build first? See the build order above — each card below also shows its build step.
Still open
Commercially important, operationally immature, and barely covered in what the firms publish. These run through the whole series.