ExplainerAIStrategy

Which AI Platform? Five Rules to Match the Tool to the Task

Climb a five-step ladder only as fast as governance, skills and data allow, and keep the choice portable.

OMBIRO4 min read
THE PACKAGEDCORELedgerProduct enginesAgentsIdentityACCOUNTABILITY STAYSJobs leave the core; accountability stays with the bank

Sam Altman, OpenAI’s chief executive, argued in 2025 that as the cost of a given level of AI falls, lower prices “lead to much more use.”1 For organisations that sentence has two consequences. Usage will spread whether or not anyone approves it, and the bill will grow faster than the unit price falls. McKinsey’s May 2026 survey found 93 percent of respondents exceeding their AI budgets, with spend rising nearly fourfold between isolated use cases and enterprise-wide adoption.2

The gap is guidance, not access

Most staff already have an AI assistant in hand. What many lack is direction. In a 2025 study by the University of Melbourne and KPMG covering more than 48,000 people in 47 countries, only 40 percent of employees said their workplace had a policy on generative AI, and 57 percent said they hid their use of it.3 The platform question is therefore not whether to start. It is how far and how fast to climb, and on whose terms.

The question of terms matters. Strive Masiyiwa, founder of Cassava Technologies, put the principle for Africa at a United Nations General Assembly side event in 2025.

Africa must be “a producer of solutions, not just a consumer of technologies developed elsewhere.” Strive Masiyiwa, founder, Cassava Technologies, 20254

It generalises. A platform choice should leave an organisation able to build, switch and own its data, not merely rent a capability.

1. Write the policy before you buy anything

Start with what the organisation already owns. A usage policy, a register of AI uses and practical training cost little, and they close the gap above. Make the policy short enough to remember: what may be used, on which data, who checks the output and who is accountable.

2. Match the tool to the task

Five steps run from an assistant on files people already open, to an assistant across organisation content, to assistants on approved content, to workflows and dashboards, and finally to predictive models. Each step needs more groundwork than the last, so climb only as far as the foundations allow. Most of the early value sits on the first three steps, where no new data engineering is needed.

Exhibit 1The five-step platform ladder
MOST EARLY VALUE SITS ON STEPS 1 TO 3 1 Assistant on files you open Draft, summarise, analyse data NOW 2 Assistant across organisation content Mail, files and meetings SELECTED ROLES 3 Assistants on approved content Policy and rules Q&A WEEKS, NO CODE 4 Workflows and dashboards Reminders and portfolio view MONTHS 5 Predictive models Early warning, pricing models YEAR 2 ON Each step needs more governance, skills and data groundwork than the last. Timings are indicative. Author's framework.

Author's framework. Timings are indicative and depend on groundwork.

3. Keep the choice portable

Single-model bets age badly. In Andreessen Horowitz’s January 2026 survey of enterprise CIOs, 81 percent used three or more model families, mainly because strengths differ by use case.5 Portability is also a cost control. McKinsey’s advice is to route each task to the lowest-cost model that meets the quality bar, which only works if applications are not hard-wired to one provider. Keep prompts, data and evaluation under the organisation’s control.

4. Treat agents as a rung, not a starting point

Gartner expects over 40 percent of agentic AI projects to be cancelled by the end of 2027, calling many “driven by hype and are often misapplied”. It also warns of agent washing, the rebranding of older chatbots as agents.6 Pursue agents where the value is clear, the content is approved and a person remains accountable for the outcome.

5. Lay three foundations and name the risks

Governance, skills and change, and groundwork come first. BCG found that nearly all future-built firms have deeply engaged leadership, against 8 percent of laggards, and that more than half run one enterprise-wide data model, against 4 percent of stagnating peers.7 The pattern is consistent: leaders who engage, and data that is consolidated, come before scale.

Exhibit 2Three foundations
SCALE ONLY ONCE THESE THREE ARE IN PLACE GOVERNANCE Who decides; rules for AI and data; risk control and assurance. IN PLACE BY QUARTER 1 Steering group, simple AI rules, register of uses, internal audit SKILLS AND CHANGE Staff who automate routine work and check AI output; change support so habits stick. IN PLACE BY QUARTER 1 Practical sessions, a champion per department, expert partners GROUNDWORK Know the systems, data, processes and gaps before building anything. IN PLACE BY QUARTER 1 A gap map of systems, data and skills; data plan agreed C-SUITE DEEPLY ENGAGED WITH AI BCG survey of 1,250 firms, 2025 Future-built nearly 100% Laggards 8% Bars drawn to scale. Pillars are the author's framework.

Author's framework. Statistic from BCG, The Widening AI Value Gap, 2025.

Name the risks early, and attach a first response to each.

Exhibit 3Five risks and first responses
RISK FIRST RESPONSE ! Unsanctioned use A usage policy and a register of AI uses ! Growth outpaces delivery Measure conversion; fix delivery first ! Fragmented data One consolidated view on core data ! Concentration and market risk Stress tests now, models later ! Technology ahead of groundwork Scale only on results Five risks, each with a first response. Author's framework.

Author's framework.

Platform strategy is a sequencing decision. The rung matters less than the groundwork beneath it.

What the evidence does not show

The McKinsey cost survey has only 75 qualified respondents. Gartner's figure is a prediction, not an outcome. The KPMG study was conducted by the University of Melbourne in collaboration with KPMG, which sells AI governance services.

Altman and Masiyiwa lead organisations that profit from wider AI use. Their statements are signals of direction, not evidence. Portability has costs too: running several models adds integration and governance work.

Sources

  1. Sam Altman, "Three Observations", 9 February 2025. blog.samaltman.com
  2. McKinsey & Company, "The cost of intelligence: How CIOs can manage AI demand at scale" (Enterprise AI FinOps survey, May 2026; 75 qualified respondents). mckinsey.com
  3. University of Melbourne and KPMG, "Trust, attitudes and use of artificial intelligence: A global study 2025", press release 9 May 2025 (survey of 48,000+ people, 47 countries, November 2024 to January 2025). kpmg.com
  4. Strive Masiyiwa, remarks at Unstoppable Africa 2025, UN General Assembly (UNGA80) side event, September 2025, reported by ITWeb Africa. itweb.africa
  5. Andreessen Horowitz, "Leaders, gainers and unexpected winners in the Enterprise AI arms race: Results from our third annual CIO Survey", 30 January 2026. a16z.news
  6. Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", 25 June 2025. gartner.com
  7. Boston Consulting Group, "The Widening AI Value Gap: Build for the Future 2025", September 2025 (survey of 1,250 firms). bcg.com

Read more on AI