The three modes of AI strategy development, and when to use each

Three modes, one strategy

Every AI strategy we have seen work in practice combines three modes of development. The failures usually pick one and stop, then wonder why the strategy sits on a shelf or never gets past a slide deck.

Mode 1: Top-down

A CEO, CIO, or board sets the direction: which markets AI must serve, what risk appetite is acceptable, and what the organisation is unwilling to do. This mode is fast, unambiguous, and essential for setting guardrails. A single afternoon with the right five people can produce a page of direction that would take a working group three months to agree by consensus.

Its weakness is detail. Executives rarely know which specific processes are ripe for automation, and top-down-only strategies tend to over-index on visible, headline-friendly use cases (a customer-facing chatbot, a flagship "AI transformation" launch) while ignoring the unglamorous back-office process that actually wastes the most hours. We have seen a board mandate an AI copilot for sales, generated largely because a competitor announced one, while the finance team's month-end reconciliation, twelve people, three days, entirely manual, went untouched because nobody at board level had ever seen it done.

Mode 2: Small leadership groups

A cross-functional group, typically eight to twelve people spanning operations, technology, risk, and a business line or two, translates direction into a portfolio. They pick target outcomes, agree the maturity investments, and shape the operating model: who owns data quality, who signs off on a model going into production, what "done" looks like for a pilot.

This mode gets the trade-offs right because the people making them own the consequences. A head of risk in the room will kill a plausible-sounding use case in minutes if it touches regulated data without controls; that same objection raised anonymously in a feedback form gets ignored for months. Its weakness is scope: a small group cannot see every opportunity or blocker inside a large organisation, and it will default to the processes its members personally touch. Twelve people cannot enumerate two thousand employees' worth of daily friction.

Mode 3: Individual completers and crowdsourcing

Ideas, edge cases, and delivery detail come from the people doing the work. Individual contributors submit opportunities, flag risks, and often complete significant chunks of the strategy artefacts, canvas contributions, process maps, project briefs, themselves. A claims handler knows exactly which three fields in a form cause 80% of rework; nobody above them does.

This mode surfaces ground truth that no leadership group or board can reconstruct from memory. Its weakness is coherence: without the first two modes, crowdsourced input becomes a backlog nobody can prioritise, a spreadsheet of two hundred ideas with no owner, no scoring, and no relationship to the organisation's actual risk appetite or budget. We have seen this exact spreadsheet. It sat untouched for a year because nobody had the authority, or the criteria, to say yes to any of it.

How to combine them

1. Top-down first. Set direction, guardrails, and non-negotiables in writing. Keep it short, a page, not a deck. State what AI must not be used for as clearly as what it should be used for.
2. Leadership group next. Translate direction into a portfolio, a maturity plan, and an operating model. Publish it somewhere the wider organisation can actually find it, not a board pack nobody below director level ever sees.
3. Open the doors. Invite the wider organisation to contribute ideas, refine canvas entries, and pick up briefs, inside the frame the first two modes established. Give people a channel and a rough sense of what "good" looks like, or you get noise instead of signal.
4. Loop. Feed what you learn from mode three back into the leadership group's next review, and occasionally back to the top for direction changes. A strategy that never gets revisited against what mode three surfaces is a strategy frozen at day one.

Choosing your starting point

If your organisation has no stated AI direction, start top-down; nothing else has anywhere to attach to. If the direction is clear but nothing is shipping, convene a leadership group; direction without a portfolio is just an aspiration. If both exist but momentum is stalling, open the process up; you likely have a coherence problem, not an ideas problem. Most organisations need to run all three concurrently within a year, not sequentially and then stop.

Watch for: a leadership group that meets once and calls itself done; a crowdsourcing channel with no visible triage, which trains people to stop submitting within a fortnight; and top-down direction so vague ("be AI-first") that mode two has nothing to translate.

Next step: pick whichever mode is currently missing entirely from your organisation and give it an owner and a date this month, not a plan for next quarter.

Where Florence helps

Florence is built around this loop rather than any single mode. Leaders shape direction directly in the Organisation profile and Strategy step. A small group can review, score, and prioritise in Projects without needing a separate offline process. And any member of the organisation can submit ideas into a governance queue that feeds the Canvas and Insights steps, so mode three's ground truth reaches the people making mode two's trade-offs, instead of dying in a spreadsheet.