A gate is a condition, not a date

I have read a lot of AI strategies in the past two years. Most of them are procurement lists with a vision statement stapled to the front.

That is not a cheap shot. It is a diagnosis, and it has three tests attached that any executive can run on their own document tonight.

Test one: name something it forbids

A strategy that permits everything has decided nothing.

Richard Rumelt's distinction is the useful one here. Strategy diagnoses a specific situation, chooses a guiding policy in response, and commits to coherent action. What he calls bad strategy substitutes ambition for diagnosis and lists goals instead of choosing between them.

So: which functions in your organisation are explicitly waiting? Which data never leaves the building regardless of the business case? Which decisions may a machine never take alone, not because of current capability but as a matter of policy?

If your document contains no such statements, it is a budget with adjectives. When I ask this in a room, the silence is usually longer than the answer, and the silence is the finding.

Test two: would it work equally well for your competitor?

Take your AI strategy and mentally hand it to a direct rival. Would it suit them?

If it would suit any bank, any insurer, any department in this country, then it diagnosed nothing specific about you. It is an industry trend you have written down and costed. Useful as an orientation document, worthless as a basis for choosing between two things you cannot both do.

Real strategy is uncomfortable to read precisely because it is legible as a bet. It says: our claims data is unusually clean and our underwriting judgement is unusually scarce, so we automate the first and protect the second, and we accept that this makes us slower than competitors in an area where we have decided not to compete.

Nobody writes that down, because writing it down means someone can be wrong about it later.

Test three: what does it sequence?

Somebody has to be first and somebody has to wait. That ordering is the only part of the document that will still matter in two years.

Support before legal. Claims before underwriting. High-volume and review-tolerant before high-stakes and irreversible. The sequencing decision encodes your actual risk appetite far more honestly than the risk appetite section does.

For public sector organisations this test carries more weight than the other two, because procurement cycles and budget votes mean you cannot buy your way out of the ordering problem. The sequence is the strategy. There is no version where everything starts at once.

Now the roadmap, which is a different failure

Strategy is about choice. Roadmaps are about time, and they fail in a distinct and more avoidable way.

Almost every enterprise AI roadmap I have seen has columns. The early ones say sensible things about assistants and embedded agents. The later ones, somewhere in the 2027 to 2029 band, promise an agent ecosystem, cross-organisational automation, agents transacting with counterparties.

Those later columns depend on things nobody in your organisation controls. Whether other firms' agents can prove who they are. Who is liable when two agents from different companies transact and one is wrong. Whether your insurer will price that exposure. In South Africa specifically, sector AI regulatory instruments are expected from 2027/28, which is the same window most of those columns occupy.

No standards body has published a date for cross-organisational agent identity. No insurer has published terms. You have.

The containerisation lesson

Malcom McLean's first container ship sailed in 1956, and the efficiency case was obvious to anyone who watched it dock. The industry could not capture that efficiency for well over a decade, because the box had to be standardised first, through the ISO process in the 1960s, before world trade reshaped itself around it in the 1970s.

The bottleneck was never the crane. It was the agreement about the size of the box.

Cross-firm agent transactions are waiting for their ISO container. Anyone putting a date on that in their own planning document is forecasting the behaviour of standards bodies, insurers and regulators, which is not a forecast they have any basis for making.

The replacement: conditions

Here is the fix, and it is small enough to do in an afternoon.

Delete every date from the later columns of your AI roadmap and replace it with the condition that would trigger it.

"2028" becomes "when counterparty agent identity standardises and liability is allocated in contract."

That single substitution changes three things.

It becomes testable every quarter. You can ask whether the condition is closer. A date is testable exactly once, on the date, when it is too late to do anything.

It becomes honest. In 2028 you are explaining a condition that has not yet been met, rather than a commitment you missed. Those are different conversations and only one of them damages you.

It becomes actionable. Once the condition is named, you can ask who is working on it and whether you should be. Several of the conditions in your roadmap are things your industry body could influence if anyone showed up.

What a gate actually is

The word gate is used loosely in delivery methodology, usually to mean a review meeting. That is not what I mean.

A gate is a condition that must be demonstrably met before the next stage consumes budget, with a named person empowered to say no, and at least one instance in your organisation's history where somebody actually said it.

If you have never stopped a project at a gate, you do not have gates. You have review meetings with a colour-coded slide. Ask your delivery leadership to name a project that was killed at a gate. If nobody can, that is the answer.

This is also why I am not arguing against failure. High failure tolerance is essential to getting anywhere with AI, and I have argued that publicly for years. But high failure tolerance is not the same as high failure cost. Gates exist so you fail in week four for the price of a sprint, instead of in quarter four for the price of a programme.

The short version

A list of use cases is not a strategy. It is a wish list with a budget attached.

A date you do not control is not a commitment. It is an excuse scheduled in advance.

And if you cannot replace the dates in your roadmap with conditions, you do not have a strategy. You have a schedule.