Disruptive and sustaining AI in logistics: telling the two apart before you fund them

Stage 1 is not ignorance of the distinction. Most leadership teams can define disruptive innovation on request, and many can name the canonical examples. What is missing is any point at which the definition touches a decision. A linehaul consolidation model, a dock-scheduling model and a speculative autonomous-yard trial arrive in the same slide pack, compete for the same line, and get judged on the same criteria — which in practice are enthusiasm, vendor credibility and how well the story survives a board slide.

The tell is the business-case format. At stage 1 there is one template and it asks every initiative for a payback period. A sustaining initiative can answer that honestly: it improves a number the operation already tracks, on volume it already handles, and the arithmetic is checkable. A probe cannot. Its output is information about whether a market exists, not a return. So it either invents a payback line or it loses to the initiative that can produce one. Both outcomes are damaging, and both are caused by the form rather than by anyone's judgement.

This is a cheap stage to leave and an expensive one to occupy, because it starves both ends at once. The sustaining work is under-funded relative to its near-certain return, because it competes against better stories. The genuinely speculative work is over-specified, because it has to promise a return nobody can know. The organisation ends up with a portfolio optimised for narrative quality, which is not a property that shows up in cost per shipment.

Stage 2 is where most competent logistics operators sit, and it is a legitimate place to be. Linehaul consolidation, dwell prediction, appointment scheduling, tender-acceptance scoring, carrier performance models and document extraction are all sustaining innovations. They all pay. They all improve a metric the operator's existing customers already buy on, delivered into the TMS or WMS where the decision actually happens. Judged on realised value rather than on column inches, this is where nearly all logistics AI value has come from.

The blind spot is structural rather than intellectual. The measurement system points exclusively at the freight you serve. Cost per shipment, OTIF, dock dwell, picks per labour hour and empty miles are all computed over accepted, executed volume. That means the requests you declined, the quotes you lost because the price took four hours, and the shippers who never called because you do not quote in seconds are invisible by construction. An operator can improve every number in the pack for three years while its addressable market quietly narrows, and no report will contain a row that says so.

Time at stage 2 is not neutral either, because the vocabulary hardens. The organisation learns that "AI" means efficiency — which is true and incomplete — and after a few cycles an "AI project" is understood internally to be a cost-reduction project. A proposal that is not a cost-reduction project then has nowhere to go: it cannot be written on the form, it cannot be argued in the forum, and the person who would have argued it stops trying.

Stage 3 begins the moment the label does work. A sustaining initiative is funded from the operating budget, owned by the operations leader whose KPI it moves, and judged against a holdout — a lane set, door bank or shift left on the previous process so the improvement is attributable rather than asserted. A disruptive probe is funded from a separate capped pot, owned by whoever would have to build the resulting business, and judged on whether it answered a stated question by a stated date. Same organisation, two different definitions of success, and neither one borrowed from the other.

Two artefacts appear at this stage, and they are what a reviewer should ask for. The first is a portfolio list with a label on every row and a different metric type by label. The second is a probe register with a kill rule per row. The kill rule is the hard part and the one most often fudged. "We will stop if it is not working" is not a rule; it is a sentence. "We stop if fewer than forty shippers accept an instant quote on this lane set within eight weeks" is a rule, because it can be false.

What stage 3 still does not have is any independent signal about whether a disruptive thesis is correct. The classification is applied to whatever ideas the operator happened to have — which came from vendors, conferences and the trade press. That is a very large improvement on stage 2, and it is still a portfolio of opinions with better paperwork around it. The next move is not more ideas; it is instrumentation.

At stage 4 the operator stops depending on the trade press to notice that the market has moved. The signals are in data it already owns: the decline register, the channel mix on inbound bookings, the time-to-firm-price distribution, the win rate banded by shipper size, and the identity of whoever won the last twenty lost tenders. External market series — spot-to-contract spread, tender rejection — are joined in as context that tells you whether a loss was cyclical or structural, not as the primary instrument. Own-data first is not a stylistic preference; published indices lag the shipper behaviour they summarise.

The discipline that separates this from a dashboard is the threshold. Each signal has a number attached, agreed before the first reading, and crossing it triggers a specific action with a named owner and a date. Without that, the panel becomes a reading exercise, and it will be read charitably — because everyone on the call has an interest in the current strategy being correct, and a chart with no threshold can always be described as noise for one more quarter.

The second thing stage 4 buys is calibration on speed, which is where most disruption arguments actually go wrong. In freight, structural change has consistently been slower than its advocates predicted and faster than incumbents' planning cycles assumed. Instrumented operators stop arguing about whether something will happen and start arguing about the rate — and the rate is a question their own quote log can settle, which makes the argument shorter and considerably less enjoyable for everyone.

Stage 5 is much narrower than "we do disruptive innovation". It is a register of two to four options, each with a stated thesis in one sentence, a cost of holding, a cost of exercising, a review date, and the signal that would trigger exercise. In freight the realistic entries are unglamorous: a productised instant-quote channel for weight breaks the network already handles; a driver-out corridor contingent on a certification path actually opening; a data product built from flows the operator already moves. Options that read like press releases tend not to survive their first pricing.

The discipline that keeps the stage honest is retirement. Options expire, and an option held past its date without a decision has quietly become a programme — one that consumes operating attention while reporting into a forum designed for probes. Reviewing the register on a calendar, and retiring something in most cycles, is what stops the portfolio silting up. The health check is simple: ask what was retired last time. A register that only ever grows is a project list wearing an option's vocabulary.

The binding constraint at stage 5 is not imagination. Every genuinely disruptive path in logistics runs into an asset, a certification or a piece of physics. Driver-out linehaul runs into automated-driving rules and a narrow operational design domain. Drone delivery runs into airspace approvals, payload and energy density. Port and terminal automation runs into capital and labour agreements. Unattended commercial commitment runs into liability and audit evidence. An option that has not written down its bound is not an option, it is a wish — and the register should state the bound and what would have to change for it to move.