The real gating factor in each vertical is which prerequisite is broken. For logistics: sensor and event-data quality. For healthcare: interop and consent. For professional services: billable-hour data hygiene. Each industry's top three blockers, the unlock that delivers payback inside 12 months, and the second-order project that follows.
Ask COOs in logistics, healthcare and professional services what stands between them and a first AI use case and you get one answer: the data is not ready. Ask what that means and the answers diverge: the logistics COO cannot trust a timestamp; the healthcare COO has careful records in systems that do not talk and may not be allowed to; the professional services COO has the timesheet, written to survive a client's review.
We have argued elsewhere for five readiness tiers and for picking the use case from the tier you are at. But the tier says how far you are from usable data, not what is keeping you there, and that is decided by your industry. Each vertical breaks one prerequisite in a characteristic way, and until it is fixed no platform moves the tier.
For ops leads in each: the top three blockers, the unlock that pays back inside 12 months, and the second-order project that follows.
Why is readiness different in every vertical?
Because the tiers describe the state of the data, and a generic assessment scores what you own rather than what is broken. It cannot see that logistics data comes from scanners, sensors and handovers, so the event stream does not describe what physically happened; that healthcare data is locked inside systems that do not interoperate and behind a consent basis written for treatment; or that the timesheet is written by a fee earner after the fact, a narrative rather than a record.
Fix the wrong prerequisite and the tier does not move: we have watched businesses connect every system and still be unable to say when a pallet left the depot. The integration was real; the readiness was not.
A readiness score tells you how far you are from usable data. Your vertical tells you what is keeping you there.
What does logistics have to fix first?
Sensor and event-data quality. Logistics generates more data per shipment than the other verticals, almost all of it machine-generated, which is why people assume it is clean. It is not clean, only voluminous, and the difference is the whole problem.
- Event timestamps that record the process, not the shipment. Scans fire in batches at shift end, the delivered event is keyed when the driver returns, and local times arrive without a zone. A model trained on it learns the depot's habits, not the network's behaviour.
- Sensors nobody has calibrated or owned. Temperature probes, door sensors and GPS pings arrive in mixed units from mixed vendors, with gaps and drift, and nobody has the authority to say what a cold-chain breach is. Alerts get silenced for noise, then ignored when they matter.
- Exceptions that live outside the systems. The delay, the refused delivery, the damaged pallet: the events you most want to predict are recorded as free text in email or on the phone, or not at all. The dataset holds every routine shipment and almost none that cost money.
The unlock is an event spine: one reconciled, timestamped sequence per shipment across transport, warehouse and carrier feeds, with exceptions captured as structured events. Scoped to one lane or depot, that is tier-three cleanup, not a platform. On top sits exception prediction and proactive customer notice, which pays back inside 12 months by removing the 'where is my shipment' work and cutting failed deliveries. The second-order project is carrier and lane performance on your own on-time record rather than the carrier's, which changes the rate negotiation.
What does healthcare have to fix first?
Interoperability and patient consent. Providers of 50 to 5,000 people, clinics, diagnostic groups and care operators, keep better records than most businesses. The records are unusable together, and sometimes for the purpose in mind.
- Systems that do not interoperate. The patient record, scheduling, the laboratory system and billing each hold their own version of the patient under their own identifier; the join is a person with several screens open. Without a canonical patient identity nothing that spans systems is possible.
- Consent captured for treatment, not for models. Data collected for treatment was collected for that purpose; tuning, training or prompting a model is a different one, and the basis for it is often undocumented. Take advice on the consent basis before any record leaves the system it was collected in.
- Clinical meaning locked in free text. Referral letters, discharge summaries, scanned PDFs and handwritten notes carry the meaning; the structured fields are sparse and inconsistently coded. Extraction is possible, but extracting into a dataset you cannot lawfully use is expensive noise.
The unlock is administrative. Referral and intake extraction with a clinician confirming the result, scheduling and no-show reduction, clinical coding and invoicing preparation, correspondence drafting: uses of data you already hold a clear basis for, inside the system it lives in. It pays back inside 12 months in administrative time, the largest cost an ops lead controls, where the consent question is simplest. The second-order project is a canonical patient identity across systems with consent designed in from the start, the door to clinical decision support.
What do professional services have to fix first?
Billable-hour data hygiene. A law firm, an accountancy practice or a consultancy has one dataset that describes the business, and it is the least trustworthy thing in the building.
- Timesheets that are a narrative, not a record. Hours entered at week or month end, rounded, moved to the matter with budget left, narrated to survive the client's review. That describes what could be billed, not what was done, and a model trained on it learns to bill rather than work.
- Client and matter identity that does not reconcile. The same client under several names across practice management, the CRM and finance; matters opened twice; work types coded differently by every partner. A benchmark across matters compares categories nobody agreed.
- Knowledge in documents nobody classified. Precedents, proposals and deliverables sit in a document management system organised by partner, with no metadata and duplicates in email. Retrieval over it returns whichever version someone last saved, with confidence.
The unlock is capture at the point of work. Calendar, email and document activity prompt the time entry, the narrative is drafted from it and confirmed by the fee earner, and one matter and work-type taxonomy is decided by someone with authority. That is hygiene, not surveillance. The use case on top is recovered time, the largest leak in any firm, and it pays back inside 12 months in hours that were worked and never billed. The second-order project is pricing: fixed-fee scoping from real effort data, then retrieval over documents that finally have metadata.
What to do next
Name the prerequisite before the use case. In logistics: can you say when a given pallet left the depot without phoning anyone? In healthcare: does the same patient have one identifier across your systems, and who can point to the consent basis for the use in mind? In professional services: what proportion of last month's hours were entered the day they were worked? The honest answer usually sits a tier below the licence.
Then scope the fix and the first use case as one piece of work, sized to pay back inside 12 months, owned by someone who can say what clean means. Scoped apart, the fix becomes a platform and the use case becomes a demo. That is a J Labs engagement: a 20 or 45-minute discovery call to place the tier and name the broken prerequisite; where it makes sense, a four-day Spec from €5,000 that puts both in one written brief you own outright; then a fixed-scope, fixed-price build; the average ships in eight weeks. We build on spec, with no product to sell.