Coverage · Ten sectors

Where failure is expensive, and why it is expensive differently

Generic resilience advice fails because the sectors do not fail the same way. A retailer’s worst day is on the calendar. A utility’s worst failure produces no alert at all. A defense estate cannot use the tooling that would find either. Below: the challenge, the cost of leaving it, and where AI helps or hurts — one row per sector.

Healthcare & Life Sciences

~$4.9T · 17.6% of GDP
The challenge
Multi-hop eligibility and adjudication paths cross organisational boundaries, so a regression three hops deep surfaces as a provider-facing timeout with no obvious owner. Validated systems cannot be modified to add instrumentation without triggering revalidation.
Cost of inaction
Downtime is clinical — paper charting and ambulance diversion, not a degraded experience. An availability incident is simultaneously a potential HIPAA reportable event, so the investigation has a regulator as its second audience.
AI upside
Parallel diagnosis compresses the phase where clinicians are waiting. Postmortems generated as evidence satisfy an obligation you already carry.
AI pitfall
Automated remediation against a clinical system on a weakly-supported conclusion. PHI reaching an agent’s context window moves the governance boundary without anyone deciding to.
Issue 03 · When Downtime Is Clinical →

Financial Services & Insurance

~$4.5T annual revenue
The challenge
A trace that starts in a modern service stops at the core-banking boundary. Half the evidence sits in a system the responder cannot query. Overnight batch failures are discovered after the settlement window has closed.
Cost of inaction
Supervised resilience means an outage is a reportable event with a deadline. Reconstructing a defensible timeline from chat logs is a week of senior time you do not have, against a clock you do not control.
AI upside
Correlation across the seam; evidence generated rather than reconstructed, which is exactly the artefact an examiner asks for.
AI pitfall
Explainability. “The model decided” is not an account a supervisor accepts, and most implementations discard the probes once a conclusion is reached.
Issue 04 · The Seam Between Two Eras →

Retail & Consumer Goods

~$7.2T retail sales
The challenge
Timeouts, retry counts and pool sizes set years ago against traffic that no longer exists. A retry policy tuned for a fast dependency becomes an amplifier when that dependency slows — the estate attacks itself at peak.
Cost of inaction
A year of engineering is graded across five days that cannot move, and the competitor is one tab away. A minute lost in late November costs a multiple of the same minute in March.
AI upside
Configuration derived from observed peak traffic rather than convention. Gaps found in the repository months ahead, at leisure.
AI pitfall
Bot traffic engineered to look like demand, so the first instinct is to scale — paying to serve the adversary faster. Model endpoints in the checkout path add latency nobody is watching.
Issue 05 · The Year Graded in Five Days →

Technology, Software & Cloud

~$2.4T sector revenue
The challenge
The baseline is already excellent, so what remains is rare, novel and cross-service. The cost has moved from downtime to senior attention — six engineers on a bridge, working hypotheses in sequence.
Cost of inaction
SLA credits are contractual and immediate. The public postmortem becomes an industry artefact that follows sales cycles for quarters.
AI upside
Parallel probing replaces the serial hypothesis loop. This is a throughput argument, and it is the honest one for this sector.
AI pitfall
Model behaviour changes without a version change — availability monitoring stays green while quality degrades. Agent systems get more expensive the worse things get.
Issue 06 · Selling Reliability You Must Also Have →

Energy & Utilities

~$1.5T · ~$500B utility revenue
The challenge
The characteristic failure is an absence, not an error. A feed stops, a historian stops writing, a collector silently drops reads. Threshold monitoring detects too much and is structurally incapable of detecting too little.
Cost of inaction
NERC CIP makes reliability federally enforceable with penalties attached. Silent data gaps are discovered in billing reconciliation or an audit, months later, when the evidence is gone.
AI upside
Baseline-relative detection surfaces a class of event that quietly disappeared — the signature threshold alerting cannot see.
AI pitfall
Automated action against systems with physical consequence. And an intrusion that manifests as slightly fewer events hides inside the silent failures the estate already tolerates.
Issue 07 · Reliability as a Federal Obligation →

Manufacturing & Industrials

~$2.9T output · 10% of GDP
The challenge
Recovery is physical. Restarting a process does not restart a furnace or a paint line, and work in progress may be scrapped. Plant culture and connected-product culture have opposite tempos under one IT function.
Cost of inaction
The clearest per-minute cost in the economy, with no partial degradation. Seasonal windows — planting, harvest, launch — cannot be compensated later.
AI upside
Resilience gaps raised as scheduled maintenance rather than discovered as incidents. Change correlation anchored to onset, which matters at high deploy velocity.
AI pitfall
OTA software is a recall-class event when wrong, not a rollback. Ransomware targets this sector precisely because interruption tolerance is the lowest in the economy.
Issue 08 · The Line Either Runs Or It Does Not →

Transportation & Logistics

~$1.9T · 8% of GDP
The challenge
Failures compound. A forty-minute sort delay displaces volume into the next cycle, misplaces crew and equipment, and runs for days. Incidents cluster at operationally significant hours and are investigated the following morning.
Cost of inaction
The recovery — expedite, reposition, overtime, credits — costs a multiple of the originating fault. Crew scheduling failures become hours-of-service violations, not delays.
AI upside
Halving diagnosis time does not halve the cost; it removes an entire branch of the cascade. The return is superlinear here in a way it is not elsewhere.
AI pitfall
Optimisation removes slack by design, so the network becomes less able to absorb a fault exactly as it becomes more dependent on software being right.
Issue 09 · Failures That Compound →

Telecommunications & Media

~$1.1T combined revenue
The challenge
Two opposite pressures in merged companies. Telecom carries statutory availability with FCC reporting; media carries an unrepeatable moment. The 5G core is moving to cloud-native functions faster than the operational practice around them.
Cost of inaction
911 availability is law, not an SLA. A live-event failure is maximally visible, completely unrecoverable, and narrated publicly while it happens. Provisioning faults convert directly into truck rolls.
AI upside
Sub-minute evaluation and correlation, which is the only cadence that matters when the response window is the event itself.
AI pitfall
Personalisation and ad-decisioning models inserted between the viewer and the stream — dependencies with variable latency in the delivery path.
Issue 10 · Statutory Uptime, Unrecoverable Moments →

Real Estate & Construction

~$2.5T construction · ~17% of GDP
The challenge
The least digitally mature sector here, with the largest individual transactions. Rate locks expire, closings have dates, milestones have liquidated damages. Field systems work offline and reconcile late.
Cost of inaction
One transaction failing at the wrong moment can exceed the annual IT budget. Low volume does not mean low stakes — it means no other transaction averages the loss away.
AI upside
Automated investigation substitutes for headcount that was never going to be hired. With lean IT the constraint is investigation capacity, not detection.
AI pitfall
Wire fraud and business email compromise have become cheap to produce convincingly, against organisations moving bank-sized sums without bank-sized security operations.
Issue 11 · Small Estates, Very Large Transactions →

Government, Defense & Aerospace

~$850B defense · ~$6.8T outlays
The challenge
Classification creates an observability gap that procurement cannot close: SaaS requires egress, and egress is what is not permitted. Integrators operate estates they did not design with limited authority to change them.
Cost of inaction
Failure is political — a hearing, an IG report, or citizens unable to access benefits. Fielded systems report hours late, so tooling anchored to “recent” cannot investigate them at all.
AI upside
Self-hosted autonomy closes a gap that exists for architectural reasons rather than lack of will. Evidence retained for oversight rather than summarised away.
AI pitfall
Adoption pressure and oversight obligation arriving together. An automated action that cannot be reconstructed afterwards is a compliance failure as well as an engineering one.
Issue 12 · Inside the Boundary →

A note on the figures

Sector scale is public. Financial impact is modelled.

Sector sizes are drawn from public reporting — BEA, CMS, Census retail trade, company filings — and rounded to establish scale rather than to be quoted as precise. Every dollar impact in our analysis is the output of a published model applied to public figures: useful for sizing a conversation, worthless as a promise. A real number requires a real baseline from your own incident record, which is where an engagement starts.