Santosh Sabnis

A city’s stated policy may not prevent access that its contract leaves open.

In a post on license-plate-reader contracts, Sabnis points to Dayton, where federal agents made 7,100 immigration-related searches despite the city’s stated opposition. They say the contract did not prohibit those searches and call for a data-sovereignty audit before a council vote, along with enforceable terms for accuracy, access, and data destruction.

Read original

Data freshness and human oversight belong in the system’s design, not as afterthoughts.

Sabnis describes an agent routing 500 containers toward a disrupted distribution hub after reading a lakehouse snapshot that was ten hours old.

• Proposed safeguards include streaming updates and a freshness check that sends actions to a person when the data is more than five minutes old.
• They also call the cost of deploying AI and then paying people to supervise it the “Automation Tax.”

Read original

The post makes the stakes concrete with examples, including a chatbot that gave business owners false information about workers’ rights and food safety. Read alongside their technical and contractual proposals, the argument is not that software never helps, but that it cannot substitute for sound governance or reform of the systems around it.

Sabnis contrasts the spread of municipal AI chatbots with the smaller share of cities said to have governance policies for them. The post also points to workforce pressures and long civil-service hiring timelines, arguing that buying new software alone cannot solve these problems.

Read original

It answers a newcomer's first question—is this a critic of the technology or of how cities buy it?—and shows the detail behind the recommendation, from a specific accuracy threshold to a specific contract clause.

Sabnis opens with a measured failure: the LAPD documented a 32.3% false-positive rate on its license-plate-reader cameras, meaning one in three alerts was wrong.

• The argument: this is not a technology failure but a contract with no accuracy floor.
• They say they reviewed 82 termination records across 28 states and found the same pattern each time: councils voted on camera count, contract term and annual cost, while data ownership, post-termination access rights and arbitration jurisdiction sat in a separate vendor-written document that city staff approved under deadline pressure.
• Data sovereignty transferred at signing, not at termination.
• A Dayton example shows the consequence: federal agents ran 7,100 immigration-related searches against the city's license-plate database, even though stated city policy opposed it, because the contract did not prohibit it.
• The proposed correction has three parts: a non-waivable data sovereignty audit before any contract reaches a council vote; a 5% false-positive accuracy floor, independently verified each year, with automatic suspension if breached in any 90-day window; and three provisions covering certified destruction of data within 30 days of termination, home-state arbitration, and a localized data silo with named, council-approved access and 48-hour public notice of unauthorized access.

For brands planning a campaign

For a govtech company working on municipal AI procurement, this post offers a ready-made talk outline: the false-positive floor, the pre-vote audit and three contract provisions give a city audience concrete steps to act on.

Readers who disagree with Sabnis should start here.

Sabnis describes a municipal problem often mislabelled as digital transformation.

• Last year, 44% of U.S. cities put AI chatbots in front of resident requests, while fewer than 16% had a governance policy for them.
• Three pressures converge: a San Francisco structural deficit projected at $640 million for FY2027; 38% of local government employees expected to retire within five years, with under 7% of civil service workers under thirty; and hiring timelines such as New York City's civil service exam-to-start-date pipeline, which exceeds 18 months.
• The example at the heart of the argument is New York City's AI chatbot, which told business owners they could legally take employees' tips, dismiss workers who reported sexual harassment, and serve food contaminated by rodents.
• The chatbot delivered false statements with the same confidence as correct ones and remained running.
• Sabnis concludes that the real constraint is a nineteenth-century civil service legal framework operating against twenty-first-century service demand—and that no procurement contract bridges the gap.

For brands planning a campaign

Useful for any pitch that frames municipal AI as a modernization purchase: Sabnis offers a counter-frame and accountability numbers vendors should be prepared to answer. The chatbot failures also make reusable evidence for a governance-first talk.

Its implementation detail helps readers assess whether Sabnis is a commentator or a builder.

Sabnis reports $684 billion of enterprise AI investment in 2025, with $547 billion producing no measurable results—a pattern they say they have observed across three consecutive enterprise deployment cycles.

• The worked example: an autonomous agent running on 10-second cycles sits over a lakehouse that processes batches every 12 to 24 hours.
• At 8:00 AM, a weather event blocks a distribution hub; at 10:00 AM, the agent reads the midnight snapshot, concludes all hubs are clear and routes 500 containers into the disruption.
• By noon, the supply chain is congested.
• The agent behaved exactly as designed, using data ten hours old.
• Sabnis calls the resulting overhead the Automation Tax: organizations use AI to cut labour costs, then hire more expensive staff to supervise its decisions.
• They cite 72% of enterprise AI infrastructure deployments failing or stalling because the data layer cannot keep pace with the model.
• The proposed correction is a three-stage data layer rebuild: log-based change data capture attached to transaction logs; change events routed through a streaming platform into an open table-format lakehouse; and a timestamp-freshness circuit breaker that hands an action to a human when the data is more than five minutes old.

For brands planning a campaign

The freshness circuit breaker and the change-data-capture-to-lakehouse sequence offer a vendor-neutral technical agenda that complements the procurement argument. The Automation Tax framing gives business audiences an accessible summary.

It says data sovereignty transfers to the vendor at signing, not termination, and cites Dayton, where federal agents ran 7,100 immigration-related searches against a city database despite a stated local policy opposing it.

For brands planning a campaign

A govtech company can sell governance as part of the product, not an afterthought. A talk or workshop should show how contract language, data ownership, and audit rights determine whether a city can control a system after signing.

Examples: Source 1

It calls this a contract with no accuracy floor and proposes a 5% false-positive accuracy floor, independently verified each year, with automatic suspension if breached in any 90-day window.

Write performance thresholds into the contract, verify them independently, and attach automatic consequences for missing them.

For brands planning a campaign

A vendor that meets a measurable accuracy threshold has a concrete differentiator. Campaign material can move beyond capability claims to offer contractual accuracy commitments, third-party verification, and public performance reporting as part of a municipal bid.

Examples: Source 1

It describes an agent deployed over a data lakehouse that processes batch cycles every 12 to 24 hours while the agent operates at 10-second intervals.

• In the example, an 8:00 AM weather event obstructs a distribution hub.
• By 10:00 AM, the agent reads the midnight snapshot and routes 500 containers into the disruption zone; by noon, the supply chain is congested.
• The post says 72% of enterprise AI infrastructure deployments are failing or stalling because the data layer cannot keep pace with the model.

Replace batch pipelines with log-based change data capture, route change events through Kafka into a medallion lakehouse, and embed a timestamp-freshness circuit breaker that sends actions to a human when data is more than five minutes old.

For brands planning a campaign

A data infrastructure company can lead with freshness rather than model performance. The practical message: replace batch pipelines with change data capture and route current operational state to agents on demand.

Examples: Source 1

It calls this the Automation Tax and says the cost remains unpriced even when the AI is operational.

The post says the AI did not malfunction; it executed as designed on data that was 10 hours old. That puts the failure in the data layer and oversight model, not the AI model.

For brands planning a campaign

Buyers evaluating AI need a total-cost model that includes oversight labor, exception handling, and stale-data incidents. A campaign can offer a workshop or assessment that prices the Automation Tax before a larger deployment.

Examples: Source 1

It calls this an accountability vacuum and cites New York City’s AI chatbot, which incorrectly told business owners they could legally take employees’ tips, dismiss workers who reported sexual harassment, and serve food contaminated by rodents, while remaining operational.

For brands planning a campaign

A govtech company selling municipal AI should pair deployment with a governance policy, escalation paths, and public correction procedures. The campaign angle is not simply “install a chatbot,” but “adopt a system you can govern.”

Examples: Source 1

It argues that cities, vendors, and reform advocates wrongly believe municipal government’s problems stem from software. The real problem, it says, is a nineteenth-century civil service legal framework operating in a twenty-first-century service-demand environment.

The post says no procurement contract can bridge that divide, setting a clear limit on what software vendors can promise.

For brands planning a campaign

A vendor that only sells software will miss the procurement and workforce constraints that determine adoption. A credible campaign should address civil service rules, hiring pipelines, and budget pressure alongside the technology.

Examples: Source 1

The ALPR post uses 82 termination records across 28 states and three structural moves.

• The AI infrastructure post lays out three stages for a real-time data correction.
• The city AI post presents three converging failures and argues that the problem is not software.

For brands planning a campaign

A brand can commission a teardown, webinar, or brief in this format: one documented failure, its contract or infrastructure cause, and three actionable corrections. It suits policy audiences and technical buyers who need a decision framework, not a product pitch.

Examples: Source 1 · Source 2 · Source 3

Mark Anthony F. analyzes total cost of ownership, vendor lock-in, and the gap between the software a city bought and the workarounds staff actually use; Santosh Sabnis argues that contract terms determine data ownership, access, and accountability.

Mark Anthony F. teaches through plain-language checklists and direct questions to municipality staff, focusing on measuring current process costs and asking vendors about data export and implementation. Santosh Sabnis focuses more on contract law, data sovereignty, and AI or data infrastructure, structuring posts as documented failures followed by specific corrective measures.

Read original · Original 2

Jorge Cardenas frames broadband as a fourth utility alongside roads, water, and electricity, and describes Brownsville’s smart-city buildout; Santosh Sabnis treats data and AI governance as infrastructure decisions with public consequences.

Jorge Cardenas’s posts are progress narratives advocating broadband investment, listing city achievements and economic-development benefits. Santosh Sabnis focuses on failure analysis and enforceable contract terms, with more technical detail on AI data pipelines and municipal governance gaps.

Read original · Original 2

Chad Smeltzer contrasts manual manhole inspections with digital capture and asks why the old process persists; Santosh Sabnis asks why cities sign technology contracts without governance terms.

Chad Smeltzer’s examples come from physical infrastructure, wastewater, and civil engineering business models, and often use personal anecdotes and price comparisons. Santosh Sabnis writes about software and AI systems, contract provisions, and municipal policy, with less emphasis on physical asset inspection.

Read original · Original 2

The posts offer a reusable teaching sequence: measure the current process cost per month in staff time before signing, then ask three exit questions about data export, what happens if the vendor is acquired, and who does implementation. For a govtech campaign on total cost of ownership or vendor management, this is closer to the buyer's decision than Sabnis's material.

The posts focus on procurement economics and contract exit terms, not who may query a database, what accuracy threshold a contract must meet, or how post-termination access is governed.

• A campaign on data sovereignty, accuracy floors, or AI-specific oversight would lose its central argument if it relied on Mark Anthony F. instead of Sabnis.
• The tradeoff is operational procurement framing in place of governance inside the technology contract.

Examples: Source 1 · Source 2

The posts argue that broadband is now a fourth core infrastructure asset alongside roads, water, and electricity, and report a specific municipal build: modernized and new data centers, roughly 100 miles of city-owned fiber, a private 5G network, and an AI effort for public safety and infrastructure insight.

• They frame the city as having moved from being ranked worst-connected to a model for urban connectivity.
• This is a strong fit for a campaign selling the case for investment or showing municipal leaders what a completed modernization program looks like from the inside.

The posts make the case for connectivity and report outcomes; they do not analyze procurement language, failure modes, or the governance gap between approving a system and controlling it.

• For a campaign whose buyer task is scrutinizing an AI or surveillance contract before a vote, this creator supplies the why-invest half and the credibility of a named municipal build, but not the what-to-require half.
• Pairing is more effective than substitution if the campaign covers both.

Examples: Source 1 · Source 2

The posts include a costed inspection comparison — roughly $500 per manhole under a manual method against about $125 using a single-pass 360-degree optical scan, a difference of about $37,500 on 100 manholes before engineering and reporting hours — and argue that firms move upstream by combining technology with engineering judgment and proprietary process instead of chasing bids and cutting fees. That directly serves a campaign about value capture and avoided cost.

These posts focus on inspection technology, engineering economics, and firm positioning.

• They do not cover data ownership, access rights, accuracy requirements, or civil-service constraints, so they cannot carry a governance-first AI procurement argument.
• They also include a direct promotional link to a sewer management platform, giving them a product-selling tone that an educational campaign would need to manage.

Examples: Source 1 · Source 2