Signoff · market brief · Flip the Switch Hackathon, Nov 2026

Agents can watch the data. People still have to own the decision.

Signoff is a pattern, shown on Citi Bike rebalancing. Agents detect data incidents and prepare the fix in the team's chat. A named owner signs off in the same thread. The decisions that depend on the data are gated on the outcome. Switch supplies the room, the roles and the sign-off gate.

The problem, in numbers

47 · 9 · 27

Citi Bike stations with 1970 timestamps, over capacity, and silent for a day, on our first pull of the live feed

Signoff, 7 Nov 2026

31%

of revenue affected by data quality issues, on average; ~15 h to resolve each incident

Monte Carlo survey, 2023

75%

of data teams take 4+ hours just to detect an incident

Monte Carlo survey, 2022

92%

of high-stakes AI projects hit data cascades

Sambasivan et al., CHI 2021

$12.9M

average yearly cost of poor data quality per organisation

Gartner, reference-customer survey

40%+

of agentic AI projects expected to be cancelled by end-2027 (costs, value, risk controls)

Gartner, June 2025

Who feels it

Any team that acts on operational data within minutes, not days: micromobility and fleet operations (rebalancing, charging, dispatch), logistics (slotting, routing), fintech operations (reconciliation, payouts), retail (pricing, inventory), energy and utilities (outage and load data). The pattern is always the same: a feed or table goes bad, an alert reaches an engineer's channel, and the people deciding keep deciding on it.

Citi Bike alone ran 46.6 million rides in 2025 across ~2,500 stations. Its rebalancing optimisation was published at AAAI 2015 and put into daily use. GBFS, the feed format, is published by 1,000+ shared-mobility systems worldwide.

Market

Segment	Size and growth	Source
Data observability	About $2–3 B in 2025, mid-teens CAGR in most forecasts (range across publishers: $1.7–3.5 B, 11–21%)	Business Research Co., SNS Insider, Spherical Insights, Market Research Future
AI agents in the workplace	By 2028, Gartner expects 15% of day-to-day work decisions to be made autonomously by agents (from 0% in 2024), and 40%+ of today's agentic projects to be cancelled first	Gartner, 2025
Regulation	EU AI Act Article 14 makes human oversight (understand, override, stop) a design requirement for high-risk systems, phasing in from 2026–2028	Regulation (EU) 2024/1689

Market-size figures come from commercial research publishers with different scopes. Treat them as an order of magnitude, not a forecast.

Landscape

Category	Examples	What they do well	What's missing
Data observability	Monte Carlo, Soda / Sodabase, Anomalo, Bigeye, Great Expectations	Detect stale, empty and wrong data; alert	The alert stops at an engineer. No decision step, and nothing gates the consumers
Incident management	PagerDuty, incident.io, Rootly	Coordinate people in Slack around an incident, with postmortems	Built for service outages. Agents don't propose data remediations, and nothing gates downstream answers
Chat assistants	Slack AI, Microsoft Copilot agents	Answer questions in chat	Answer from whatever data they're given, with no trust state and no owner sign-off
Signoff on Switch	this project	Agents detect, propose with a preview, verify and write up. A named owner signs off, enforced by Switch. Downstream answers are gated on the result	One data owner and one feed today (see roadmap)

Research it rests on

Idea	Where it shows up in Signoff	Source
Levels of automation: the machine suggests, the human approves	The fixer proposes and executes only on sign-off (Sheridan's level 5); detection and verification are automatic	Parasuraman, Sheridan & Wickens, IEEE SMC-A 2000
Appropriate reliance: trust should match reliability	The analyst says "untrusted", "caveated" or "trusted", with reasons, instead of always sounding sure	Lee & See, Human Factors 2004
Human-AI guidelines: explain why, support correction, scope by context	Every proposal shows why and what changes; reject offers the alternative; caveats name the incident	Amershi et al., CHI 2019
Data validation as code	A written, tested contract; rejected rows quarantined with a reason; constraints repeated in the database	Breck et al., MLSys 2019; Schelter et al., PVLDB 2018
Data cascades	Bad data is stopped before the dispatch decision, not discovered after the truck has left	Sambasivan et al., CHI 2021

How it could make money

Open core on Switch. The rooms, roles and agents are free and self-hosted. A paid tier adds connectors (warehouses, dbt, Snowflake), two-person sign-off, rule expiry and audit export.
Per data owner, not per seat. Value scales with the number of datasets that have a named owner and a gate.
Wedge: operations teams in mobility and logistics, where a wrong decision has a visible cost (a truck, a crew, a missed delivery) within the hour.

The bet: the next data-quality product isn't a better alert. It puts the incident, the evidence, the decision and the person who decides in the room where the work happens, and makes every consumer of the data respect the outcome.

Sources

Monte Carlo, State of Data Quality 2023 and the 2022 survey · N. Sambasivan et al., Data Cascades in High-Stakes AI, CHI 2021 · Gartner, agentic AI projects, June 2025; cost of poor data quality ($12.9 M, reference-customer survey) · EU AI Act, Article 14 · E. O'Mahony & D. Shmoys, Data Analysis and Optimization for (Citi)Bike Sharing, AAAI 2015 · Citi Bike and monthly operating reports · MobilityData, GBFS · R. Parasuraman, T. Sheridan, C. Wickens, Types and levels of human interaction with automation, 2000 · J. Lee & K. See, Trust in Automation, 2004 · S. Amershi et al., Guidelines for Human-AI Interaction, CHI 2019 · E. Breck et al., Data Validation for Machine Learning, MLSys 2019 · S. Schelter et al., Automating Large-Scale Data Quality Verification, PVLDB 2018 · Data observability market: SNS Insider, Spherical Insights, The Business Research Company