Cyrus Safaie worked at Opex Analytics and Convoy before joining DoorDash, where he eventually led engineering and AI/ML efforts focused on marketplace operations. His view of DoorDash starts with a point that is easy to miss from the outside. The company is not operating a single marketplace but thousands of small markets, each with different demand patterns, regulations, taxes, pay rules, merchants, consumers, and dashers. A driver working in one part of Chicago may never work in another, so optimization must start locally before it can be coordinated across the broader network.
DoorDash initially handled many of those problems manually. Safaie describes teams assigning deliveries and planning incentives in spreadsheets before building more sophisticated systems, and, in some cases, using different ETA models across markets. The logic was to first understand what worked in each market, then codify those decisions and gradually increase the number of variables the system could optimize. That approach reduced the risk of imposing a single global model on markets with fundamentally different economic and operating conditions.
The harder problem is that DoorDash has to optimize for three groups at once. Dashers want higher pay, consumers want lower prices and reliable delivery, and merchants need enough volume and margin to justify participating. DoorDash therefore created teams focused separately on each side of the marketplace before reconciling those objectives at a broader financial level. Safaie says the long-term objective ultimately comes back to free cash flow, but that only works if drivers continue returning, consumers keep ordering, and merchants remain available on the platform.
Supply-demand balancing shows how quickly those decisions can change. DoorDash has to decide when and where to offer additional incentives to bring dashers online, often at a very granular market level. During the Netflix Tyson fight, demand spiked beyond peak-holiday levels, and the company hadn't planned for it, forcing teams to quickly increase spending to bring more dashers online. Safaie argues that this is also why human judgment works better as an input to automated systems than as an override after the decision is made. If operators simply override a system, the reason is often lost. If their judgment is provided as input to the model, the system can incorporate that context into future decisions.
For analytics leaders, the lesson is that marketplace optimization depends as much on system design as on model quality. DoorDash improved ETAs by predicting the errors in its own models and automatically correcting them, while its supply systems evolved from manual spreadsheets into automated decision systems that still accept human input. Start with human judgment, capture the operating logic, automate what can be codified, and keep expertise focused on the decisions the system cannot make.
The DecideWise Insider
Benjamin Baer, Founding Member of the DecideWise community, argues that improving decision velocity starts with separating decisions by frequency and risk. High-frequency, low-risk decisions, such as lead routing or routine approvals, are strong candidates for full automation, while high-frequency, high-risk decisions, such as credit underwriting or fraud detection, require automated execution with safeguards and escalation paths. Low-frequency, high-risk choices such as acquisitions, executive hires, and major platform migrations should remain human-led, with technology supporting analysis. Baer’s framework uses a risk-versus-frequency matrix to determine where automation adds value and where human judgment remains essential, with human oversight shifting toward governance, audits, and intervention as more operational decisions are automated.
Market Pulse
SAS is bringing Decision Builder to Microsoft Fabric, enabling teams to combine business rules, predictive models, and operational constraints within governed decision workflows. The system can incorporate eligibility rules, regulatory requirements, and risk thresholds, then execute an action automatically or route it for human approval. The integration is designed to connect analytics directly with repeatable business decisions inside Fabric.
IBM has added Business Value Alignment to watsonx.governance, providing organizations with a way to connect AI initiatives to business cases, strategic objectives, and standardized KPIs. Teams can compare expected value with realized outcomes using measures such as cost per decision, processing time, and manual rework. The feature extends governance beyond model risk to whether deployed AI is producing the intended business result.
Cognizant has deployed an AI agent network across more than 10,000 delivery projects to monitor obligations, risks, customer sentiment, audits, and institutional knowledge. More than 8,000 employees were using the system by June 2026. One assessment agent has reached a 42% recommendation acceptance rate, while humans continue to make the final decisions.
Strata Decision Technology has launched Financial Decision Intelligence inside StrataJazz, combining financial data, analytics, and AI for healthcare finance teams. The system monitors KPIs, identifies changes and possible root causes, generates forward-looking predictions, and supports natural-language investigation. It uses the financial definitions, calculations, security controls, and governed data already maintained inside StrataJazz.
Resources and Events
📅 Generative AI Boston (Boston, MA - October 29, 2026)
Generative AI Summit Boston brings together AI engineers, researchers, technical leaders, and executives at the Westin Boston Seaport for a day of workshops, breakout sessions, roundtables, and technical discussions on deploying generative and agentic AI in production. The event expects 500+ attendees and 60+ speakers, with sessions covering agentic stacks, MCP frameworks, governance, autonomous workflows, and enterprise AI deployment.
📅 The European Chief Data & Analytics Officer Conference 2026 (Amsterdam, Netherlands - November 19, 2026)
The European CDAO Conference 2026 brings together senior data, analytics, and AI leaders at Rewire Amsterdam for panel discussions, roundtables, and networking focused on codified knowledge and agentic AI. The agenda will examine how enterprises can scale AI and data capabilities, strengthen governance, and turn organizational knowledge into systems that support better decisions and automation. Speakers include data and AI leaders from Picnic, KLM, Unilever, adidas, Mondelēz International, Elsevier, Decathlon, and Rewire.
📊 Report Spotlight: The Decision-Confidence Gap (TheyDo)
TheyDo surveyed 1,000 enterprise decision-makers and found that 68% say AI has made decisions faster, but speed has not removed uncertainty. 77% say dashboards, CRM data, and customer feedback sometimes tell different stories, while 43% delay decisions at least once a week due to insufficient context. 38% have already seen an AI-driven decision create an unintended customer experience problem, and 86% say they would trust AI recommendations more if they could inspect the underlying data and context.
Join the conversation: Is AI making decision-making better or worse?
How should organizations think about AI’s real impact on decision quality, governance, and operations? Join the DecideWise community discussion and weigh in.
For the Commute
The Politics of Analytics (Decision Intelligence Lab)
Zahir Balaporia argues that the hardest part of analytics is often not building or deploying the model, but getting people to trust it enough to change how they work. He distinguishes the “last mile” of putting decision tools into production from the “extra mile” of navigating culture, incentives, organizational politics, and existing beliefs that determine whether those tools are actually used. Drawing on his work at Schneider, he describes how dispatch optimization required redefining the dispatcher role, changing accountability, and allowing humans to intervene only when real constraints or bad data justified it.
