Uber has published the most complete public account so far of what it costs to run agents across an entire software organization, and the number that matters is not the adoption number. Between February and mid-August 2026, weekly agentic requests at Uber grew 9.4 times and weekly active users grew seven times, while total AI spend, in Uber’s own words, has relatively stabilized since April. With one model held constant, cost per session fell 52 percent from its June peak and cost per thousand requests fell 34 percent. Uber did not get there by capping engineers. It got there by moving work out of engineers’ terminals and into managed agents it could benchmark, route and gate. That is a governance decision in an engineering costume, and it is the clearest public illustration I have seen of the middle levels of my Agentic AI Roadmap.
Agentic AI is the next enterprise workload abstraction: a governed, goal-driven software layer where LLM-powered agents autonomously plan and execute multi-step business processes.
That is my definition, and the word carrying Uber’s case is governed. Instructions in, results out was IT. Intent in, outcomes out is agentic AI. Uber’s post is about keeping the outcomes coming without the bill coming with them.
What Uber actually built
The scale first. More than 70 percent of Uber’s pull requests are now attributed to agents, engineers have written more than 3,600 agent skills, and those skills run more than 30,000 times a day, according to the engineering post Uday Kiran Medisetty published on August 27. A growing share of sessions are not started by a person at all. Managed agents review code, repair failing CI, finish end-to-end pull requests with visual checks, triage on-call alerts and debug incoming bugs, with humans reviewing and catching escalations.
Uber sorts all of this into four layers, from the most specialized, a managed agent doing one job, to the most general, an engineer at a terminal. The more specialized the layer, the more control the platform team holds over cost, quality and which model runs. Then it writes cost as six multiplied terms: users, sessions per user, turns per session, requests per turn, tokens per request, price per token. The first two are adoption, which Uber wants more of. The middle three are the agent’s own overhead, the searching and re-sending and retrying it does on top of what the engineer asked for. That is where the effort went.
The cost equation is a control
I want to be precise about why this travels beyond Uber. A cost equation with named terms is a governance instrument. Every term gets an owner, a metric and a lever. Uber’s levers are not exotic. Benchmark real work before choosing a model. Default subagents to a cheaper model, because they do bounded tasks with known inputs. Compact context at 400,000 tokens even on a million-token model. Default reasoning effort to medium. Stretch the prompt-cache window to an hour for interactive sessions, because engineers walk away from the keyboard and a five-minute cache expires while they get coffee. Route every MCP server, more than 1,000 of them, through one gateway and expose them as shell commands so tool schemas stop riding along in every prompt.
That last one deserves its own paragraph. With more than 100 tools installed, standard MCP was loading 50,000 to 70,000 tokens of tool schema into a session before anyone typed a word, and re-sending it on every turn. One SaaS vendor’s server bundled 49 tools and roughly 22,000 tokens of schema on its own. Running the same warehouse query through a shell script instead of the model’s polling loop cut tokens by more than half on small result sets and by more than 90 percent on bulk jobs. None of that is a model upgrade. All of it is work on the harness, the layer of controls around the model, which is the one layer an enterprise actually owns.
The numbers a P&L reader will want
Every figure below is Uber’s own, self-reported, and Uber says plainly that the specific reductions are unique to its environment.
| Measure | Uber’s figure | Why a CFO cares |
|---|---|---|
| Weekly agent requests, Feb to mid-Aug 2026 | Up 9.4 times | Demand grew almost tenfold |
| Weekly active users, same period | Up 7 times | Adoption, the term Uber wants to grow |
| Total AI spend since April | “Relatively stabilized” | Volume up, bill roughly held |
| Cost per 1,000 model requests, model held constant | Down 34 percent from peak | Unit cost, not vendor discount |
| Cost per session, model held constant | Down 52 percent from June peak | The headline unit-economics number |
| Cached prompt reads | 0.1 times the input rate | Why the one-hour cache pays despite a 2x write premium |
| Tool schema overhead per session, 100+ tools | 50,000 to 70,000 tokens, re-sent every turn | Pure waste before the first prompt |
| Warehouse query via code-mode, small result sets | 55 to 71 percent fewer tokens | Same answer, half the bill |
| Wide-table query via code-mode | 1,431,594 tokens to 900 | One query, near-total elimination |
| Bulk workflows via code-mode | More than 90 percent saved | N model turns become one script |
| Grounded versus ungrounded agent, same prompt | 38 seconds and correct, versus 20 minutes and wrong | Time and tokens, plus a wrong answer avoided |
| One user’s session dashboard | $4,162 spent, 433 sessions, 95 percent cache hit rate, $1,214 savings identified | Waste is findable per person |
Read the table top to bottom and the shape of the P&L is clear. Demand rose 9.4 times. Unit cost fell by a third to a half. The levers that did it were not price cuts from the vendor. They were a longer cache window, tool schemas kept out of the prompt, polling loops moved into a script, and an agent that knew where to look. In my per-decision cost framing, Uber shrank the cost of a completed task while the number of tasks exploded. That is the only ROI story in agentic AI that survives an audit.
Where Uber sits on the roadmap
My Agentic AI Roadmap has five levels, Prompted, Piloted, Governed, Assured and Autonomous, and each is earned by institutionalizing the one below it. So where is Uber, honestly?
The interactive terminals are Level 1 and Level 2 by nature. But Uber wrapped that layer in governed defaults. One wrapper for installation, authentication and cost visibility. One spend tier across every harness. Slack nudges at 50, 80 and 100 percent of expected spend. Manager sign-off to raise the tier. That is a Level 3 control applied to Level 1 usage, and it is the right move. You do not stop engineers prompting. You make prompting governed.
The managed agents are the real story. Each one follows the same path: define the outcome metric, build a benchmark from the agent’s actual work, pick the Pareto-efficient model, and keep re-picking as the frontier moves every few weeks.
What Uber has not shown is Level 5. Its own next-steps list says so. Recording papercuts from skill executions and generating skill updates from the traces is still in progress, and dynamic model routing is still being built out. Until the loop closes on its own, the honest placement is Governed across the organization, Assured on the managed-agent fleet, and Autonomous as a stated direction. That is a strong position. Most of the fifty enterprise deployments I mapped are still stuck at Piloted.
More on agentic AI cost and control
- The Roadmap Itself: The five levels from Prompted to Autonomous, with the exit gate each one needs before moving up.
- Compute Is Not Fixed: Why agentic AI cost has to be governed per decision, not per line item.
- Govern the Harness: The controls around the model, and who should own them, stage by stage.
- The Agentic AI Hook: The one control in the stack that decides an outcome instead of influencing one.
- Ten Facts First: The cost and eval facts I wish I had known before building my first agentic system.
What I do with this in loan origination
I write the cost equation down before the first agent ships, with a named owner for each term. I benchmark every managed agent against real cases with known answers, graded easy, medium and hard, and I hold the model constant when I measure my own improvements, because otherwise every vendor upgrade gets booked as my work. And I keep interactive use governed rather than banned. A shared tier, a visible counter, an approval path. Engineers who can see what a session costs make better calls than engineers who are told not to have one.
The point most people will miss
Uber’s closing argument is the one I would put on the wall. A fleet of specialized managed agents, each with its own benchmark and its own model, is cheaper to optimize and easier to scale than thousands of individual terminal sessions. Read that as an engineering claim and it is about cost. Read it as a governance claim and it says something larger. You cannot govern what runs in ten thousand terminals. You can govern a factory. Every workload Uber moves from an interactive session into a managed agent picks up an identity, a benchmark, a model policy, a budget and a trace. That is what climbing the roadmap looks like when nobody is calling it a roadmap.
The 52 percent is not the headline. The headline is that Uber can prove the 52 percent with the model held still, which means the gain belongs to the organization and not to the vendor. Instructions in, results out was IT. Intent in, outcomes out is agentic AI, and here the outcome includes the bill. I wrote the longer version in Intent In, Outcomes Out and the case for earned autonomy in Earned Autonomy.
Which of your agent workloads is still running in a terminal that nobody can benchmark?
Go deeper
- Agentic AI Architect: control design for enterprise agents.
- Agentic AI Case Studies: deployment evidence, one teardown at a time.
- Agentic AI P&L: cost, payback, and risk in a CFO’s voice.
- Agentic AI Governance: who owns the harness.
- Substack first access: agenticaiarch.substack.com/subscribe
© Dr. Harish Kotadia, Ph.D., All Rights Reserved, 2026.
Dr. Harish Kotadia, Ph.D., is an Enterprise AI Architect with 20+ years of IT consulting experience serving Fortune 100 clients, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock, and Google Vertex AI.
Disclaimer: This blog post is based on publicly available academic publications, vendor documentation, open standards, and news items from reputed media sources linked above. This post is intended for educational purposes, to help the enterprise agentic AI community build a shared vocabulary from public, authoritative sources.
Views and opinions expressed here are my own and do not represent those of any employer or client, past or present. The analysis presented is my independent interpretation of the published sources linked above and does not constitute legal, financial, or consulting advice of any kind.