# The agents shipped. The operating model did not.

> Genpact finds 79% of large companies run AI agents in production, but only 9% have changed their operating model. That gap is where the value goes missing.

- Type: Opinion
- Published: 2026-11-13
- Author: Sentynel
- URL: https://sentynel.ai/sygnals/agents-shipped-operating-model-did-not/
- Language: en

Most large companies now run AI agents in production, and very few have changed how the company runs around them. A new Genpact survey puts numbers on that gap, and **the gap, not the agents, explains why so little of the value reaches the P&L.** Adding agents to an unchanged operation gives you the same operation with a new layer on top.

## Four numbers from the same companies

Genpact published "Autonomy in Practice" on 12 November 2026. It surveyed more than 2,000 senior executives at companies with annual revenues from $1 billion to more than $50 billion, across 17 industries. The findings are self-reported, and Genpact sells agentic operations services, so read it as a vendor's survey. The pattern is still worth a close look.

- 79% of the enterprises surveyed have AI agents in production.
- On average, 49% of the AI initiatives funded in the past 24 months delivered the value expected.
- Only 15% have fully tied AI investment to P&L outcomes.
- Only 9% have completed significant operating model transformation.

**Agents deployed vs operations changed**

- 79%: have AI agents in production
- 15%: have fully tied AI investment to P&L outcomes
- 9%: have completed significant operating model transformation

Source: Genpact, "Autonomy in Practice", press release of 12 November 2026. Survey of 2,000+ executives at companies with $1B+ revenue; self-reported, rounded.

Genpact's chief technology and innovation officer, Sanjeev Vohra, summed it up in the release: "Enterprises have become effective at deploying AI agents. Running them is the hard part."

## The easy reading is wrong

The easy reading of these numbers is that it is early. Give the agents better models, give finance a better ROI dashboard, and the value will show up.

We disagree. **A measurement problem would look the same at every company. This one tracks the operating model.** You cannot tie an agent to a P&L line if nothing on that line changed: the same people approve the same things, the same report gets built, the same queue waits for the same owner. The agent did real work, and that work landed in a structure that had no place to put it.

That is what "deploying" without redesigning means in practice. The agent was added. Nothing was removed.

## An added layer looks like this

Two more numbers in the survey describe the layer. Genpact reports that AI now leads or fully runs 40% of decisions, more than double a year ago, yet humans still decide or audit in at least 87% of cases at every decision stage. And 92% of respondents with agents in production kept their human workforce the same size or grew it.

Neither number is bad on its own. Keeping people in the loop is often the right call, and Sentynel does not count success by heads removed. But put the two together and a picture appears: the agent produces a decision, and then a person does roughly what they did before, at every stage. The work did not move. It doubled.

**Value reaches the P&L when someone's work changes, not when an agent is switched on.** That change has to be designed: which decision the agent now owns, which step the person stops doing, what they do with the time, and who answers for the result.

## What an operating model change actually is

"Operating model transformation" sounds like a multi-year programme. At the level of one process, it is a short list of decisions that somebody has to write down and own.

**From an agent in production to a line on the P&L**

1. Pick one line — A cost, a revenue or a cycle time someone already reports
2. Name the owner — One person answers for that number
3. Move the decision — Write down what the agent now decides
4. Retire the old step — The work a person stops doing
5. Measure on the line — Before and after, same definition

The fourth step is the one most companies skip. Retiring the old step is uncomfortable, because it changes someone's job and someone's approval. It is also the only step that turns agent output into a number finance will recognise.

## Start from the question the business is asking

Sentynel's Campoverde case started at the business end. Sales had one view and analytics had another; the work was to make them a shared decision layer, and then build a Walmart campaign from what the data revealed. Sentynel's case record reports a **+18% reported sales uplift**. Sales uplift is a reported campaign outcome; the chart uses a normalized baseline.

The useful part for this argument is the order. The question came first, the people who had to act on the answer were named, and the system was built to serve that decision. The result could be stated as a sales number because the work was designed around a sales number from the start.

## A test for your agent list

Take the list of agents your company has in production. Next to each one, write two things: the P&L line it moves, and the step a person no longer does because of it.

If you can fill in both columns, you have changed the operation, and you can measure it. If the second column is empty, you have added a layer. That is not a failure, but it is the reason the value is hard to find, and it tells you exactly where the redesign has to start.

## Sources

- https://www.prnewswire.com/news-releases/79-of-enterprises-report-ai-agents-in-production-but-only-15-have-fully-tied-ai-investment-to-pl-outcomes-302937385.html
- https://www.genpact.com/insight/autonomy-in-practice
