Canonical: https://deeplight.ae/insights/simulating-agents-in-workflows
[Reading guide](https://deeplight.ae/llms.txt): company facts, services, published articles and optional visual assets.
Author: Sam Patel
Published: 2026-10-19
Format: Markdown representation of the public HTML page. Interactive controls and diagrams are available in the HTML version.

[All Insights](https://deeplight.ae/insights)

# Simulating agents in a workflow: price each step before you build

Strategy Part 2 of 2 19 October 2026 12 min read

By Sam Patel

Part 1 mapped a workflow and found the steps where an agent fits. This part asks what an agent at those steps would return, before you build it. It explains why faster steps don’t always make faster workflows, the four relationships a simulation needs, the inputs to measure and what a case really costs. An interactive simulator then runs part 1’s hypothetical know-your-customer (KYC) review with and without the agent.

Read part 1 first. [Mapping operational workflows: find the steps where agents pay off](https://deeplight.ae/insights/mapping-operational-workflows)

## In short

- [**The slowest step sets the pace** Speeding up another step doesn’t raise output. In Faros AI’s data, waits for a first code review more than doubled at peak AI adoption.](https://deeplight.ae/insights/simulating-agents-in-workflows#section-1)
- [**Price the review, not only the model** Each agent output costs review time and the redo of what review catches, plus the chance an error gets through times the work to fix it.](https://deeplight.ae/insights/simulating-agents-in-workflows#section-4)
- [**Analyst hours fall 24%; days fall 4%** In the hypothetical worked example, the agent saves analysts’ time, but the customer’s reply still sets the pace.](https://deeplight.ae/insights/simulating-agents-in-workflows#section-5)
- [**The queue moves to the next constraint** In the hypothetical example, at 1,400 reviews a month, the agent keeps analysts within capacity, but approvals fall behind by about 255 a month.](https://deeplight.ae/insights/simulating-agents-in-workflows#section-5)

## Why a faster step doesn’t always make a faster workflow

A faster step doesn’t always make a faster workflow. Faros AI’s telemetry covers 22,000 developers in more than 4,000 teams, and compares each organisation’s periods of lowest and highest AI adoption. At the highest, tasks completed per developer rose 33.7%. The median wait for a first code review rose 156.6%, and changes merged with no review at all rose 31.3% [1]. Faros sells engineering analytics, and these are correlations, but the pattern is an old one.

Operations research explains it. In the theory of constraints, the slowest step sets the output of the whole chain. Speeding up any other step doesn’t raise output, and once the constraint is relieved, another step becomes the constraint. A queue of work in front of a step is a classic sign that it’s the constraint [2]. In Faros’s data, the queue moved to the people who review the code [1].

So an agent’s return depends on the rest of the workflow. It depends on how much review the output needs, who has time for that review and what happens to errors that get through. A simulation answers those questions before you build, using the map from [part 1](https://deeplight.ae/insights/mapping-operational-workflows).

## Four relationships a simulation needs

Four relationships do most of the work in a workflow simulation, and you can state each one in words:

- **Work in progress.** The cases in progress equal the cases that arrive each day times the days each case takes. This is Little’s law, and it holds for the long-run averages of any stable workflow [3]. Use it to check that a promised cut in time per case matches a fall in the backlog.
- **Waiting under load.** Waiting time rises steeply as a team’s utilisation, the share of its working time spent busy, approaches 100%. It rises further as the time each task takes varies more [4]. So an AI step can cut the average handling time and still lengthen queues. In one model’s worked example, AI-routed work was faster on average, but rework from missed errors made it far more variable. It waited longer than manual work whenever demand was below about 76% of the team’s manual capacity [4].
- **Reviewer capacity.** The cases that must go to a person to keep risk within limits, times the minutes each takes, must stay below the minutes reviewers have. Past that point, in one theoretical model, no escalation policy is both safe and stable: either the queue grows without limit, or risky cases go unreviewed [5].
- **Human time per agent case.** Each case an agent handles still costs a person the review time, plus the chance that an error gets through times the time to fix it later [4]. If reviewers redo the wrong outputs they catch, add that time too.

None of these needs a model of the agent’s internals. They need the map’s volumes, handling times and waits, plus the inputs in the next section.

## Inputs to measure, not assume

A simulation is only as good as its inputs. Four inputs decide most results, and each one is easy to get wrong:

- **Reliability on every try.** A high-volume step needs an agent that gets the same case right every time, not once. One benchmark, τ-bench, has agents serve simulated customers. On it, an agent with 61% success on one try succeeded on all eight tries of a task less than 25% of the time [6]. Measure success across repeated runs of your own cases.
- **What review catches.** In Anthropic’s own data on its Claude Code coding tool, users approve 97% of permission prompts. In Anthropic’s controlled test with 1,053 paid testers, people caught 13.6% of dangerous commands, against 89% for an automated classifier. Users rejected 39% of plans but only 3% of single actions [7]. Review a whole output or plan, and assume that some errors pass.
- **The share that reaches a person.** Salesforce’s data from businesses with agents in production puts the escalation rate from AI agents to human agents at about 32%. Salesforce says the rate is holding steady as conversations grow [8]. Plan for a lasting human share, and staff for it.
- **The hours people really give.** Van der Aalst lists 15 ways that business-process simulations go wrong. One is modelling a manager who spends 20% of their time on a process as if always present, which makes the results too optimistic. People also work in batches and change speed under pressure [9]. Model the hours a day each team really gives the queue, and treat even that as a simplification.

## What a case really costs

An agent’s cost per case has four parts: the model, the review, the escalations and the rework. Review is a person’s time on each output, including redoing what the review catches. Escalations are cases the agent hands to a person, and rework fixes errors that get through. If you leave any of them out, the agent looks cheaper than it is.

List prices leave most of that out. Intercom charges $0.99 per resolution for its Fin agent. It counts a conversation as resolved when the customer confirms the answer helped, and also when the customer leaves without asking for more help [10]. Gartner expects the cost per resolution of generative AI in customer service to exceed $3 by 2030, more than many offshore human agents cost [11]. In McKinsey’s 2026 survey, about one respondent in five said AI operating costs, including tokens, limited their use of AI [12].

The model’s cost per task depends on the model and where it runs. Compare hosted and self-hosted options at the quality your task needs, including the cost of operation and support. The other three parts depend on the workflow, which is what the simulation in the next section measures.

## The hypothetical example, simulated

This section simulates the hypothetical KYC review from part 1: 1,000 reviews a month, 22 analysts giving reviews 6 hours a day, and 5 approvers giving them 3. By default, the agent is right on every try for 90% of outputs. An analyst spends 10 minutes reviewing each one, and 15% of wrong outputs get past the review. Each error that gets through costs 3 hours of rework, and the model costs $3 a review. Every number is an assumption for illustration.

Figure 1 compares six versions of the review.

| Version | Analyst hours a month | Approvers’ time needed | Working days, mean (90th percentile) | Cost per review | Errors found after approval a month |
| --- | --- | --- | --- | --- | --- |
| No agent | 2,375 | 87% | 8.4 (15.1) | $146 | 0 |
| The agent checks documents and drafts the memo | 1,807 | 87% | 8.1 (14.8) | $121 | 30 |
| The agent also drafts document requests | 1,638 | 87% | 7.3 (13.5) | $112 | 30 |
| The two steps, with a less reliable agent | 2,262 | 87% | 8.2 (15.0) | $144 | 105 |
| The two steps, at 1,400 reviews a month | 2,530 | 122% | No limit | $121 | 42 |
| The same, with 7 approvers | 2,530 | 87% | 8.4 (15.1) | $121 | 42 |

**Figure 1.** Six versions of the hypothetical KYC review, simulated. The less reliable agent is right on every try for 65% of outputs. “No limit” means a pool can’t keep up, so the queue grows for as long as the run lasts.

**The agent saves hours, not days.** When it checks documents and drafts memos, analyst hours fall 24%, from 2,375 to 1,807 a month, and a review costs $121 instead of $146. But a review still takes 8.1 working days, against 8.4. Waiting for the customer takes 7.4 of those days, and neither step touches it.

**The step that saves days is the request.** When the agent also drafts the document request, assume that 15% of replies miss a document instead of 35%. A review then takes 7.3 working days, and analysts need 1,638 hours a month.

**Reliability decides the saving.** At 65% instead of 90%, analysts redo so many outputs that their hours fall only to 2,262. About 105 errors a month get past review, against 30, and their rework takes back part of the saving.

**Capacity moves the queue.** At 1,400 reviews a month, analysts with the agent are busy 91% of their review hours; without it, they would need 120%. But approvers now need 122% of theirs, so the approval queue grows by about 255 reviews a month. With 7 approvers, a review takes 8.4 working days.

None of these numbers forecasts anything for a real bank. They show which questions a simulation answers, and how much the answers depend on reliability, review and capacity. To test your own assumptions, change the controls in the following simulator. It reruns the same workflow with and without the agent each time, and Figure 2 shows where a review’s time goes.

These results use the default assumptions: the agent checks documents and drafts the memo, at 1,000 reviews a month. Turn on JavaScript to change them.

Time from opening a review to approval

8.4 → 8.1 working days 90% finish within 15.1 → 14.8 working days.

Analyst hours a month

2,375 → 1,807 The work fills 86% → 65% of the 6 hours a day analysts give reviews.

Approvers’ time needed

87%The work fills this share of the 3 hours a day approvers give reviews. Above 100%, the queue never clears.

Cost per review

$146 → $121 Analyst time at $50 an hour and approver time at $100, plus $3 of model cost.

Errors found after approval

30 a month Errors in checks and memos that passed an analyst’s review, at 3 hours of rework each.

Longest queue

Approve the rating A review waits 2.3 working hours for this step on average.

Without the agent 8.4 days

With the agent 8.1 days

Waiting for the customer Waiting in a queue Being worked on

| Version | Waiting for the customer | Waiting in queues | Being worked on | Total |
| --- | --- | --- | --- | --- |
| Without the agent | 7.4 | 0.7 | 0.3 | 8.4 |
| With the agent | 7.4 | 0.4 | 0.2 | 8.1 |

**Figure 2.** Where a review’s time goes, in working days from opening to approval. Each run simulates 20 months of reviews and leaves out the first 2.

## Teams that simulated first

Nubank screened its customer-service agents in a simulator before releasing them. The figures come from Nubank and Guardrails AI, which makes the simulator. Nubank tested 29 model configurations across more than 16,000 simulated conversations. Across four deployed versions, failure scores in simulation tracked failure scores in production, with a Pearson correlation of 0.74. An automated judge scored the simulated conversations. By its scores, changes Nubank found in simulation cut unnecessary transfers to people from 22.4% to 16.0%. In a live A/B test, which splits real traffic between two versions, the redesigned agent beat the agent it replaced. Its self-service rate was 4.9 percentage points higher [13]. The simulated customers wrote longer messages than real ones, and the correlation rests on four versions. So the simulator screened options rather than predicting results.

Researchers at a university hospital simulated its chest X-ray reporting workflow, using the hospital’s own exam rates, reporting rates and mix of findings. In the simulation, ordering the worklist by an AI’s findings cut the mean time to report a collapsed lung (pneumothorax) from 80.1 minutes to 35.6. But the AI missed some collapsed lungs, and those scans went to the end of the list. The longest wait rose from 890 minutes to 1,178. A cap on the maximum wait brought the longest wait down to 979 minutes, and the mean rose only to 38.5 [14]. The simulation exposed the risk and a fix without touching a real worklist; the paper reports no deployment.

You don’t need a research team to do this. BPSim, an open standard, attaches simulation parameters for time, resources, cost and branching to a process model. One scenario can inherit from another and change only what differs [15]. Apromore’s process-mining software also runs what-if simulations: its assistant compares fully or partly automating an activity with a baseline model [16].

## From simulation to pilot

A simulation turns a business case into a short list of numbers to measure. When you pilot the agent, measure the numbers the simulation assumed:

- The share of your own cases the agent gets right every time, across repeated runs.
- The minutes a reviewer spends on each output, and the share of planted errors reviewers catch.
- The share of cases sent to a person, and the time each one takes.
- The hours a day reviewers really give the queue.

Then rerun the simulation with the measured values before you scale. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, because of rising costs, unclear business value or weak risk controls [17]. This process puts a number on each of those before you build. Our [agentic workflow services](https://deeplight.ae/services/agentic-workflows) connect this evaluation to the design, build and operating decisions.

## How the simulation works

The simulator is a discrete-event simulation written for this article: it follows each review from step to step, one event at a time. It works as follows:

- Reviews arrive at random times, at the monthly rate you set.
- Handling times, review times and customer replies vary around their averages, with a few taking much longer than the rest (lognormal distributions). Customer replies vary most.
- Analysts and approvers work only the hours a day they give reviews, and take work in the order it arrives.
- An analyst reviews every agent output and redoes the wrong ones they catch. A missed error in a document request brings back an incomplete reply. A missed error in a check or a memo is found after approval and costs rework.
- Each run covers 20 months of 21 working days, and leaves out the first 2 so queues can settle.
- Each review draws the same random numbers in every version, so differences between versions come mainly from the change, not from chance.

We checked the simulator against two standard queueing formulas: Pollaczek-Khinchine for a single worker, and Erlang C for three workers sharing a queue. The model leaves out absence, priorities, batching, changes of speed under pressure, approvers returning a review because of an agent’s error, and people getting faster with practice. It also leaves out the agent’s help on screening alerts and ownership, where part 1 has it help analysts rather than do the step. The cost figures count staff time at $50 an hour for analysts and $100 for approvers, plus the model’s cost.

## Sources

Figures and dates were checked against each source on 19 October 2026. Vendor and company results are their own figures. The worked example is hypothetical.

1. [Faros AI, Ten takeaways from the AI Engineering Report 2026](https://faros.ai/blog/ai-acceleration-whiplash-takeaways). 12 April 2026. Faros’s own telemetry; the figures are correlations.
2. [Tomaszewska, Constraint management: the review](https://doi.org/10.26552/tac.C.2020.1.7). Transport and Communications 8(1), 2020.
3. [Little, Little’s law as viewed on its 50th anniversary](https://doi.org/10.1287/opre.1110.0940). Operations Research 59(3), 2011.
4. [Bartolucci and Vivo, Queue and AI: when faster tasks slow down the workflow](https://arxiv.org/abs/2605.27202). arXiv 2605.27202, May 2026. An analytical model with assumed parameters.
5. [Wang and Rachev, Operating imperfect AI: reliability drift and human congestion](https://arxiv.org/abs/2601.22295). arXiv 2601.22295, January 2026. A theoretical model.
6. [Yao and others, τ-bench: a benchmark for tool-agent-user interaction in real-world domains](https://arxiv.org/abs/2406.12045). ICLR 2025.
7. [Anthropic, Auto mode is now the default in Claude Code](https://claude.com/blog/auto-mode-default-in-claude-code). 7 August 2026. Anthropic’s own study, on coding actions.
8. [Salesforce, Agentic Enterprise Index](https://www.salesforce.com/news/stories/agentic-enterprise-index-insights-2026/). 7 August 2026. Salesforce’s own data from businesses with agents in production every month, February 2025 to April 2026.
9. [van der Aalst, Business process simulation survival guide](https://www.vdaalst.com/publications/p808.pdf). Handbook on Business Process Management 1, Springer, 2015.
10. [Intercom, Fin AI Agent outcomes](https://www.intercom.com/help/en/articles/8205718-fin-ai-agent-outcomes). Help Center, accessed 19 October 2026.
11. [Gartner, GenAI cost per resolution for customer service will exceed offshore human agent costs by 2030](https://www.gartner.com/en/newsroom/press-releases/2026-01-26-gartner-predicts-genai-cost-per-resolution-for-customer-service-will-exceed-offshore-human-agent-costs-by-2030). Press release, 26 January 2026. A forecast.
12. [McKinsey, The state of AI in 2026: on the road to ROI](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai). Global survey of 1,719 respondents, May to June 2026; published 25 August 2026.
13. [Nubank and Guardrails AI, Screen before you serve: simulation for production customer experience AI agents at 140M scale](https://arxiv.org/abs/2609.30137). arXiv 2609.30137, September 2026. Reported by the bank and the simulator’s maker.
14. [Baltruschat and others, Smart chest X-ray worklist prioritization using artificial intelligence: a clinical workflow simulation](https://arxiv.org/abs/2001.08625). European Radiology 31(6), 2021.
15. [Workflow Management Coalition, Business Process Simulation (BPSim) specification 2.0](https://www.bpsim.org/specifications/2.0/WFMC-BPSWG-2016-01.pdf). December 2016.
16. [Apromore, Copilot for simulation](https://documentation.apromore.org/simulation/copilotforsimulation.html). Documentation, version 10.6.
17. [Gartner, Over 40% of agentic AI projects will be cancelled by the end of 2027](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027). Press release, 25 June 2025. A prediction, not a measured outcome.

[Talk to us](https://deeplight.ae/contact)

## Put this thinking to work

[**Programmes** Choose an AI use case or build and test one.](https://deeplight.ae/services/programmes)

## Explore the offer

[**Discovery**](https://deeplight.ae/services/discovery) [**Accelerate**](https://deeplight.ae/services/accelerate)

## Related reading

[All Insights](https://deeplight.ae/insights)

Strategy

### [Mapping operational workflows: find the steps where agents pay off](https://deeplight.ae/insights/mapping-operational-workflows)

Read the article

Practice

### [Agentic workflows in practice](https://deeplight.ae/insights/agentic-workflows-in-practice)

Read the article
