Skip to content
Talk to us
MenuClose
All Insights

Mapping operational workflows: find the steps where agents pay off

By Sam Patel

An agent’s return differs from step to step, and the hard part is finding the steps where it pays off. This is the first of two parts. It shows how to map an operational workflow as it really runs and score each step against six traits from the research. It then shows how to decide whether an agent does each step, helps a person or stays out. A worked example maps a hypothetical know-your-customer (KYC) review at a bank, and part 2 simulates that map before anything is built.

In this article

In short

Why the step decides the return

Using AI and profiting from it are different things. In McKinsey’s 2026 survey, only 37% of respondents attribute any effect on earnings before interest and taxes (EBIT) to AI, about the same as in 2025. About 6% are high performers: they attribute 5% or more of EBIT to AI and call its impact significant. Meanwhile, 40% of respondents from companies with more than $1 billion in revenue report scaling agents, up from 27% a year earlier [1].

The Gulf shows the same gap between use and value. In a survey of 139 executives and board directors in the Gulf Cooperation Council (GCC), 84% said their organisations had adopted AI. Only 11% said their organisations were scaling AI and could attribute 5% or more of earnings to it [2].

The high performers in McKinsey’s 2026 survey differ in how they treat the workflow. Nearly three-quarters of them say they fundamentally redesigned workflows, against about a quarter of other respondents [1].

Returns also differ from step to step inside one workflow. EY describes agentic AI in three steps of an anti-money-laundering investigation for an unnamed financial intelligence unit. It summarises the case background, helps investigators analyse transactions through a chat interface and automates the case documentation. EY reports time savings of 25% on the summary, 70% on transaction analysis and 50% on documentation, more than two hours a case in all. The decision to file stayed with the investigator [3]. These are EY’s own figures, without a stated method.

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 [4]. Each of those is a property of a step: what it costs to run, what it returns and what happens when it goes wrong. So the work starts with a map of the steps.

Map the work as it really runs

A process document describes how work should run; event logs show how it does run. Process mining rebuilds the real paths from the timestamps your systems already record. In one analysis of a public purchase-to-pay log from a multinational paints and coatings company, 251,734 order items followed 11,973 different paths, one for every 21 items [5]. Exceptions, rework and hand-offs sit in those paths, and so does much of the work an agent would take on.

Build the map from four sources, because each one misses something:

  • Event logs. Your case, workflow and core systems record the order of steps, their volume, how long each takes and how long work waits between them.
  • Object-centric logs. A review touches a customer, documents, owners and alerts. Flattening them into one case can produce a misleading map, so object-centric logs keep the links between objects [6].
  • Task mining. Recordings of desktop work show what happens between system events, where much of the handling time sits [7].
  • Interviews and observation. Workarounds create mismatches between event logs and what people really do [8]. Interviews and observation show judgement and workarounds, and process mining complements them rather than replacing them [9].

Record the same measures for every step, so steps can be compared (Figure 1).

MeasureWhere it comes fromWhy it matters for an agent
VolumeEvent logsHigh volume spreads the cost of building and checking an agent.
Handling timeEvent logs and task miningSets the hours an agent can save.
Waiting timeEvent logsShows whether faster work shortens the case at all.
Rework and variantsEvent logsExceptions are where agents stall, and where review effort goes.
Hand-offsObject-centric logsEach hand-off is a queue, and a place where context gets lost.
Decision rightsInterviews and policiesSome decisions must stay with a named role.
Data accessInterviews and system ownersAn agent can use only the data it’s allowed to read.
Cost of an errorInterviews, risk and control documentsDecides how much review a step needs.
Figure 1. What to record for each step of a workflow before you place an agent in it.

The last three measures come from people and policies, not logs. Business process management researchers call the rules that bound what an agent may see and do in a process its frame. A frame includes roles, segregation of duties and approval rules, such as four-eyes checks that need a second person’s approval. It also limits the data an agent can see [10].

Automation vendors now offer this map as an input for agents. In preview, Microsoft lets a Copilot Studio agent query activity metrics such as case frequency, waiting time and rework from its process mining [11]. UiPath’s Cartographer builds a map of how a process runs, with its exceptions and judgement calls, from documents and a guided conversation. UiPath presents it as the starting point for building agents [12]. Neither publishes evidence that agents placed with these tools return more, so use them to build the map, not to choose the steps.

Six traits of a step an agent does well

Research on agent performance points to six traits of a step that an agent does well (Figure 2). As a rule of thumb, a step with all six is a candidate for the agent to do. A step that lacks one or two is usually a place where the agent helps a person instead.

  • Output you can check

    Can a person or a test confirm the result quickly? One benchmark, τ-bench, has agents serve simulated customers, and scores each agent by comparing the final state of the database with the goal [13]. In OpenAI’s GDPval, GPT-5 produced deliverables about 90 times faster than experts. Counting an expert’s review, and the expert redoing tasks where the output fell short, OpenAI estimates it was only 1.12 to 1.39 times faster [14].

  • Short enough to finish reliably

    On METR’s software tasks, the task length an agent completes 80% of the time is four to six times shorter than the length it completes half the time [15]. Anthropic used Claude to classify its own API traffic, which mostly comes from enterprises. By Claude’s estimate, success falls from about 60% on tasks under an hour to about 45% on tasks that would take people five hours or more [16].

  • Tools the agent controls

    In a telecom support benchmark, success fell by 18 to 25 percentage points when a simulated customer also had to take actions. In that mode, success approached zero for tasks that needed more than seven actions [17]. In a simulated company, the best agent completed 41% of tasks on a project tracker and 13% on an office-documents platform [18].

  • Mistakes you can undo

    METR rates tasks on 16 factors of messiness, such as mistakes that can’t be reversed or no way to check the result. At the same task length, each extra factor lowers success by about 8 percentage points [15].

  • Inputs that are fully specified

    GDPval’s tasks were precisely specified and left out work that needs extensive tacit knowledge. When GPT-5 was given under-specified versions of the same tasks, it did worse [14].

  • Volume, and people who want the help

    An AI assistant raised the issues that customer-support staff resolved per hour by 15% on average, and by about 30% for less experienced staff [20]. In a Stanford study, workers favoured automating about 46% of their tasks [21].

Figure 2. Six traits of a step that an agent does well, with the evidence for each.

Score steps, not whole jobs. In mid-2026 the best agent completed about 16% of real freelance projects to a standard a client would accept [19]. Single steps inside those projects are shorter, and agents succeed more often on shorter tasks [15][16].

Score them against real cases, too, not against the step’s description. A step just outside an agent’s ability can look like one inside it. In a field experiment with 758 BCG consultants, those using AI completed 12.2% more tasks and finished 25.1% faster on tasks inside the AI’s ability. On a task chosen to sit just outside it, they were 19 percentage points less likely to reach a correct answer [22].

Decide who does each step

For each step, choose one of four answers:

  • The agent does it, and a person reviews it. The step has all six traits.
  • The agent helps a person. The step lacks a trait or two, but the agent can gather, draft or summarise.
  • A person decides. Decision rights or the cost of an error keep the step with a named role.
  • Ordinary code does it. The step follows fixed rules, so code without a model is cheaper and more predictable.

Doing and helping are different jobs. In 200,000 conversations with Microsoft’s Bing Copilot, the user’s goal and the AI’s own action were different activities in 40% of conversations [23]. Label each step as doing or helping.

Ask the people who do the work, too. Stanford’s WORKBank study compares what workers want automated with what experts rate agents able to do. It sorts tasks into the following four zones:

  • Green light. Workers want automation, and agents are able.
  • Red light. Agents are able, but workers don’t want it.
  • Research and development. Workers want it, but agents aren’t able yet.
  • Low priority. Workers don’t want it, and agents aren’t able.

Of the tasks that start-ups funded by Y Combinator address, 41% fall in the red-light or low-priority zones [21]. You can still automate a red-light step, but plan the change with the team.

In regulated work, some steps stay with people. KPMG surveyed 100 senior leaders at US banks in the third quarter of 2026. Of those, 75% said they don’t let agents access sensitive data without human oversight, up from 51% in the second quarter [24]. Octopus Energy’s email agent answers only routine enquiries, and vulnerable customers, sensitive cases and complex complaints always go to staff. In a three-month trial on about 4% of its UK customer emails, Octopus reports a 76% customer satisfaction score for the agent’s replies. Comparable replies from staff scored 72% [25]. Aviva’s claims journey defaults to a person for personal-injury claims [26]. A clear boundary is what lets an agent take the rest of the volume.

A worked map: a hypothetical KYC review

This section applies the method to a hypothetical workflow: the periodic know-your-customer (KYC) review of corporate customers at a mid-sized bank. Every number is an assumption for illustration, not client data. The bank reviews about 1,000 corporate customers a month. It has 22 analysts, who spend 6 hours a day on reviews, and 5 compliance approvers, who spend 3 hours a day approving them. Figure 3 is the map.

Step and who does itWork per reviewWaitingReworkCost of an errorWho decides
Open the reviewRules in the case systemNoneNoneNoneLowRisk policy sets the date
Request documentsAnalyst20 minCustomer: 6 working days on average35% of replies miss a document: 10 min and about 4 days moreMedium: a slower reviewAnalyst
Check documentsAnalyst35 minQueue for an analystRareHigh: a bad document stays on fileAnalyst
Clear screening alertsAnalyst40 min, on the 30% of reviews with alertsQueue for an analystRareSevere: a missed sanctions matchAnalyst, with escalation
Trace ownershipAnalyst40 min, varying widelyQueue for an analystSome, when registries disagreeHigh: a wrong beneficial ownerAnalyst
Write the risk memoAnalyst30 minQueue for an analystPart of returned reviewsMedium: an unsupported ratingAnalyst
Approve the ratingCompliance approver15 minQueue for an approver, open 3 hours a day10% come back for a 20 min fixSevere: the bank’s risk decisionApprover, never the analyst
Update recordsIntegration with core bankingNoneNoneNoneLowRules
Figure 3. A hypothetical periodic KYC review, mapped step by step. Work times are averages; the assumptions are hypothetical.

The map shows where effort and time go, and they go to different places. Analysts spend about 2.4 hours on each review, most of it checking documents, tracing ownership and writing the memo. But most of a review’s elapsed time is waiting. When part 2 simulates this map, a review takes 8.4 working days on average, and 7.4 of them are spent waiting for the customer.

Figure 4 scores the six steps that involve judgement against the six traits. Opening and updating the review are rules, so they stay with ordinary automation.

StepOutput you can checkShort enough to finishTools the agent controlsMistakes you can undoFully specified inputsVolume, and people who want helpAnswer
Request documentsYesYesYesYesYesYesAgent does it, analyst reviews
Check documentsYesYesYesYesYesYesAgent does it, analyst reviews
Clear screening alertsPartlyYesPartlyNoPartlyPartlyAgent helps the analyst
Trace ownershipPartlyNoPartlyYesNoYesAgent helps the analyst
Write the risk memoYesYesYesYesYesYesAgent does it, analyst reviews
Approve the ratingNoYesNoNoPartlyYesA person decides
Figure 4. The six judgement steps scored against the six traits, with the answer for each.

Three steps have all six traits: requesting documents, checking them and writing the memo. The agent does those, and an analyst reviews every output. Clearing screening alerts fails because a mistake can’t be undone, and tracing ownership fails because the work is long and depends on judgement. For those two, the agent helps the analyst. Approval stays with a person who didn’t prepare the review.

The map also shows that the steps that save hours and the steps that save days are different. Checking documents and writing the memo hold much of the analysts’ time. The customer’s reply holds most of the elapsed time, and the only agent step that touches it is the document request. A request that’s complete the first time means fewer second rounds. Part 2 measures both.

From the map to a shortlist

To turn a map into a shortlist, follow these steps:

  1. Record every step’s owner, work, waiting, rework, decision rights, data access and cost of an error.
  2. Take out the steps that follow fixed rules, and automate them with ordinary code.
  3. Score the remaining steps on the six traits, using real cases.
  4. For each step, decide whether the agent does it, helps a person, or leaves the decision to a person.
  5. Rank the candidates by the hours and the days they can return, then test them in a simulation before you build.

Choosing which workflow to map first is a separate decision: start with a clear operating problem, an accountable owner and evidence you can access. Our Discovery work maps workflows like this one with your team. Part 2 takes this map and simulates it, to price each agent step before anything is built.

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. McKinsey, The state of AI in 2026: on the road to ROI. Global survey of 1,719 respondents, May to June 2026; published 25 August 2026.
  2. McKinsey and the GCC Board Directors Institute, The state of AI in GCC countries. 139 respondents, August to September 2025; November 2025.
  3. EY Canada, How agentic AI can improve anti-money laundering investigations. Discussion paper, 2025. Figures are EY’s own, for an unnamed client.
  4. Gartner, Over 40% of agentic AI projects will be cancelled by the end of 2027. Press release, 25 June 2025. A prediction, not a measured outcome.
  5. Gutermuth and others, Efficient and compliant purchase order handling. BPI Challenge 2019 submission; the variant count is for the unfiltered log.
  6. Berti and others, OCEL 2.0 specification for object-centric event logs. arXiv 2403.01975, 2024.
  7. Leno and others, Robotic process mining: vision and challenges. Business & Information Systems Engineering 63(3), 2021.
  8. Bartelheimer and others, Workarounds as a cause of mismatches in business processes. Business & Information Systems Engineering 67(3), June 2025.
  9. Mannhardt and Toussaint, Revealing work practices in hospitals using process mining. Studies in Health Technology and Informatics 247, 2018.
  10. Calvanese and others, Agentic business process management: a research manifesto. arXiv 2603.18916, version 3, April 2026.
  11. Microsoft, Create a Copilot Studio agent with process mining. Preview feature, April 2026.
  12. UiPath, UiPath launches UiPath Cartographer, helping enterprises turn scattered process knowledge into an actionable view of how their business runs. Press release, 23 September 2026.
  13. Yao and others, τ-bench: a benchmark for tool-agent-user interaction in real-world domains. ICLR 2025.
  14. Patwardhan and others, OpenAI, GDPval: evaluating AI model performance on real-world economically valuable tasks. arXiv 2510.04374, October 2025.
  15. Kwa and others, METR, Measuring AI ability to complete long software tasks. NeurIPS 2025; version 4, July 2026.
  16. Anthropic Economic Index report: economic primitives. 15 January 2026. Anthropic’s classification of its own traffic, sampled 13 to 20 November 2025.
  17. Barres and others, τ²-bench: evaluating conversational agents in a dual-control environment. arXiv 2506.07982, June 2025.
  18. Xu and others, TheAgentCompany: benchmarking LLM agents on consequential real-world tasks. arXiv 2412.14161, version 3, September 2025.
  19. Center for AI Safety, A significant increase in digital labor automation. Remote Labor Index update, 1 July 2026.
  20. Brynjolfsson, Li and Raymond, Generative AI at work. Quarterly Journal of Economics 140(2), 2025.
  21. Shao and others, Future of work with AI agents: auditing automation and augmentation potential across the U.S. workforce. arXiv 2506.06576, version 3, February 2026.
  22. Dell’Acqua and others, Navigating the jagged technological frontier. Harvard Business School working paper 24-013; field experiment with 758 BCG consultants.
  23. Tomlinson and others, Microsoft Research, Working with AI: measuring the applicability of generative AI to occupations. arXiv 2507.07935, version 6, December 2025.
  24. KPMG, AI Quarterly Pulse Survey, Q3 2026. US leaders at companies with $1 billion or more in revenue, including 100 in banking; 24 September 2026.
  25. Octopus Energy, AI trial wins customer approval. Press release, 7 July 2026. Octopus’s own figures.
  26. McKinsey, Aviva: rewiring the insurance claims journey with AI. Case study by the consultancy that did the work.
Talk to us