Workflow Optimization Framework

Why This Matters Now¶

Enterprises have crossed the adoption threshold for AI but not the value threshold. By 2025, 78% of organizations reported using AI in at least one business function, up from 72% in early 2024 and 55% in 2023 (McKinsey, 2024). Yet adoption has badly outrun realized return. In the same survey, only 39% of organizations could attribute any enterprise-level EBIT impact to AI — and most of those put the contribution at less than 5% of EBIT (McKinsey, 2025). The work has been deployed almost everywhere and is paying off almost nowhere at scale.

The gap is not a technology gap. It is a workflow gap. AI tools get bolted onto tasks while the surrounding process — the sequence of steps, handoffs, decisions, exceptions, and systems that actually constitutes the work — is left untouched. A faster draft of an email inside a process that still routes through four manual approvals does not move the P&L. The organizations that capture value are the ones that treat the workflow, not the tool, as the unit of optimization: they find the workflows worth changing, decide how far to push automation on each, and instrument the result so the value is visible.

This page is the organizing framework for that work. It covers how to identify and prioritize AI-enabled workflows, introduces the four levels of workflow AI integration (Assist → Automate → Augment → Agent), lays out a process discovery methodology, frames how to size and defend the investment, and shows how workflow optimization depends on the data and platform tracks. It is the what to optimize and in what order layer; the sibling pages tell you how far to push each workflow, what tools to use, and how to run the program.

The Value Gap: Why Most Automation Never Pays Off¶

Before prioritizing workflows, it helps to understand precisely how value leaks out, because the failure modes are remarkably consistent across two decades of automation waves — and they shape the framework that follows.

The headline number for the current wave is stark. MIT's Project NANDA reviewed more than 300 AI initiatives and concluded that roughly 95% of enterprise generative-AI pilots were delivering no measurable impact on the P&L, with only about 5% achieving rapid value (MIT NANDA, 2025). This is not an isolated finding. Gartner predicted that at least 30% of generative-AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs, and unclear business value (Gartner, 2024). BCG found that 74% of companies struggle to achieve and scale value from AI, with only about 4% generating substantial value across functions (BCG, 2024).

~95% of enterprise generative-AI pilots were delivering no measurable impact on the P&L (MIT NANDA, 2025).

None of this is new. The previous automation wave — robotic process automation (RPA) — failed in the same shape. EY estimated that 30% to 50% of initial RPA projects failed (EY, 2016). A survey of 400 senior executives found that 38% said RPA projects failed because they were too complex and 30% said they failed because the process or the tools were not understood (ABBYY, 2020). And scaling was vanishingly rare: Deloitte's global RPA survey found that only about 3% of organizations had scaled their digital workforce beyond 50 bots (Deloitte, 2018). The pattern repeats because the root cause is the same: organizations automate a process they have not actually understood, standardized, or instrumented. The technology changes; the discipline that is missing does not.

The practical lesson for prioritization is that the binding constraint is rarely model capability — it is process readiness and the path to scale. A workflow that is poorly understood, full of undocumented exceptions, dependent on dirty data, or lacking an owner will not produce value no matter how capable the AI. The framework below is built to surface those conditions before the investment, not after.

How to Identify and Prioritize AI-Enabled Workflows¶

Start by separating tasks, processes, and workflows¶

The first discipline is to be precise about the unit you are optimizing, because conflating these three is a primary source of disappointment. APQC's Process Classification Framework offers a clean taxonomy: a five-level hierarchy running Category → Process Group → Process → Activity → Task, where a task is the atomic "element of work that goes into executing an activity" (APQC, PCF). The PCF is deliberately a classification taxonomy, not a sequence model.

For our purposes:

A task is the atomic unit — drafting a paragraph, extracting a field from an invoice, classifying a ticket. AI tools are very good at tasks.
A process is the named, repeatable end-to-end objective — "procure-to-pay," "claims adjudication," "employee onboarding."
A workflow is the orchestration layer the PCF omits: the actual sequence of tasks, handoffs between people and systems, decision branches, and exception paths through which a process runs.

The reason this matters: most AI value claims are measured at the task level (a 25% faster draft), but P&L value is realized at the workflow level (a cycle that completes faster, with fewer errors, at lower cost). McKinsey's work reinforces the point — its analysis scores work activities, not occupations, finding that current technologies "have the potential to automate work activities that absorb 60 to 70 percent of employees' time today" (McKinsey, 2023). A 60–70% activity-automation ceiling does not translate into 60–70% of jobs or even of any single workflow; it accumulates only when the surrounding workflow is redesigned to capture the freed-up time.

Prioritize on two axes: value and feasibility¶

With the right unit in view, prioritize candidate workflows on two independent axes.

Value — how much is this workflow worth changing? Size it on:

Volume and frequency — how many times the workflow runs per period. High-volume, high-frequency workflows compound small per-run gains into material totals.
Time and cost per run — labor hours, cycle time, and cost-per-transaction consumed.
Error and rework rate — the cost of mistakes, exceptions, and downstream correction.
Strategic weight — customer-facing impact, risk exposure, or bottleneck status (a workflow that gates others is worth more than its own cost suggests).

Feasibility — how ready is this workflow to be changed? Drawing on the canonical automation-screening criteria from the research literature — standardization, rule-based logic versus judgment, availability of digital input, process stability, and exception rate (Wellmann et al., 2020) — assess:

Standardization and stability — is the workflow consistent, or does every run differ?
Rule-based vs. judgment — deterministic, rule-bound steps are far easier than open-ended judgment (though generative AI extends the frontier into language-based judgment, covering activities at roughly 25% of work time per (McKinsey, 2023)).
Digital, accessible input — is the data the workflow consumes already digital, structured, and reachable? This is where the data-readiness track becomes a hard dependency.
Exception density — workflows that are 80% standard and 20% chaotic exceptions are deceptively hard; the exceptions, not the happy path, determine effort.

Plot candidates on a value-versus-feasibility grid. The sequencing rule is unglamorous but reliable: high-value, high-feasibility workflows first — these fund the program and build credibility — followed by high-value, lower-feasibility workflows that justify the harder data and process work. Low-value workflows are deprioritized regardless of how easy they are to automate, which is the single most common prioritization error: automating what is easy rather than what matters.

Prioritization quadrant	Value	Feasibility	Action
Quick wins	High	High	Do first — fund the program, prove the model
Strategic bets	High	Low	Sequence next — invest in data/process readiness to unlock
Fill-ins	Low	High	Only if near-zero marginal effort; beware the "easy but pointless" trap
Avoid	Low	Low	Do not pursue

The Four Levels of Workflow AI Integration¶

For each prioritized workflow, the next decision is how far to push the integration. A single binary "automate or not" obscures a spectrum. This framework uses four levels — Assist → Automate → Augment → Agent — each representing a greater degree of AI autonomy and a different risk, governance, and value profile.

Level	What it means	AI's role	Human's role
Assist	AI helps a human do a step faster or better	Suggests, drafts, summarizes	Performs and decides; does every step
Automate	AI executes a defined, bounded step end-to-end	Executes deterministic, rule-bound steps	Designs, supervises, handles exceptions
Augment	AI and human jointly own the workflow, each on their strengths	Handles routine volume and analysis	Owns judgment, escalations, edge cases
Agent	AI plans and executes a multi-step workflow toward a goal	Plans, acts across systems, self-corrects	Sets goals, sets guardrails, audits outcomes

The evidence base spans the spectrum. At the Assist level, controlled studies are strong: consultants using GPT-4 completed 12.2% more tasks, 25.1% faster, with results rated over 40% higher in quality (Dell'Acqua et al., 2023), and developers using GitHub Copilot completed a coding task 55.8% faster than a control group (Peng et al., 2023). At the Augment level, a field study of customer-support agents found a 14% average increase in issues resolved per hour, rising to 34% for novice agents (Brynjolfsson et al., 2023). The Agent level is where the market is now racing — Gartner predicts that 40% of enterprise applications will feature task-specific AI agents by 2026, up from less than 5% in 2025 (Gartner, 2025) — and also where governance risk concentrates, because the AI is acting across systems rather than suggesting to a human.

Choosing the right level is itself a value-and-risk judgment: pushing a high-judgment, high-stakes workflow straight to Agent is how pilots become incidents, while leaving a high-volume deterministic workflow at Assist leaves most of the value on the table. The full decision model — how to choose a level per workflow, the governance implications of each, and how to progress a workflow up the levels over time — is detailed in the four levels of workflow AI integration.

Process Discovery Methodology¶

Prioritization is only as good as the picture of how work actually flows — and that picture is almost never the one in the process documentation. Process discovery is the discipline of reconstructing the real workflow before you change it. There are three complementary methods, best used together.

1. Process mapping (the intended process). Workshops, interviews, and BPMN diagramming capture how people believe the workflow runs. This is fast and essential for shared understanding, but it is subjective, a snapshot in time, and systematically blind to the variants and rework that nobody admits to.

2. Process mining (the real end-to-end process). Process mining reconstructs the actual execution from event-log data — the timestamps and activity records left in ERP, CRM, ticketing, and other transactional systems. Academically, it "aims to discover, monitor and improve real processes by extracting knowledge from event logs," via three operations: discovery, conformance checking, and enhancement (IEEE, 2012). It exposes the variants, bottlenecks, and rework loops that mapping misses. Its value is well evidenced: a Forrester study of a composite Celonis customer found 383% three-year ROI with payback in under six months, including one workflow where no-touch processing rose 53 percentage points (from 33% to 86%) (Forrester, 2025). Named results are larger still — Deutsche Telekom reported over €66M saved in procure-to-pay and GE Healthcare a $1.3B free-cash-flow increase in a single year (Celonis) — though as single-vendor figures these are illustrative rather than benchmarks. The catch is universal: mining is only as good as the event log, and extracting, cleaning, and mapping that log (case ID, activity, timestamp) is the bulk of the work — another reason data readiness gates the whole effort.

3. Task mining (the desktop reality). Where process mining draws on transactional system logs, task mining captures user-interaction data — clicks, keystrokes, copy-paste, application time — to reconstruct how work is performed at the desktop, below the level any system log records (Celonis, What is Task Mining). It is how you find the invisible swivel-chair work between systems that is often the richest automation candidate.

The methodology in practice: map to align, mine to verify, task-mine to find the hidden work — then prioritize against the value/feasibility grid above using evidence, not anecdote. Discovery is where the "process not understood" failure mode (the 30% of RPA failures cited earlier) is defeated.

ROI Framing for Workflow Investment¶

A workflow investment has to be sized and defended in financial terms, both to win funding and to make value realization measurable rather than aspirational.

Sizing the value¶

Build the value case from the workflow's own measurable mechanics, not from headline industry figures:

Labor/capacity freed — FTE-hours saved per run × runs per period, valued at loaded cost. McKinsey's value-by-function work is a useful sanity check on magnitude: it estimates generative AI could add $2.6–4.4 trillion annually across use cases, with productivity gains of 30–45% of function cost in customer operations and 20–45% in software engineering (McKinsey, 2023). Treat these as ceilings for whole functions, not promises for one workflow.
Cycle-time reduction — faster completion converts to revenue (faster cash, better experience) or cost (less work-in-progress). Klarna reported its AI assistant cut customer-service resolution time from 11 minutes to under 2 minutes while handling the equivalent of 700 FTEs of work (Klarna, 2024) — a self-reported example, but a clear illustration of cycle-time value.
Error and rework reduction — fewer exceptions and corrections. Quantify this from your own discovery data; credible cross-industry percentages for error reduction are genuinely scarce in primary research, so size it from your measured baseline rather than an imported figure.

Framing the return — and the payback reality¶

On the optimistic side, IDC's analysis found organizations realizing an average of $3.70 in return per $1 invested in generative AI, rising to $10.30 for the top adopters (IDC, 2024), and a Google Cloud study reported 74% of executives achieving ROI within the first year (Google Cloud, 2025). Both are sponsored studies and skew toward leaders — useful as directional ceilings, not base cases.

The sober counterweight matters more for credible planning. Deloitte found that organizations scaling automation beyond piloting achieved an average 32% cost reduction in the areas targeted (up from 24% in 2020) — but also that 36% had not calculated their payback at all, and that average payback for pilot-stage intelligent automation had lengthened to about 22 months, up from 16 months in 2020, as the easy wins were exhausted (Deloitte, 2022). The honest ROI framing, therefore, is a barbell: quick wins can pay back in well under a year, but the strategic, harder workflows carry multi-year paybacks and demand the data and platform investment to be counted in the case.

The discipline that separates the 39% (McKinsey, 2025) who see EBIT impact from the rest is instrumenting the baseline before you start — current cycle time, cost-per-transaction, error rate, and volume — so the delta is measurable. Value that is not measured against a baseline is value that, by Deloitte's evidence, a third of organizations never actually calculate. The measurement methodology itself is the subject of Measurement & Value Realization.

How Workflow Optimization Connects to the Other Tracks¶

Workflow optimization is not a standalone effort; it sits on top of two enabling tracks and feeds two governing ones. Treating it as independent is itself a failure mode.

Data readiness is a hard dependency. Every method above — process mining's event logs, an AI step's input data, an agent's context — depends on data that is accessible, structured, and trustworthy. Gartner named poor data quality among the top reasons projects are abandoned after POC (Gartner, 2024). A workflow's feasibility score is, in large part, a data-readiness score. See Data Readiness.
The platform is how it runs at scale. Moving a workflow from a clever pilot to reliable production — orchestration, integration across systems, monitoring, evaluation — is platform work. The 3–4% RPA scaling rate (Deloitte, 2018) and the pilot-to-production gap are platform failures as much as anything. See Technology Architecture & Platform.
Governance gates the higher levels. As workflows move up the Assist → Agent spectrum, the AI takes more autonomous action and governance risk rises sharply. Any workflow that automates a decision-bearing or customer-facing step falls under AI Governance & Risk.
Measurement closes the loop. Sizing value up front and proving it after deployment is measurement-track work; without it, ROI claims stay aspirational. See Measurement & Value Realization.

For the tools that support each method here — mapping, mining, automation platforms, and agent frameworks — see the tooling landscape. To run an actual program end to end, see the practitioner guide, and to score your own portfolio, use the workflow maturity and opportunity assessment.

Common Failure Modes¶

The research above points to a recurring set of traps. Each maps to a discipline in this framework.

Automating the easy, not the valuable. Chasing high-feasibility, low-value workflows because they are simple. This is why prioritization runs on two axes, not one — feasibility alone is a trap.
Optimizing the task, ignoring the workflow. Deploying a faster tool inside an unchanged process. Task-level gains do not reach the P&L until the surrounding workflow is redesigned; this is the core of the 39%-EBIT-impact gap.
Automating a process you never understood. Skipping discovery and automating the documented (and fictional) happy path while exceptions break in production — the 30% of RPA failures attributed to the process not being understood (ABBYY, 2020).
Pilot purgatory. Running endless POCs with no path to production. Roughly 95% of generative-AI pilots showed no P&L impact (MIT NANDA, 2025) largely because scaling was never designed in — a platform and ownership failure, not a model one.
Pushing the autonomy level too far, too fast. Sending a high-judgment workflow straight to Agent without the governance to contain it — turning pilots into incidents.
No instrumented baseline. Launching without measuring current cycle time, cost, error rate, and volume — leaving value uncalculated, as 36% of organizations do (Deloitte, 2022).
Treating the program as standalone. Ignoring the data, platform, and governance dependencies that determine whether anything scales.

Closing Checklist¶

Before committing to a workflow optimization investment, confirm:

[ ] Unit is clear — you are optimizing a workflow (orchestration), not just a task, and you know which process it belongs to.
[ ] Discovery is evidence-based — the real workflow has been mapped, mined, and task-mined; you are not working from the documented fiction.
[ ] Prioritized on two axes — the workflow ranks high on both value and feasibility, or is a deliberate strategic bet with a readiness plan.
[ ] Integration level chosen deliberately — the Assist → Automate → Augment → Agent level matches the workflow's judgment and risk profile.
[ ] Baseline instrumented — current cycle time, cost-per-transaction, error rate, and volume are measured, so the delta will be provable.
[ ] ROI case is honest — value sized from the workflow's own mechanics, with realistic (often multi-year) payback for the harder bets, not headline ceilings.
[ ] Dependencies confirmed — data is accessible and trustworthy (see Data Readiness), there is a path to production at scale (see Technology Architecture & Platform), governance covers the autonomy level (see AI Governance & Risk), and measurement is wired in (see Measurement & Value Realization).
[ ] Ownership assigned — a named owner accountable for the workflow's outcome, not just its launch.

Sources¶

McKinsey — The State of AI: How Organizations Are Rewiring to Capture Value, 2024. 78% of organizations now use AI in at least one business function (up from 72% in early 2024 and 55% in 2023). View source · verified 2026-06-22 · primary
McKinsey — The State of AI, 2025. About 39% of organizations report any enterprise-level EBIT impact from AI; most of those say less than 5% of EBIT is attributable to AI use. View source · verified 2026-06-20 · primary
MIT Project NANDA — The GenAI Divide: State of AI in Business 2025, 2025. Just 5% of integrated AI pilots are extracting millions in value, while the vast majority remain stuck with no measurable P&L impact. View source · verified 2026-06-20 · ⚠ secondary mirror
Gartner — Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025, 2024. At least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, escalating costs or unclear business value. View source · verified 2026-06-20 · primary
Boston Consulting Group (BCG) — AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value, 2024. Seventy-four percent of companies have yet to show tangible value from their use of AI; just 4% of companies have developed cutting-edge AI capabilities across functions and consistently generate significant value. View source · verified 2026-06-20 · primary
EY — Get ready for robots: Why planning makes the difference between success and disappointment, 2016. as many as 30 to 50% of initial RPA projects fail. View source · verified 2026-06-20 · primary
ABBYY — ABBYY survey of 400 senior executives on RPA (reported via CIO Dive), 2020. 38% of executives say projects fail because they are too complex; three in 10 projects fail because the intended automation processes are not understood or there's not enough understanding of the underlying automation tools. View source · verified 2026-06-20 · ⚠ secondary mirror
Deloitte — The robots are ready. Are you? Untapped advantage in your digital workforce (Global RPA Survey), 2018. only 3% of organizations have managed to scale RPA to a level of 50 or more robots. View source · verified 2026-06-20 · ⚠ secondary mirror
APQC — Understanding the Elements of APQC's Process Classification Framework (PCF). Level 5-Task: an element of work that goes into executing an activity (within the hierarchy Category, Process group, Process, Activity, Task). View source · verified 2026-06-20 · primary
McKinsey (McKinsey Global Institute) — The economic potential of generative AI: The next productivity frontier, 2023. Current generative AI and other technologies have the potential to automate work activities that absorb 60 to 70 percent of employees' time today. View source · verified 2026-06-20 · ⚠ secondary mirror
Wellmann, Stierle, Dunzer & Matzner (arXiv) — A framework to evaluate the viability of robotic process automation for business process activities, 2020. Ideal candidate processes for automation must be standardized; the process follows a rule-based flow; candidate processes suited for RPA show little or no amount of exceptions; the data needs to be in a structured and digital form. View source · verified 2026-06-20 · primary
McKinsey (McKinsey Global Institute) — The economic potential of generative AI: The next productivity frontier, 2023. the acceleration in the potential for technical automation is largely due to generative AI's increased ability to understand natural language, which is required for work activities that account for 25 percent of total work time. View source · verified 2026-06-20 · ⚠ secondary mirror
Dell'Acqua et al. — Navigating the Jagged Technological Frontier (HBS Working Paper 24-013), 2023. In a pre-registered experiment with 758 consultants, those using AI inside the frontier completed 12.2% more tasks, 25.1% more quickly, with more than 40% higher quality; on a task outside the frontier AI users were 19 percentage points less likely to be correct. View source · verified 2026-06-20 · primary
Peng et al. — The Impact of AI on Developer Productivity: Evidence from GitHub Copilot, 2023. The treatment group with access to the AI pair programmer completed the task 55.8% faster than the control group; developers with GitHub Copilot took on average 1 hour 11 minutes versus 2 hours 41 minutes, across 95 professional programmers. View source · verified 2026-06-20 · primary
Brynjolfsson, Li & Raymond — Generative AI at Work (NBER Working Paper 31161), 2023. Access to the tool increases productivity, measured by issues resolved per hour, by 14% on average, including a 34% improvement for novice and low-skilled workers but with minimal impact on experienced and highly skilled workers. View source · verified 2026-06-20 · primary
Gartner — Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, 2025. Forty percent of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% today. View source · verified 2026-06-20 · primary
IEEE Task Force on Process Mining — Process Mining Manifesto, 2012. The idea of process mining is to discover, monitor and improve real processes by extracting knowledge from event logs readily available in today's systems; the three main types of process mining: (a) discovery, (b) conformance checking, and (c) enhancement. View source · verified 2026-06-20 · primary
Forrester Consulting (commissioned by Celonis) — The Total Economic Impact of Celonis, 2025. an ROI of 383%; [orders processed] without any human intervention increased from 33% to 86%. View source · verified 2026-06-20 · primary
Celonis — Customer Stories (Deutsche Telekom procure-to-pay; GE Healthcare cash flow). Deutsche Telekom saved over EUR 66M by maximizing the execution capacity of their Procure-to-Pay process; GE Healthcare increased free cash flow by $1.3 billion in just one year. View source · verified 2026-06-20 · ⚠ secondary mirror
McKinsey (McKinsey Global Institute) — The economic potential of generative AI: The next productivity frontier, 2023. the direct impact of AI on the productivity of software engineering could range from 20 to 45 percent of current annual spending on the function; [the analyzed use cases total] $2.6 trillion to $4.4 trillion annually. View source · verified 2026-06-20 · ⚠ secondary mirror
Klarna — Klarna AI assistant handles two-thirds of customer service chats in its first month, 2024. It is doing the equivalent work of 700 full-time agents; customers now resolve their errands in less than 2 mins compared to 11 mins previously. View source · verified 2026-06-22 · primary
IDC — The Business Opportunity of AI (IDC InfoBrief, sponsored by Microsoft), 2024. For every $1 a company invests in generative AI, the average ROI is 3.7x. View source · verified 2026-06-20 · ⚠ secondary mirror
Google Cloud — The ROI of AI 2025, 2025. 74% of executives report achieving ROI within the first year. View source · verified 2026-06-20 · primary
Deloitte — Automation with intelligence (Global Intelligent Automation survey), 2022. organisations that moved beyond piloting intelligent automation have achieved an average cost reduction of 32 per cent, up from 24 per cent in 2020; over a third (36 percent) have not calculated the payback period; the average payback period increased from 16 months in 2020 to 22 months. View source · verified 2026-06-20 · primary