Process Automation Before AI: Fix the Basics That Block Value

Adoption statistics can flatter. Eurostat reports that in 2025, 17% of small, 30.36% of medium and 55.03% of large EU enterprises used AI, and that 26 EU countries recorded higher shares than in 2024. In the UK, the Office for National Statistics reports that AI adoption among businesses increased from around 12% to around 35% between late 2023 and June 2026. Take-up is real and rising.

Depth is another matter. The ONS notes that the depth of AI use rose only modestly, from around 1.4 to around 1.6 over the same period, so most firms use a small number of AI technologies in a limited way. McKinsey's global survey says 44% of respondents report that AI is scaling across their enterprise, yet only 37% attribute at least some EBIT impact to AI use, and about 6% qualify as high performers with an EBIT impact of 5% or more.

For a COO this means the question has moved from whether to adopt AI to why value is not following adoption. The honest answer in most operations is that AI has been applied to processes that are not ready: unclear ownership, inconsistent inputs, poorly defined exceptions and weak measurement. Those weaknesses are tolerable when humans compensate informally, and become costly when an automated system amplifies them at speed and scale.

Context matters for expectations too. Executives often see AI results from the minority of high performers and assume they are typical. The McKinsey data suggest otherwise: sustained financial impact remains concentrated among a small group that has changed how work is done. A realistic plan assumes your own results will track your process maturity, and sets milestones accordingly, with early wins in the processes that are already the cleanest.

There is also a competitive angle. Firms that find the discipline to repair basics first tend to compound: each fixed process generates cleaner data, which makes the next automation cheaper and the next AI use case more reliable. Firms that skip the basics accumulate pilots that do not connect. Over two or three years the difference shows up in unit cost and in the speed with which new capabilities can be rolled out.

What to do: before approving the next AI investment, ask for a one-page readiness check of the process it will touch, covering ownership, volumes, exception rates, data sources and current cycle time. If the sponsor cannot answer, that is the first piece of work. The readiness check costs little and often reveals that conventional automation or simple redesign delivers most of the benefit without any AI at all.

McKinsey's State of AI research offers the clearest evidence on what separates leaders from the rest. It reports that nearly three-quarters of AI high performers say they have fundamentally redesigned workflows, up from 55% the year before, while around one-quarter of other respondents report doing the same. High performers do not simply deploy tools into existing steps; they change the steps.

Celonis, a process-intelligence vendor, surveyed 1,620 business leaders for its 2025 Process Optimization Report. It found that 89% say AI needs the context of how their business runs, 81% plan to use AI specifically to improve business processes in the next 12 months, and 58% worry that process shortcomings could limit AI's contribution. The source has a commercial interest in process tooling, so treat the figures as directional, but they align with the independent McKinsey finding.

The implication is that the workflow, not the algorithm, is the unit of value. An AI assistant that drafts a reply faster saves little if the reply waits three days for approval. A model that classifies invoices accurately adds nothing if exceptions still travel through email. Value arrives when the whole path from trigger to outcome is redesigned, with AI as one component, and with clear roles for the people who handle what the system cannot.

Process intelligence and process mining tools can speed up this work by reconstructing real flows from system logs. They are useful for finding where cases wait, loop back or take unexpected routes. They are not a substitute for talking to the people who run the process, since logs miss work done in email, chat and spreadsheets. Combine both views and validate the redesign with a small pilot before scaling.

Choose the first processes carefully. The best candidates are high in volume, painful to the customer or the cost base, reasonably stable in their rules and owned by a manager who wants change. Avoid starting with the most politically complex process because failure there will taint the whole programme. A visible, uncontroversial success creates the permission to tackle harder ones, and gives the team practice in the method.

What to do: pick two or three end-to-end processes where cost, delay or error is visible, such as order-to-cash, procure-to-pay or customer onboarding. Map the real flow, including workarounds, and redesign it on paper before choosing technology. Set targets for cycle time, touch time and exception rate. Any AI or business process automation tool that does not move one of those numbers should be questioned.

Gartner reports that 63% of organisations either do not have or are unsure whether they have the right data management practices for AI, based on a survey of data management leaders. It predicts that through 2026 organisations will abandon 60% of AI projects unsupported by AI-ready data. The prediction is a forecast, not a measurement, but the survey result shows how widespread the weakness is.

AI-ready data differs from traditional reporting data. It must be accessible where the process runs, consistent across systems, labelled with enough context to be interpreted, and governed so that its quality is monitored. McKinsey's supply chain work identifies three main ingredients for resilient operations: end-to-end visibility, high-quality master data and effective scenario planning. The same ingredients apply to any process that AI will touch.

Master data is the quiet culprit in many failed projects. Duplicate customer records, inconsistent product codes, free-text fields and conflicting supplier identifiers force models and rules to guess. Teams then compensate with manual review, which erodes the time savings that justified the project. Fixing the data once, at source and with an owner, usually pays back across every downstream system, including those that never use AI.

Governance of data should be tied to business outcomes rather than technical hygiene. A monthly data-quality score is more likely to be taken seriously if it is connected to something a business leader cares about, such as invoice error rates, order-delivery accuracy or customer complaints. Linking the two gives owners a reason to act and gives the COO a way to show the value of the work.

Think also about where data lives. In many European groups, critical process data is split across enterprise resource planning systems, local databases, email and spreadsheets maintained at site level. Consolidating everything into one platform is rarely feasible, but agreeing a small set of shared definitions and an integration layer often is. Pragmatic interoperability, rather than a grand migration, gets processes ready for automation in months rather than years.

What to do: appoint business owners for the five or six data domains that most processes depend on, such as customer, product, supplier, employee, asset and contract. Give each a quality target and a monthly report. Before building anything, confirm that the data the process needs is available in usable form. If it is not, the first project is a data project, and it should be planned, funded and measured as such.

Autonomous agents are the current frontier, and they depend even more on well-defined processes than assistants do. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, and estimates that only about 130 of the thousands of agentic AI vendors are real. The warning is as much about expectations as about technology.

The operational reason is straightforward. An agent that acts across systems needs defined inputs, boundaries, escalation rules and audit trails. If the process has undocumented exceptions or inconsistent approvals, the agent either stalls or acts unpredictably. McKinsey's supply chain survey found that three quarters of respondents are planning, blueprinting or piloting AI use cases, but only 19 percent say they are deploying AI tools at scale. Pilots are easy; production is where process quality shows.

For a COO the practical implication is to treat agent readiness as a function of process maturity. A process with clear steps, owners, controls and metrics is a candidate for an agent. A process that depends on tacit knowledge is not, until it has been described and tested. The boundary between what the agent may decide and what a person must approve should be written down in business terms and agreed by risk and compliance.

Cost and vendor claims deserve scrutiny in this area. With so many vendors describing their products as agents, buyers should ask for evidence of production deployments, the share of tasks completed without human intervention and the cost of supervision. Insist on a defined pilot with success criteria and exit terms. If a vendor cannot explain how its product handles exceptions in your process, the demonstration is not enough.

Risk teams should be partners rather than gatekeepers in this work. Agree early what level of autonomy is acceptable for each type of decision, how errors will be detected and who is accountable. Frameworks for model risk, data protection and operational resilience already exist in most regulated sectors and can be adapted. Early involvement shortens approval cycles because concerns are designed out instead of discovered late.

What to do: apply a simple gate. Before any agentic pilot, require a documented process, an owner, a defined set of permitted actions, a human escalation route, a log of decisions and a measurable success criterion. Start with low-risk, reversible tasks. Treat the first agent as a test of your process quality as much as a test of the technology, and expect to find gaps.

Evidence on how businesses use AI points towards operations. The ONS reports that improving business operations is the most common use of AI, reported by over 60% of larger businesses. By technology, large language models were used by 18% of UK businesses with 10 or more employees in June 2026, visual content creation by 16%, data processing using machine learning by 12%, and robotics by 2%. Exhibit 4 sets out the pattern.

The distribution tells a practical story. Text-based tools are widely used because they are easy to start. Machine-learning data processing is less common because it needs reliable data pipelines and well-defined tasks. Robotics is rare because physical automation requires stable processes, capital and integration. The easy tools spread first; the ones that depend on process quality lag. That is a signal about where the constraint lies, not about where the value is.

Conventional workflow automation, including rules engines, robotic process automation, system integration and electronic forms, still addresses a large share of repetitive office work and is more predictable than AI. It handles high-volume, stable, structured steps such as data entry, reconciliations, approvals routing and report generation. Doing this first reduces the noise in the process, creates clean event data and leaves a smaller, more interesting set of cases for AI to handle.

Quick wins build the case for deeper change. A visible reduction in manual hours or error rates in one process gives the programme credibility and frees capacity for the next. Be careful to count savings honestly: time saved is only valuable if it is redeployed, absorbed into growth or removed through planned attrition. Finance should validate the figures so that the programme is trusted when it asks for further investment.

Measure the real cost of ownership too. Automation reduces labour in a process but adds licences, maintenance, monitoring and change management. Bots and integrations break when upstream systems change, and AI models drift. Include these running costs in the business case, with a named owner for maintenance, so that savings reported at launch are still there two years later. Honest accounting protects the credibility of the whole programme.

What to do: list the ten highest-volume processes by hours consumed and error rate. For each, classify steps as rules-based, judgement-based or mixed. Automate rules-based steps with conventional tools first and measure the saving. Direct AI at the judgement-based or unstructured steps, such as reading varied documents or drafting responses, where it can add value on top of a stable base. Review the split every quarter.

Eurostat data show how uneven European adoption is. In 2025 the highest shares of enterprises using AI were in Denmark (42.03%), Finland (37.82%) and Sweden (35.04%), while the lowest were in Romania (5.21%), Poland (8.36%) and Bulgaria (8.55%) (Exhibit 1). Size matters as well: 17% of small enterprises used AI, against 30.36% of medium and 55.03% of large enterprises (Exhibit 2).

These gaps have operational consequences for pan-European groups. A group with sites in both high-adoption and low-adoption countries cannot assume the same digital baseline, skills or supplier ecosystem everywhere. Some plants will already use structured workflows and digital records, while others rely on paper and local spreadsheets. A single group-wide AI programme that ignores this will succeed in some places and stall in others, and the stalled sites will be blamed for a design problem.

Benchmarks should therefore be handled with care. A national average says little about your sector or your process maturity, and a leader's share says little about the quality of its use. The ONS finding that depth of use has barely changed while breadth has tripled is a reminder that the useful benchmark is not who has deployed a tool, but who has changed a process and measured the result.

Regulation adds another layer for European groups. Data protection rules, works-council consultation requirements and sector-specific controls differ by country and shape what can be automated and how decisions must be explained. Build these requirements into the design from the start, and keep a register of automated decisions that affect customers or employees. Doing so avoids rework and makes later audits straightforward.

For smaller organisations the picture is different but the logic is the same. A mid-sized business does not need a large programme office to apply the sequence; it needs one executive who owns the agenda, a short list of processes and a bias towards simple tools. Because smaller firms have fewer systems and fewer layers, they can often standardise faster than large groups, which is an under-used advantage.

What to do: assess process maturity site by site on a simple five-level scale, from ad hoc to instrumented and automated. Sequence the group programme by readiness, starting where processes are stable and data are clean, and use those sites to create templates others can copy. Set expectations that the later sites will need foundation work first, and fund it accordingly.

The practical sequence has five steps. Map the process as it really runs, including exceptions. Standardise it so that the same inputs lead to the same path. Instrument it, so that volumes, delays and errors are measured automatically. Automate the stable, rules-based steps. Then apply AI to the variable remainder. Each step reduces the cost and risk of the next, and the instrumentation generated along the way provides the evidence needed to judge whether AI is working.

Governance should be light but real. Name a business owner for each process, a data owner for each domain and a single accountable executive for the programme. Hold a monthly review of a small scorecard that includes cycle time, cost per transaction, exception rate, automation coverage and realised savings. Retire pilots that do not move a number within a defined period, and publish what was learned. Visible discipline builds confidence among employees and the board.

People are part of the process. Skills are generally built rather than bought, so plan for training alongside the technology. Involve front-line staff in mapping and redesign, because they know the workarounds, and communicate openly about how roles will change. Adoption fails more often through quiet resistance to an unexplained change than through technical shortcomings.

Finally, expect the programme to change its own priorities. As early processes are automated, the bottlenecks move elsewhere, often to approvals, hand-offs between departments or supplier inputs. A standing review that re-ranks processes by remaining cost and delay keeps effort focused where it still matters, and prevents the team from over-polishing areas that have already delivered most of their benefit.

Communicate the logic to the board in plain terms: the investment in basics is not a delay to AI, it is the shortest route to value from it. Show the sequence, the expected savings at each stage and the decision points where AI will be considered. Boards that understand why the early work looks modest tend to support it, and are less likely to push for a disconnected showcase project.

What to do in the next 90 days: choose two processes, complete the readiness check, map and redesign them, fix the critical data issues, automate the stable steps and set up measurement. At day 90, decide where AI should go next, using the measured baseline. Where the programme touches regulated data or employee decisions, involve legal, risk and works-council representatives early, because they shape what may be automated.

Karter McKinley supports AI and Technology implementation for operations leaders: process readiness assessments, workflow redesign, automation roadmaps and the governance that links pilots to measured savings. We work with your teams and your chosen technology vendors, and we focus on outcomes you can verify rather than on tool adoption for its own sake.