Stop Automating Bad Processes: An AI Workflow Triage Method

The automation trap: making failure move faster

AI does not repair a badly designed workflow. It accelerates whatever logic, incentives and omissions already exist. If a supplier invoice currently passes through four inboxes, waits for an absent manager and is rejected because the purchase-order field is unreliable, adding an AI document reader may save three minutes of data entry while leaving the ten-day delay untouched. Worse, faster ingestion can create a larger queue downstream, giving the appearance of productivity while increasing work in progress.

The first discipline of AI workflow design is therefore triage, not tooling. Before selecting a model, classify the operational problem. Is the constraint slow interpretation, missing information, unclear ownership, excessive control, conflicting policy or limited capacity? AI is useful when judgement depends on extracting, comparing, summarising or generating information at scale. It is far less useful when nobody has authority to decide, two departments follow incompatible rules, or a required field is never captured.

A practical test is to remove AI from the proposal and ask what outcome remains. If the business case says only that staff will complete the same steps more quickly, the process may still be structurally unsound. A credible case should identify a measurable constraint: reducing average claims triage from 18 minutes to six, increasing first-pass accuracy from 72% to 90%, or cutting unresolved service cases older than seven days by half. The target exposes whether the technology is addressing the bottleneck or merely decorating it.

Step one: map the work as it happens

Begin with a case-level map built from observation, system records and interviews with the people doing the work. Do not rely on a policy diagram produced for audit. Select 20 to 50 recent cases, including routine, delayed and failed examples, then trace every event from trigger to closure. Record who touched the case, which system they used, what information they needed, how long the action took and how long the item waited. Separating touch time from elapsed time is essential: a refund may require 14 minutes of labour but take nine days to approve.

The map should include unofficial work. Staff frequently copy data into spreadsheets because the case-management system cannot filter it, send screenshots because external users lack access, or keep personal templates because approved wording is obsolete. These workarounds are not peripheral. They are evidence that the formal process does not meet operational needs. An AI assistant trained only on the official procedure will reproduce the gap and may suppress the practical knowledge that keeps work moving.

Mark each handoff and decision with a simple record: input, owner, rule, output and evidence. A customer complaint, for example, may move from a web form to a shared inbox, then to a regional team, legal review and finance. If the regional team reclassifies 35% of cases because the web form uses different categories, that handoff is the first redesign target. Better classification at intake may help; automating legal summaries will not resolve the upstream mismatch.

Step two: diagnose handoffs and queues

Handoffs create delay because context is lost and responsibility becomes ambiguous. For every transfer, ask whether the recipient receives a complete package, whether acceptance is explicit and whether the sender remains accountable after dispatch. Measure bounce rates: the proportion of items returned, reassigned or reopened. A procurement request that is sent back in 28% of cases for missing cost centres is not an approval problem. It is an intake-control problem.

Queue data often changes investment priorities. Suppose a maintenance request spends four hours awaiting AI-assisted fault classification, 36 hours waiting for a supervisor to assign an engineer and two days waiting for a part. Reducing classification to ten minutes produces little customer benefit unless assignment and inventory are redesigned. The relevant measures are end-to-end cycle time, queue age, first-time-right rate and the number of handoffs, not simply model latency or minutes saved per task.

Eliminate avoidable transfers before automating necessary ones. Combine roles where risk permits, route work directly from verified data and establish one accountable case owner. Where handoffs must remain, define a minimum information contract. A sales-to-operations transfer might require customer identity, signed scope, delivery location, service tier and exception notes. AI can then validate completeness or summarise context, but it should not invent absent terms or infer contractual commitments from informal email.

Step three: challenge approvals and controls

Approvals accumulate after incidents, audits and management changes, yet they are rarely removed when circumstances change. Inventory every approval and identify the risk it controls, the value threshold, the decision criteria and the frequency with which the approver rejects or modifies requests. If a director approves 600 monthly expenses and changes fewer than 1%, the control may be ceremonial. Replacing the director with an AI approval agent would automate theatre, not governance.

Use tiered controls instead. Low-value, policy-compliant transactions can pass automatically; unusual or high-impact cases receive human review. A business might auto-approve travel claims below £250 when receipts, dates and permitted categories match, sample 5% for audit, and escalate duplicates, overseas cash payments or policy exceptions. This design reduces delay while preserving oversight. AI can extract receipt data and flag anomalies, but deterministic rules should govern clear financial limits.

Tradeoffs must be explicit. Removing an approval may increase residual risk; retaining it may impose delay, labour cost and customer harm. Quantify both sides. If a second review costs £80,000 annually and prevents £12,000 of recoverable error, it requires a stronger regulatory or safety justification. Conversely, a low-frequency medical or security decision may warrant two-person control because the consequence of one mistake is severe. Triage should distinguish controls that manage material risk from habits that merely signal caution.

Step four: test data readiness and policy coherence

AI workflows fail quietly when inputs are incomplete, inconsistent or inaccessible. For every decision, list the required data, its source, format, owner, freshness and permissible use. Then calculate field completion and agreement rates. If customer identifiers match across billing and support systems in only 81% of records, an AI service agent cannot reliably assemble a complete history. Entity resolution may help, but the organisation must first decide which system is authoritative and how disputed matches are handled.

Missing data should not be disguised by model inference. Predicting a product code from free text may be appropriate for routing, provided confidence thresholds and correction paths exist. Inferring consent, legal status, safety certification or contractual entitlement is different: those facts require evidence. The workflow should pause, request the missing information or escalate. A fluent answer is not a valid record.

Policies require the same scrutiny. Ask three experienced staff to resolve ten representative cases using current guidance. If their decisions diverge materially, the problem is not model performance but policy coherence. Conflicts often appear between sales promises and fulfilment rules, global standards and local regulation, or fraud controls and customer-service targets. Resolve those contradictions, document precedence and version the rules before using them in prompts, retrieval systems or automated decisions.

Step five: separate rules, judgement and generation

Once the process is stable enough to automate, decompose each step by the kind of reasoning required. Deterministic rules suit calculations, thresholds, eligibility checks and mandatory fields. Conventional workflow software suits routing, timers and system updates. AI is strongest where inputs are unstructured or language-intensive: classifying emails, extracting clauses, comparing evidence, drafting responses and summarising case histories. Human judgement remains necessary for ambiguous, contested or high-consequence decisions.

This separation prevents a common design error: asking one model to perform an entire process in a single prompt. Consider insurance claims. AI may extract damage details from documents, compare them with policy wording and draft a recommendation. Rules can verify coverage dates and excess amounts. A claims handler should decide cases involving suspected fraud, conflicting evidence or substantial loss. Each component has a defined responsibility, making testing and accountability possible.

Set confidence and consequence thresholds independently. A 95% confidence score may be adequate for sorting internal mail but unacceptable for denying a benefit. Low-confidence cases should route to review, while high-confidence outputs still need escalation when impact is severe. Also design for abstention: the system must be able to say that evidence is missing or policies conflict. The cost is that some promised automation will remain manual; the benefit is a workflow that fails visibly rather than producing authoritative-looking errors.

Step six: pilot against operational outcomes

Pilot the redesigned workflow with a bounded population, such as one product line, region or case type, and establish a baseline before deployment. Track at least five measures: end-to-end cycle time, touch time, first-pass accuracy, escalation rate and rework. Add risk-specific measures such as incorrect payments, missed regulatory deadlines or customer complaints. Model accuracy alone is insufficient. A classifier can score 94% overall while repeatedly mishandling the rare cases that create most liability.

Use a shadow phase before granting action rights. The AI should produce classifications, summaries or recommendations alongside the existing process without sending messages, approving payments or changing records. Compare outputs with final decisions, then investigate disagreement rather than assuming the human or model is correct. After performance stabilises, release autonomy in stages: drafting first, then low-risk execution, then broader action if evidence supports it.

Account for hidden operational costs. Human review may take longer when staff must verify dense AI-generated text than when they create a short note themselves. Integration, monitoring, incident response and model updates can erase apparent savings. A pilot that saves 400 staff hours but creates 250 hours of review and 100 hours of support has delivered only 50 net hours. The business case should include these costs, along with failure recovery and vendor dependence.

Step seven: govern the workflow, not just the model

AI governance should follow the complete chain from input to outcome. Assign an operational owner for performance, a policy owner for decision rules and a technical owner for reliability. Define who can suspend automation, how affected users appeal and what evidence is retained. Logs should capture source material, relevant policy version, model or system version, output, human intervention and final action. Without that lineage, disputes become guesswork.

Review triggers should be based on change as well as time. A new regulation, product, fraud pattern, data source or organisational structure can invalidate the workflow even if model metrics appear stable. Monitor input drift, override rates and exception volumes. If human overrides rise from 8% to 19% after a pricing change, investigate immediately; do not wait for the quarterly model review.

The triage decision should ultimately place each candidate into one of four categories: eliminate, redesign, automate or retain for human handling. Eliminate steps with no current purpose. Redesign broken handoffs, controls and data capture. Automate stable, measurable work with appropriate tools. Retain human authority where ambiguity or consequence demands it. This sequence may delay an AI launch by several weeks, but it avoids spending months industrialising confusion—and produces automation that improves the operation rather than merely speeding up its defects.