AI: The Decisions People Are Getting Wrong — An Operator’s Field Guide: Operator Field Guide
The expensive AI mistakes are rarely model mistakes. They are management mistakes: automating unstable work, buying before diagnosing, trusting fluent output, ignoring adoption, and measuring activity instead of operating value.
Theo MarchettiInvestigations editorFirst published 8/6/2026 · last revised 8/8/2026 with fresh sources, corrections, and new context. Reader corrections are reviewed and folded into future versions.
Summary
Most organizations are not failing at AI because they chose the second-best model. They are failing because they framed the decision as a software purchase rather than a redesign of work, controls, and accountability. Leaders routinely automate broken processes, confuse demonstrations with dependable systems, and calculate ROI without counting review, integration, exception handling, security, or adoption. The better operating question is not ‘Where can we add AI?’ but ‘Which decisions and workflows can be improved measurably, under what controls, and who owns the result?’
Key takeaways
Explain like I'm 5
Imagine hiring an extremely fast junior colleague who has read a vast library, writes confidently, and sometimes invents details. You would not hand that colleague the company bank account, customer database, and authority to sign contracts on day one. You would provide a clear task, examples, limited access, review rules, and a manager responsible for the outcome. Business AI should be deployed the same way. A chatbot mainly produces an answer; an agent may also retrieve records, update systems, send messages, or trigger payments. Every added action increases usefulness and risk. The sensible progression is observe, recommend, draft, act with approval, then act within tightly bounded conditions. Autonomy is not the objective; better economics and service, with acceptable risk, are.
Deep dive
Wrong decision 1: buying AI before diagnosing the workflow
Executives often begin with ‘Which platform should we buy?’ The useful unit of analysis is a workflow: trigger, inputs, decisions, handoffs, systems, exceptions, output, and accountable owner. Consider inbound sales. ‘Deploy an SDR agent’ hides several jobs—enrichment, qualification, routing, research, drafting, scheduling, CRM updates, and escalation—whose data quality and risk differ. Map the current process first. Record volume, labor minutes, waiting time, rework, error cost, conversion rate, and exception frequency. If the operating procedure changes weekly or employees disagree about a qualified lead, automation will scale ambiguity rather than remove it.
Wrong decision 2: treating a polished demo as production evidence
A demo is usually run on selected examples, with attentive operators and forgiving latency. Production introduces malformed files, missing permissions, prompt injection, API outages, duplicate records, multilingual customers, policy changes, and users who behave unexpectedly. Evaluate on a representative test set, including adversarial and edge cases. Define acceptance criteria before the pilot: factual accuracy where measurable, task completion, human acceptance, latency, unit cost, escalation rate, and severe-error frequency. Then run in shadow mode, where the system recommends but does not act, before granting write access. A 90% aggregate score can still be unacceptable if the remaining 10% includes discriminatory hiring recommendations or incorrect bank details.
Wrong decision 3: maximizing autonomy instead of system reliability
Agentic AI is valuable when work requires interpretation, tool use, and adaptation. It is wasteful when a deterministic rule, database query, or conventional workflow engine can do the job more cheaply and predictably. Strong systems are usually hybrid: rules enforce hard policy; models classify, extract, summarize, or draft; humans resolve high-impact ambiguity. Use an autonomy ladder—read-only retrieval, recommendation, drafting, action with approval, bounded action, broader delegation. Advancement should depend on observed evidence, not vendor vocabulary. Restrict tools, records, spending, recipients, execution frequency, and reversibility. The best agent is not the one with the most freedom; it is the one that completes valuable work inside an explicit operating envelope.
Wrong decision 4: calculating ROI from wages alone
The familiar calculation—hours saved multiplied by salary—overstates value when saved minutes do not become additional capacity, revenue, lower headcount, or faster service. It also omits integration, model usage, observability, evaluation, security review, training, supervision, retries, and incident response. Use cost per accepted outcome: total operating cost divided by outputs that pass business quality standards. Compare it with the current baseline and an alternative such as process simplification or conventional automation. Benefits should be traced to a financial mechanism: more selling time, shorter onboarding, fewer support escalations, lower rework, faster cash collection, or avoided compliance loss. Count displaced effort only when management has a credible plan for redeploying it.
Wrong decision 5: making AI governance either theatrical or paralyzing
A blanket ban pushes employees toward unsanctioned tools; unrestricted experimentation exposes confidential data and creates invisible dependencies. A practical control plane classifies use cases by data sensitivity, decision consequence, reversibility, customer exposure, and degree of autonomy. Low-risk drafting may receive pre-approved tools and lightweight logging. Employment, credit, health, legal, payment, or safety decisions warrant stronger evaluation, human review, evidence retention, and sometimes prohibition. NIST’s AI Risk Management Framework organizes work around Govern, Map, Measure, and Manage; ISO/IEC 42001 provides an AI management-system standard. Neither replaces operational ownership. Someone must be able to pause the workflow, revoke credentials, reconstruct actions, notify affected parties, and remediate damage.
Wrong decision 6: underfunding adoption and accountability
AI changes who gathers information, who judges exceptions, and what managers inspect. That is operating-model change, not merely training. Frontline employees may reject a system that creates more checking than it removes, while leaders may misread high usage as value. Design the human interface around uncertainty: show sources, disclose missing evidence, make corrections easy, and route exceptions to named roles. Track override reasons because they reveal bad instructions, missing data, policy conflict, or model weakness. Give every deployment one business owner, one technical owner, and explicit security and compliance partners. Scale only after the workflow demonstrates sustained quality, user adoption, control effectiveness, and a credible economic result.
- 1956The Dartmouth Summer Research Project, organized by John McCarthy and others, helps establish ‘artificial intelligence’ as a field.
- 1997IBM Deep Blue defeats chess world champion Garry Kasparov, demonstrating narrow machine superiority under defined rules.
- 2012AlexNet’s ImageNet performance accelerates commercial investment in deep learning and GPU-based AI.
- 2016DeepMind’s AlphaGo defeats Lee Sedol, highlighting systems that combine learned models, search, and reinforcement learning.
- 2017Google researchers publish ‘Attention Is All You Need,’ introducing the Transformer architecture behind modern language models.
- 2020OpenAI releases GPT-3 through an API, making general-purpose language generation accessible to software teams.
- 2022ChatGPT launches publicly on November 30, making conversational generative AI a mass-market product.
- 2023NIST publishes AI RMF 1.0; Microsoft releases Copilot broadly, while enterprises accelerate governed pilots.
- 2024The European Union’s AI Act enters into force on August 1, beginning a phased, risk-based regulatory rollout.
- 2025Organizations increasingly shift from standalone copilots toward agents that use tools, retrieve enterprise data, and execute bounded workflows.
Glossary
- AI agent
- A software system that uses an AI model to interpret a goal, select steps, use permitted tools, and act within an environment, often with memory or state.
- Copilot
- An assistive interface that drafts, summarizes, searches, or recommends while leaving execution and accountability primarily with a person.
- Agentic workflow
- A process in which models perform multi-step reasoning or tool use, usually constrained by rules, approvals, and deterministic software.
- Grounding
- Connecting model output to supplied evidence, such as approved documents, database records, or retrieved passages, to improve relevance and verifiability.
- Hallucination
- A fluent but unsupported or incorrect model output. The business concern is not the label but the probability and consequence of an undetected error.
- Evaluation
- A repeatable test of system performance using representative tasks, scoring criteria, human judgments, automated checks, and often adversarial cases.
- Human in the loop
- A control pattern in which a person reviews, approves, corrects, or handles exceptions before or after an AI-supported action.
- Least privilege
- The security principle of granting an agent only the data access and actions necessary for its assigned task, for no longer than needed.
- Cost per accepted outcome
- Total workflow cost—including models, infrastructure, review, failures, and operations—divided by outputs that meet defined business standards.
- Shadow mode
- A deployment stage in which the AI produces recommendations alongside the live process but cannot execute consequential actions.
FAQs
Should an organization choose a model or a workflow first?+
Choose the workflow first. Its data, latency, accuracy, privacy, tool-use, and cost requirements should determine whether you need a frontier model, a smaller model, conventional automation, or no AI at all.
What is the safest first use case for an AI agent?+
Start with a frequent, measurable, reversible workflow using non-sensitive or well-controlled data. Internal research, case summarization, meeting preparation, and draft generation are usually safer than payments, hiring decisions, or customer commitments.
How should executives measure AI ROI?+
Compare the controlled deployment with a credible baseline using cost per accepted outcome, cycle time, quality, conversion, rework, and risk. Translate gains into a financial mechanism rather than counting all theoretical hours saved as cash.
When is human approval necessary?+
Require approval when actions are consequential, hard to reverse, legally sensitive, externally binding, or based on uncertain evidence. Approval can be sampled or exception-based only after measured performance and control effectiveness justify the change.
Does retrieval-augmented generation eliminate hallucinations?+
No. Retrieval can provide relevant evidence, but the model may retrieve the wrong material, misinterpret it, omit qualifications, or make claims beyond the sources. Citations and deterministic validation remain important.
Is using an enterprise AI plan enough to protect confidential data?+
No. Buyers must verify retention, training use, subprocessors, access controls, encryption, regional processing, deletion, logging, and contractual terms. Internal permissions and employee behavior can still expose data even when the vendor’s platform is secure.
Should companies build or buy AI agents?+
Buy commodity capabilities when vendor controls and integration fit; build differentiated workflow logic when it creates strategic value. Many strong deployments combine purchased models and platforms with proprietary data, evaluations, rules, and orchestration.
How long should a pilot run?+
Long enough to include normal volume, edge cases, operational handoffs, and user behavior—often several weeks rather than a single workshop. Set a decision date and success thresholds beforehand to prevent permanent pilot limbo.
Predictions
{"items":["Through 2027, more enterprises are likely to purchase agent capabilities inside existing CRM, service, productivity, and ERP suites rather than deploy separate general-purpose agents.","Model choice may become less strategically important than proprietary evaluations, permission architecture, process data, and integration quality as capable models commoditize.","Regulated and high-consequence workflows will probably adopt tiered autonomy, with routine cases handled automatically and uncertain cases routed to specialists.","AI observability may develop into a standard operating layer covering prompts, tool calls, evidence, cost, latency, overrides, and incident reconstruction.","Boards are likely to demand portfolio-level AI reporting that separates experimentation, controlled production, realized economics, and material risk exposure."}]}
Risks
{"items":["Silent decision errors: plausible output can pass quickly through sales, finance, legal, or operations unless evidence and validation are designed into the workflow.","Permission amplification: an agent with broad credentials can expose data or perform harmful actions faster than a person, particularly under prompt injection or account compromise.","Automation debt: brittle integrations, unowned prompts, undocumented exceptions, and model changes can accumulate into an expensive shadow operating system.","Regulatory and contractual exposure: personal data, employment, credit, intellectual property, records retention, and sector rules may apply even when a project is called a pilot.","Human deskilling and automation bias: employees may lose process knowledge or defer to confident recommendations, reducing their ability to identify unusual failures."}]}
Opportunities
{"items":["Compress revenue-cycle friction by researching accounts, preparing outreach, updating CRM records, and escalating buying signals while preserving human control of claims and commitments.","Turn unstructured operational material—emails, calls, PDFs, tickets, and notes—into classified work queues with evidence, confidence indicators, and exception routing.","Improve management leverage by generating daily exception briefs that identify blocked orders, at-risk renewals, overdue approvals, and inconsistent forecasts.","Embed institutional knowledge in governed retrieval systems so frontline teams can find current policies, product details, and prior resolutions without relying on tribal memory.","Use process telemetry from AI deployments to reveal unnecessary handoffs, duplicate entry, weak policies, and data defects that should be fixed before further automation."}]}
For professionals
For an enterprise architecture review, model the agent as a non-human service principal rather than an intelligent employee. Specify identity, authentication, authorization, network boundaries, secrets, permitted tools, record-level access, state persistence, retention, rate limits, transaction limits, and kill mechanisms. Separate probabilistic interpretation from deterministic enforcement: the model may propose an action, but policy code should verify schema, entitlement, amount, destination, duplication, and approval status. Preserve traceability across user request, retrieved evidence, model and version, prompt or policy version, tool calls, outputs, approvals, and resulting system changes. Red-team indirect prompt injection through documents, websites, emails, CRM fields, and attachments—not only direct chat inputs. Govern the workflow as a production service with a service-level objective and an error budget. Evaluation sets should be versioned, representative, stratified by consequence, and refreshed when products, policies, attackers, or user behavior change. Monitor severe-error rate separately from averages, because a rare unauthorized refund is not equivalent to a slightly weak summary. Commercial analysis should include vendor concentration, model portability, data egress, contractual indemnities, intellectual-property terms, and exit costs. The decisive capability is not prompt engineering; it is the institutional ability to diagnose work, define acceptance, constrain authority, observe behavior, and continuously improve a sociotechnical system.
Sources & references
- NIST Artificial Intelligence Risk Management Framework (AI RMF 1.0)
- NIST AI 600-1: Artificial Intelligence Risk Management Framework—Generative AI Profile
- ISO/IEC 42001:2023 — Artificial intelligence management systems
- Regulation (EU) 2024/1689 — Artificial Intelligence Act
- OECD AI Principles
- Attention Is All You Need
- Stanford AI Index Report 2024
- McKinsey Global Survey: The State of AI in Early 2024
| Copilot | Approval-gated agent | Bounded autonomous agent | |
|---|---|---|---|
| Primary role | Drafts or recommends | Prepares and proposes an action | Executes predefined classes of action |
| Human control | Human performs the action | Human approves before execution | Human reviews exceptions and monitoring |
| Best fit | Novel, judgment-heavy work | Repeatable work with material consequences | High-volume, well-understood, reversible work |
| Typical example | Drafting an account plan | Preparing a refund for approval | Routing standard support cases |
| Failure containment | High; no direct execution | Medium-high; approval can block errors | Depends on limits, monitoring, and rollback |
| Evidence required to scale | User acceptance and quality | Approval accuracy and lower cycle time | Low severe-error rate and reliable exception handling |
Agent Oracle examines The AI Chief of Staff Playbook through AI agents, workflow automation, sales intelligence, executive decisions, compliance, and measurable business ROI, with practical signals, risks, examples, and a reason for readers to return as the story changes.
Agent Oracle examines AI Agent ROI Scorecards for Small Teams through AI agents, workflow automation, sales intelligence, executive decisions, compliance, and measurable business ROI, with practical signals, risks, examples, and a reason for readers to return as the story changes.
Agent Oracle examines Workflow Bottleneck Mapping With Voice Agents through AI agents, workflow automation, sales intelligence, executive decisions, compliance, and measurable business ROI, with practical signals, risks, examples, and a reason for readers to return as the story changes.
The August 2026 scorecard favors companies turning capable models into dependable systems—and punishes vendors selling intelligence without control, distribution, or measurable workflow economics.
A field guide to separating AI capability from AI theater—and turning agents, automation, and human judgment into measurable operating leverage.
A boardroom-clear briefing on the releases, policy signals, infrastructure economics, and implementation lessons that matter now—and the operating system for separating durable shifts from weekly AI noise.