The Open Questions That Will Define AI Next: An Operator’s Field Guide: Operator Field Guide
The decisive AI questions are shifting from model intelligence to agent reliability, workflow economics, control, liability, and organizational design. Here is what business leaders should watch—and test—before placing the next large bet.
Saoirse MulliganBooks & ideasFirst published 8/16/2026 · monitored for updates; the next revision publishes a new version and appears here. Reader corrections are reviewed and folded into future versions.
Summary
AI’s next chapter will not be determined by benchmark scores alone. The consequential questions concern whether agents can execute long workflows reliably, whether businesses can verify their decisions, who bears liability when they fail, and whether automation creates durable operating leverage rather than expensive demonstrations. For executives, the practical task is to separate capabilities that improve quickly from constraints—data quality, process ambiguity, security, and accountability—that remain stubbornly organizational. The winners are likely to treat AI as a governed operating system for work, not a collection of isolated chat tools.
Key takeaways
- Agent reliability—not conversational fluency—is becoming the gating factor for production adoption.
- Workflow diagnosis should precede vendor selection: automate stable decisions before ambiguous exceptions.
- The relevant ROI unit is the complete process, including review, integration, failures, and change management.
- Enterprises will need explicit authority boundaries defining what an agent may read, recommend, write, spend, send, or approve.
- Model choice will become more dynamic as firms route tasks among frontier, smaller, open-weight, and specialist systems.
- Security is moving from prompt filtering toward identity, permissions, provenance, sandboxing, and continuous evaluation.
- Regulation will matter, but contracts, insurance, procurement controls, and litigation may shape behavior faster.
- Human advantage will increasingly reside in judgment, relationship ownership, process design, and responsibility for outcomes.
Explain like I'm 5
Imagine hiring a brilliant new assistant who has read an enormous library but sometimes guesses, forgets instructions, or trusts a fraudulent email. That assistant can draft quickly, yet you would not immediately give it permission to sign contracts, transfer money, or email every customer. AI agents create the same management problem at software speed. The central question is no longer simply, ‘Can AI produce a good answer?’ It is, ‘Can an AI system complete useful work repeatedly, prove what it did, stop when uncertain, and call a person at the right moment?’ Businesses that answer those questions carefully can automate meaningful workflows. Those that skip them may merely automate mistakes, security incidents, and customer confusion.
Deep dive
1. Will agents become dependable operators?
Models can already draft proposals, classify tickets, search documents, and write code. Production work, however, requires persistence across many steps: maintaining state, using tools correctly, interpreting exceptions, and verifying completion. A sales agent that identifies accounts but contaminates the CRM, invents contact details, or sends an off-brand message creates negative leverage. Reliability is also multiplicative: if each of ten dependent steps succeeds 95% of the time, the probability that all ten succeed is only about 60%. Better models help, but production systems also need constrained tools, deterministic checks, retry policies, observability, and escalation paths. Operators should test complete workflows using real edge cases, not judge an agent from its best conversation.
2. Where does the automation ROI actually accrue?
The attractive demo often occupies only a fraction of a process. Drafting a quote may take seconds, while validating pricing, securing approval, updating the CRM, delivering the document, and handling revisions consume the real cycle time. A credible business case therefore measures baseline labor, waiting time, error costs, conversion effects, model and integration spend, review effort, and exception rates. AI can be valuable without eliminating jobs: reducing lead-response time, expanding account coverage, or preventing revenue leakage may matter more than hours saved. The unresolved economic question is how quickly inference prices fall relative to demand for more capable reasoning—and how much of the gain vendors capture through subscriptions, usage fees, and platform dependence.
3. How much autonomy should an enterprise permit?
Autonomy is not binary. An agent can observe, recommend, draft, execute reversible actions, or make consequential commitments. Each level needs different controls. A procurement copilot may safely compare approved suppliers but should not alter bank details; a support agent may issue a small credit within policy but escalate threats, regulated claims, or unusual refunds. Mature deployments will use an authority matrix tied to user identity, data sensitivity, transaction value, reversibility, and confidence. The open governance question is whether organizations can make these boundaries legible enough for employees, auditors, customers, and increasingly autonomous software acting on behalf of other software.
4. Can security keep pace with connected agents?
A chatbot that only drafts text has a limited blast radius. An agent connected to email, cloud storage, browsers, payment systems, and customer records becomes a privileged digital worker. Prompt injection can arrive through a webpage, attachment, calendar invitation, or retrieved document and attempt to redirect the agent or exfiltrate secrets. Traditional application controls remain essential: least privilege, short-lived credentials, network segmentation, data-loss prevention, allow-listed actions, approval gates, and immutable logs. Yet agents add ambiguity because natural-language instructions mix trusted policy with untrusted content. The unresolved technical challenge is robustly separating commands from data while preserving enough flexibility to perform useful work.
5. Who is accountable when AI causes harm?
Responsibility is distributed among model developers, application vendors, deployers, data providers, employees, and customers. The European Union’s AI Act creates risk-based obligations, while sector rules, privacy law, employment law, product-liability doctrine, and contractual warranties also apply. But compliance does not settle accountability. If an agent rejects a qualified applicant, exposes confidential information, or promises an unauthorized discount, boards will ask who approved the workflow, what testing occurred, and whether a human could intervene. Procurement terms—indemnities, audit rights, data-use restrictions, incident notification, and service levels—will be as important as regulatory labels.
6. What becomes of work, competition, and trust?
AI may compress the cost of analysis, software creation, personalization, and administrative coordination. That could empower small firms, but scale advantages in compute, distribution, proprietary data, and customer access may concentrate value. Within companies, entry-level tasks could shrink before training systems adapt, weakening the apprenticeship through which people acquire judgment. Externally, abundant synthetic content raises the cost of deciding what is authentic. Brands may respond with verified channels, provenance standards, named human accountability, and premium service. The defining management question is not whether humans stay ‘in the loop,’ but which human capabilities must remain strong enough to challenge the system and own the outcome.
- 2012AlexNet’s ImageNet victory demonstrates the commercial potential of deep neural networks trained with GPUs.
- 2017Google researchers publish ‘Attention Is All You Need,’ introducing the Transformer architecture.
- 2020OpenAI releases GPT-3, making few-shot language capabilities visible through a large general model.
- 2022ChatGPT launches on November 30 and brings generative AI into mainstream workplace use.
- 2023Microsoft expands Copilot across its product portfolio, while tool use and retrieval become core enterprise patterns.
- 2024The EU AI Act enters into force on August 1, beginning a phased implementation of risk-based obligations.
- 2025Agent platforms increasingly emphasize computer use, multi-step research, tool calling, and enterprise orchestration.
- 2026Organizations focus more sharply on measurable agent reliability, identity, auditability, and process-level economics.
Glossary
- AI agent
- A software system that interprets goals, chooses actions, uses tools, and maintains state across multiple steps.
- Agentic workflow
- A process combining model judgment with tools, business rules, memory, verification, and human escalation.
- Hallucination
- A plausible-looking output unsupported by the model’s evidence or the underlying facts.
- Retrieval-augmented generation (RAG)
- Supplying a model with retrieved documents or records at runtime so its response can use current, relevant evidence.
- Prompt injection
- Malicious or accidental instructions embedded in content that attempt to override an agent’s intended policy.
- Human-in-the-loop
- A design in which a person reviews, approves, corrects, or takes over designated decisions.
- Evaluation (eval)
- A repeatable test used to measure quality, safety, cost, latency, or task completion under defined conditions.
- Model routing
- Directing each request to a model selected for capability, risk, speed, privacy, or price.
- Provenance
- Evidence about the origin, modification history, and authenticity of data, content, or an AI output.
FAQs
Are AI agents ready to run core business processes?+
Some bounded agents are ready for production, particularly where inputs are structured, actions are reversible, and success is machine-checkable. High-consequence or exception-heavy processes still require strict permissions, evaluation, monitoring, and human escalation.
What is the best first workflow to automate?+
Choose a frequent, measurable process with stable rules and accessible data, such as lead research, meeting follow-up, ticket triage, or document intake. Avoid starting with a politically sensitive transformation spanning many systems and undefined ownership.
How should a company calculate agent ROI?+
Measure the entire process before and after deployment: labor, waiting time, quality, errors, conversion, review, inference, integration, support, and incident costs. Report both average performance and the cost of exceptions, because rare failures can dominate the economics.
Should we buy an agent platform or build internally?+
Buy when the workflow is common and vendor controls meet requirements; build when proprietary logic, integration, or differentiation justifies ownership. Many enterprises will use a hybrid architecture: commercial models and orchestration paired with internal policy, data, evaluations, and interfaces.
Will smaller or open-weight models replace frontier models?+
Not universally. Smaller and open-weight models can offer lower cost, lower latency, deployment control, and sufficient quality for narrow tasks, while frontier systems remain useful for difficult reasoning and multimodal work. Routing is usually more resilient than committing every task to one model.
What is the largest security mistake with agents?+
Granting broad credentials before defining a narrow authority envelope. An agent should receive the minimum data and tools needed, with separate identities, short-lived access, action limits, logs, and approval gates for consequential operations.
Does human review eliminate AI risk?+
No. Reviewers can become complacent, lack necessary context, or face too many outputs to inspect carefully. Human oversight works only when reviewers have time, authority, evidence, and clear escalation criteria.
What should a board ask management?+
Ask which workflows use AI, what decisions agents can make, how performance is evaluated, which incidents have occurred, and who owns residual risk. The board should also understand vendor concentration, regulatory exposure, data use, and the economics after supervision costs.
Predictions
- By 2028, many enterprises will likely maintain formal agent identity and access policies analogous to those for employees and service accounts.
- Model routing may become standard as buyers allocate routine work to cheaper systems and reserve frontier models for difficult cases.
- Agent evaluation could evolve into a continuous operational discipline, with regression testing tied to every model, prompt, tool, or policy change.
- Procurement may shift from per-seat pricing toward mixtures of consumption, completed outcomes, and risk-adjusted service commitments.
- Verified human and corporate provenance could command a premium as synthetic communication becomes cheaper and more abundant.
Risks
- Silent process corruption: plausible agent outputs can propagate incorrect data across CRM, finance, support, and analytics systems.
- Privilege amplification: compromised instructions may turn connected agents into channels for data theft or unauthorized transactions.
- Automation debt: rapid pilots can create brittle prompt chains, duplicated logic, and undocumented dependencies that are costly to maintain.
- Accountability gaps: unclear ownership among vendors, implementers, business teams, and reviewers can delay response when harm occurs.
- Capability atrophy: removing junior tasks without redesigning training may weaken the future supply of experienced operators and managers.
Opportunities
- Revenue operations can use agents to research accounts, detect stalled opportunities, prepare calls, and enforce CRM hygiene without replacing relationship ownership.
- Operations teams can reduce cycle time by combining document extraction, policy lookup, exception detection, and approval routing.
- Consultancies can productize workflow diagnosis, evaluation libraries, governance design, and post-deployment optimization rather than selling generic AI strategy.
- Security and compliance teams can create reusable permission templates, audit trails, and control evidence for agent deployments.
- Founders can compete with larger firms by automating coordination-heavy back-office work while preserving a high-trust customer interface.
For professionals
For implementation leaders, the right abstraction is the controlled decision loop: sense, interpret, propose, authorize, act, verify, and learn. Map every workflow at that level, then assign a control to each transition. Inputs need provenance and sensitivity labels; interpretation needs grounded context; proposals need policy checks; authorization needs an accountable identity; actions need scoped tools and transaction limits; verification needs independent evidence; learning needs versioned evaluation rather than unreviewed memory. This architecture is more durable than betting on a particular model because models, prices, and capabilities will change faster than the organization’s fiduciary duties. Governance should also be expressed as operating telemetry, not a policy document. Track task completion, unsupported-claim rate, exception frequency, human override, approval latency, unit economics, permission denials, and severity-weighted incidents by workflow and model version. Establish a release gate for changes to prompts, retrieval sources, tools, models, and permissions, with rollback procedures and named owners. For material workflows, maintain a failure-mode register that considers confidentiality, integrity, availability, discrimination, financial loss, customer harm, and regulatory breach. The strategic advantage is not maximum autonomy; it is the ability to increase autonomy safely as evidence accumulates.
Sources & references
- Attention Is All You Need
- NIST AI Risk Management Framework (AI RMF 1.0)
- EU Artificial Intelligence Act
- Stanford AI Index Report 2024
- International AI Safety Report
- OWASP Top 10 for Large Language Model Applications
- Artificial Intelligence, Automation, and Work — Acemoglu and Restrepo
- C2PA Technical Specification
| Copilot | Supervised agent | Autonomous agent | |
|---|---|---|---|
| Primary role | Drafts or recommends for a person | Executes steps with checkpoints | Completes an end-to-end objective |
| Typical authority | Read and suggest | Write within limits; approval for key actions | Act across tools inside a policy envelope |
| Best-fit work | Ambiguous, judgment-heavy tasks | Repeatable workflows with meaningful exceptions | Stable, high-volume, machine-verifiable processes |
| Control burden | Moderate review discipline | Strong identity, logs, checks, and escalation | Highest: sandboxing, transaction limits, continuous monitoring |
| Failure blast radius | Usually local to reviewer | Can affect records or customers | Can propagate across systems rapidly |
| ROI pattern | Individual productivity | Cycle-time and quality improvement | Potential structural cost reduction at sufficient scale |
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 center of gravity in artificial intelligence is moving from models that answer questions to systems that pursue goals, use tools, and complete workflows. The competitive question is no longer who has a chatbot, but who can redesign work around bounded, observable agency.
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.