Before you buy an AI agent, fix the workflow
An agent cannot rescue a process that nobody understands. Start with the hand-offs, decisions and exceptions that define the work.
Workflow redesignAgentic technology earns authority only after the workflow has clear states, tested controls and an accountable route for exceptions.
01Use conventional automation for deterministic work.
02Separate drafting, recommendation and action authority.
03Expand autonomy only after measured reliability.
Start with the work
Agent demonstrations compress a complex process into a smooth sequence. Real operations contain incomplete inputs, policy conflicts and exceptions that require negotiation. Automating the demonstration instead of the process creates fragile systems.
Follow a real item from arrival to completion. Record states, decisions, evidence, wait time and ownership. This map reveals whether the constraint is interpretation, missing data, unclear accountability or an ordinary integration problem.
Use the least complex reliable tool
Fixed calculations, validated field movement and known notifications belong in deterministic software. Use AI where language, variation or interpretation creates a genuine advantage.
A combined design is usually stronger than a single general agent. Explicit rules provide stability, while AI handles bounded ambiguity.
Match authority to consequence
Drafting an internal summary can tolerate review. Releasing payment, changing a customer record or sending regulated advice requires explicit approval and a durable audit trail.
Define three levels: draft, recommend and act. Each level needs evidence thresholds, permissions and a route to a person.
Earn autonomy through evidence
Build an evaluation set from normal work, edge cases and known failures. Track task completion, corrections, time saved and exceptions. Review failures by cause rather than reporting one blended accuracy score.
The best first system is rarely the most autonomous. It is the one the team can understand, operate and improve.
What the wider evidence says
Findings are paraphrased from the linked original publications. Their scope and populations differ, so they inform the thesis rather than prove a universal outcome.
Experienced open-source developers took longer with early-2025 AI tools in one randomised study
Capability claims must be tested in the real context, especially where experts hold substantial tacit knowledge.
Source: METR, Experienced Developer Productivity Study, 2025Field research finds that AI performance varies materially across task boundaries
Authority should be assigned by task and consequence, not by a general belief that the model is capable.
Source: Harvard Business School, Navigating the Jagged Technological FrontierAgentic AI remains the least adopted AI category in current UK research
Leaders should treat agent deployment as controlled operating-model change, not a mature default.
Source: UK Government, AI Adoption Research 2026Google DORA recommends shorter delivery batches as AI increases the velocity of change
A bounded release with fast feedback and explicit rollback is a stronger path to autonomy than a large agent rollout.
Source: Google DORA, AI Capabilities Model, 2025A controlled route from thesis to operating evidence
Each work package ends with an explicit decision and a tangible output. The sequence keeps delivery connected to operating evidence.
Observe real work
- Decision
- What actually happens between request and completion?
- Work
- Trace normal cases, incomplete inputs, policy conflicts, rework and informal escalation.
- Output
- A state, decision and exception map grounded in evidence.
Simplify before automating
- Decision
- Which steps can be removed, standardised or validated?
- Work
- Clarify ownership, reduce duplicate fields and establish a controlled system of record.
- Output
- A cleaner workflow with deterministic foundations.
Bound AI authority
- Decision
- Should the system draft, recommend or act?
- Work
- Classify consequences, define evidence thresholds and route approvals by risk.
- Output
- An authority matrix tied to tools and permissions.
Earn greater autonomy
- Decision
- Does live evidence justify more independent action?
- Work
- Review evaluation results, overrides, incidents and exception recovery at release gates.
- Output
- An explicit expand, hold, restrict or stop decision.
The failure modes leadership should watch before scale
Risks become manageable when the early signal, control and accountable owner are agreed before release.
A scorecard that connects activity to management action
Measures are useful only when their definition is stable and their movement changes a management decision.
Authority should rise more slowly than model capability
Illustrative control intensity by action consequence.
An authority gate separates suggestion from action
01proposal = agent.plan(task, context)02risk = controls.classify(proposal)03evidence = evaluator.check(proposal)04approval = authority.route(risk, evidence)05executor.run(approval.allowedActions)