Learning objectives. Convert a business process into an agent specification; apply the process-to-agent method; identify decision points, controls and KPIs; and construct a defensible ROI case.

Core lesson

This module is the commercial heart of the programme. Everything before it was capability; this is application.

The process-to-agent method

Business Problem › Process › Decision Points › Agent Responsibilities › Tools › Workflow › Controls › Evaluation › ROI

1. Name the business problem as a cost. Not “our reporting is manual” but “producing forty monthly client reports consumes sixty analyst hours, delivery slips past the fifth working day about a third of the time, and the inconsistency generates client queries.”

2. Map the process as it actually is. Not the SOP — the reality, including the exceptions, the tribal knowledge and the rework. Standardise before you automate. An unstandardised process automated is a defect factory.

3. Identify decision points. At each step, ask: is this a rule (deterministic), a judgement (agent-suitable), or a responsibility (human-only)? Responsibilities include final client recommendations, pricing, legal positions, financial commitments, public statements and sensitive communications. These do not move.

4. Assign responsibilities. Write, explicitly, what the agent does, what a human does, and where they meet. Ambiguity here is where accountability gets lost.

5. Specify tools. With reversibility classification (Module 6).

6. Design the workflow. Deterministic where possible; agentic only where the Agency Test justifies it. Most real business “agents” are workflows with two or three agentic steps inside them, and that is a mature outcome rather than a compromise.

7. Define controls. Approval gates, spend limits, scope restrictions, audit logging.

8. Define evaluation. Test set, metrics, thresholds, monitoring (Module 12).

9. Build the ROI case. Below.

PRACTITIONER — the business agent portfolio

Agent	Realistic scope today	Human retains
Sales agent	Lead enrichment, qualification, meeting prep, CRM hygiene, follow-up drafting	Pricing, negotiation, commitments, relationship
Marketing agent	Research, briefs, first drafts, scheduling support, performance summaries	Positioning, brand judgement, publication approval
SEO agent	Crawl analysis, issue triage, keyword clustering, internal-link suggestions, schema drafting, report assembly	Strategy, client recommendations, anything touching a live site
Research agent	Multi-source gathering, synthesis, citation, contradiction flagging	Interpretation, decisions taken on it
Customer service agent	Classification, retrieval of policy, drafted responses, order status, deflection of routine queries	Complaints, exceptions, goodwill, anything emotionally loaded
Finance intelligence agent	Reconciliation proposals, anomaly detection, variance narratives, invoice preparation	Payment authorisation, statutory reporting, judgement calls
HR agent	Job-post drafting, scheduling, policy answers, onboarding coordination	Hiring decisions, performance, grievances, anything about a named individual
Operations agent	Task creation, status chasing, SOP guidance, meeting notes and actions	Prioritisation, escalation decisions
Compliance agent	Monitoring, checklist verification, evidence collection, drafting registers	Legal interpretation, sign-off, regulatory submissions

Two patterns run through the whole table. First, agents are strongest at preparation, synthesis, drafting, checking and coordination — the work around decisions. Second, the human retains the decision, the commitment and the accountability. A firm that respects that division captures most of the value with a fraction of the risk.

The ROI case

Benefit = (hours saved × loaded hourly cost) + (value of speed) + (value of consistency and reduced rework) + (capacity released for revenue-generating work).

Cost = model and tool spend + platform costs + build effort amortised + maintenance (budget twenty to thirty per cent of build effort annually) + human review time, which is the line most business cases omit and the one most likely to erase the gain.

Risk-adjusted = benefit × expected success rate − cost of failures × their frequency.

Be honest about the review line. An agent that saves ninety minutes and requires forty minutes of careful checking has saved fifty minutes, not ninety — and if the checking is skipped because it is tedious, you have not saved time, you have transferred risk.

Business example — the monthly reporting agent

Problem: sixty analyst hours a month, late delivery a third of the time. Process: pull data from analytics, search and rank sources; identify significant changes; write commentary; assemble the deck; review; deliver. Decision points: data pull is a rule; identifying significant change is judgement; commentary is judgement; the client recommendation is a human responsibility. Design: deterministic data collection; an agent that detects and explains significant movements with citations back to the data; drafted commentary in house style; automatic assembly; a strategist who writes the recommendations and approves. Result: preparation time fell from roughly ninety minutes to twenty-five per report, delivery moved to the second working day, and the strategist’s time shifted from assembly to recommendations — which is the part the client actually pays for. KPIs: hours per report, on-time rate, rework rate, client query volume, cost per report.

Common mistakes

Automating a broken process rather than fixing it first.
Building the agent that is interesting rather than the one that is expensive to do manually.
Omitting review time from the business case.
No baseline. If you did not measure the manual process, you cannot prove improvement.
Letting an agent hold a responsibility rather than a task.
Declaring victory at the pilot. Adoption, not capability, is where most projects actually die.

Expert insight

The highest-value business agents are usually unglamorous: preparation, triage, reconciliation, chasing, formatting, first drafts. They are valuable because they are high-volume, low-judgement and currently done by expensive people at the wrong end of their skill range. Ask not “where could AI be impressive?” but “where are qualified people doing work beneath their qualification?”

Knowledge check

State the nine stages of the process-to-agent method.
Distinguish a rule, a judgement and a responsibility, with an example of each.
Why must a process be standardised before it is automated?
Which cost line is most often omitted from agent business cases?
Why are most successful business “agents” actually workflows with agentic steps?
Give three tasks in your function that are high-volume and low-judgement.
Why is a baseline measurement essential before building?

Challenge

Select one process. Produce a complete one-page specification: problem stated as a cost, current process, decision points classified, agent and human responsibilities, tool list with reversibility, controls, KPIs, and a risk-adjusted ROI estimate including review time. This document, not a prototype, is the deliverable that gets projects approved.

Building Business Agents