AI Strategy and Transformation: AI Agents
The four agents that assemble evidence for this function, what each may read, and the five decisions no agent is allowed to make.
Markdown. No sign-up, no email.
A transformation function that automates its own judgement has removed the only thing it was there for. The split below is deliberate and holds throughout: agents assemble the evidence, people make the call.
The four agents#
Process-mining agent#
Reads event logs from the systems a process touches and reconstructs what actually happens: the real sequence, the real durations, where cases wait, and how often the path taken differs from the path in the written procedure.
It is routinely the first thing to contradict the organisation's account of itself. A process everyone describes as three steps turns out to have eleven, four of which exist to correct the other seven.
Access: read-only, on logs only. It never touches the systems it is reading. Output: a ranked view of where hours and rework actually go, with the evidence attached.
Candidate-scoring agent#
Applies the four-axis model from the SOPs, volume, structure, reversibility, tolerance, to a proposed change, and returns the score with every input shown.
Showing the inputs is the point. A score that arrives as a single number cannot be argued with, and an unarguable score is a decision dressed up as an analysis.
Access: the process inventory. Read-only. Output: a score, its inputs, and the axis that constrains it.
Benefit-tracking agent#
Holds the frozen baseline and reports the post-change measure on schedule, at 90 days and again at 12 months.
It reports the number when the number is bad. That is the entire reason it is separated from delivery: the team that built the change is the worst-placed to judge whether it worked, and the most motivated to round up.
Access: the measurement store, read-only. It cannot alter a baseline. A baseline that can be adjusted after the fact is not a baseline. Output: measured against forecast, published either way.
Capability-scanning agent#
Watches what has become buyable since the last build-or-buy decision, and flags where a sound decision has been overtaken by the market.
It has an explicit bias against novelty. Most new capability is not yet worth its switching cost, and the agent's job is to say when that changes, not to relay announcements.
Access: public sources. Read-only. Output: a dated assessment with sources, reviewed by a person before it reaches the portfolio.
What no agent does here#
- Choose which processes change. The ranking encodes the company's risk appetite. Risk appetite is owned, not computed.
- Sign off a benefit claim. Measurement and delivery stay apart, and an agent that certifies its own programme collapses that separation.
- Decide to stop a programme. Kill decisions are political before they are analytical, and the political part is exactly what a model cannot see.
- Anything that changes what a person does all day. Where automation reshapes a job, a person decides, with HR in the room.
- Alter a baseline, a refusal-list entry, or a published result. All three are append-only by design.
The rule underneath all of it#
The test for automating a decision is not whether it can be automated. It is reversibility and accountability: if a mistake is hard to undo, or if someone must answer for the outcome, a person stays in the loop regardless of what the technology can do.
Applied to this function's own work, that rule produces exactly the split above. Assembling evidence is reversible and unattributed. Choosing what the company does with that evidence is neither.
Where this sits against the governance frameworks#
Checked 9 August 2026. Worth knowing that the three frameworks most enterprises run were all written before autonomous agents existed:
| Framework | Role it plays | Covers agents? |
|---|---|---|
| OECD AI Principles | The values statement | No |
| NIST AI RMF | The internal risk operating model | No |
| ISO/IEC 42001 | The certifiable management system procurement asks for | No |
None of the three was designed for agentic systems, so anyone deploying agents has to extend them by hand to cover cascading failures, scope creep and attribution gaps.
Singapore's Model AI Governance Framework for Agentic AI, launched 22 January 2026, is the first to address autonomous agents directly. Its central principle is that an AI agent cannot be a principal: even when the agent acts on its own, the organisation remains responsible, a human remains accountable, and any delegation must be explicit and bounded. It expects each agent to carry a verifiable identity and an audit trail showing which agent acted under whose authorisation.
That is the same conclusion this framework reached independently, and it is why every department page here carries a "what is NOT delegated to an agent" section rather than a list of capabilities. The useful question was never how much an agent can do. It is who answers when it is wrong.
Note also that EU AI Act Article 50 transparency obligations have applied since 2 August 2026: if a system interacts directly with a person, that person must be told they are dealing with AI, and generated content must be marked. Penalties reach €15 million or 3% of worldwide turnover, and a transition period closing on 2 December 2026 covers anything that was already on the market before May 2026.