Govern the Agents

Introduction

A new class of technology is moving quietly onto the retail floor and into the warehouse. Agentic artificial intelligence is software that does not merely report a problem. It proposes a resolution and, within defined limits, acts on it.

We see this capability starting to take on the operational work that determines whether a retailer makes or loses money. That includes keeping stock in the right place, marking down ageing product in time, and catching fraud and error. It also means chasing the supplier and fulfilment failures that erode margin one small decision at a time.

The value and the risk, in numbers

• $240 to 390 billion. McKinsey’s estimate of the annual value generative AI could unlock for retail, worth roughly 1.2 to 1.9 percentage points of industry profit margin.

• 2 to 5% revenue and 2 to 3% gross margin improvement where merchants are equipped to make better, faster decisions, and 4 to 10% EBITDA uplift from end-to-end AI transformation.

• More than 40% of agentic AI projects will be cancelled by 2027, Gartner forecasts, on escalating cost, unclear value or inadequate risk controls. The differentiator is governance, not ambition.

• Retail and consumer goods rank second across all industries for agentic AI adoption, at roughly 47%. The technology is arriving whether a given board is ready for it.

For management, this is an operational story. For a board, it is something else entirely. We think the questions that matter are the ones every director is trained to ask. How is it governed? What risk does it introduce, and what risk does it remove? What is the board’s role in overseeing it? And where does it change the liabilities carried by the organisation and its directors? In this article, we take those four lenses in turn.

Governance: who decides, and how it is controlled

The first duty of a board approaching any powerful capability is to satisfy itself that decision rights and controls are clear. Agentic AI makes this concrete, because the technology can, in principle, act. We believe the governing question is not how clever the model is, but what it is permitted to do on its own, and what it must bring to a human.

A well governed system answers that question by design, not by policy written after the fact. Agents surface and recommend, and act only within explicit, preset limits. Routine, reversible steps proceed automatically. Anything that moves money or commits stock beyond a defined threshold is held for human approval. Every action, whether automated or approved, is logged and attributable, and we can pause or switch off the whole system at any time. We deliberately keep the design philosophy conservative: brakes matter as much as acceleration.

Governance also requires visibility, and here agentic AI offers boards something retail has rarely had: a single live view of operational health. Every agent’s findings roll into one place, showing what needs attention, how urgent it is, and what has already been handled. We group related issues together and drop self-resolving matters, so the view stays an honest read on the state of operations rather than a growing to do list. For a board, that is the difference between assurance built on last month’s report and assurance built on what is happening now.

1. Risk: the cost of acting badly, and of not acting

Directors are trained to weigh risk on both sides of the ledger, and agentic AI demands the same discipline.

The risk of adopting badly is real. Gartner warns that more than 40% of agentic AI projects will be cancelled by 2027, over escalating cost, unclear value or weak risk controls.

Every board should take that caution seriously. Autonomous software introduces its own risks. It can act on a wrong inference, touch data it should not, or drift beyond its intended scope. We see these as governable risks, but only if the controls are designed in from the start. That means scope limits, human approval on anything material, comprehensive logging, and the ability to withdraw access instantly.

The risk of not acting is just as real, and easier to overlook. In retail, a slow decision is usually a wrong decision. A markdown taken three weeks late clears stock at a fraction of its value. A replenishment missed on Monday is a lost sale by Friday. These losses rarely appear on a single line of the P&L, which is exactly why they persist. A board that treats doing nothing as the safe option may be accepting a larger, quieter risk than the one it is avoiding.

We build our mitigations to be concrete, and they belong in any risk committee’s assessment. We design and operate our platform to be SOC 2 ready. Its security, access and audit controls meet the standard expected of an enterprise system that handles operational and customer data. We build on frontier models from established providers, using Google and Anthropic, with OpenAI’s vision models for visual understanding, so the underlying intelligence rests on their substantial safety and security investment. We connect to existing systems in a controlled way, using an API first approach built on open standards. Agents operate on scoped, revocable security keys, and user access ties into the organisation’s existing single sign on and identity controls. In plain terms, the business can see exactly what each agent is permitted to touch and revoke it in an instant.

2. The board’s role: oversight, not abdication

The arrival of software that can act on the organisation’s behalf raises an obvious governance temptation: to treat it as a purely operational matter delegated to management. We think that would be a mistake. A board can delegate the work; it cannot delegate the accountability. The directors’ duty of care and diligence extends to satisfying themselves that autonomous systems operating inside the business are properly bounded, monitored and reversible.

That does not require directors to become experts in artificial intelligence. It requires them to ask a short, disciplined set of questions, and to be satisfied with the answers.

• Limits. What is the system permitted to do on its own, and what must it escalate to a human?

• Approval. Who signs off anything that moves money or commits stock, and is that person accountable?

• Audit. Is every action logged and reconstructable after the fact?

• Off switch. Can we pause or stop it immediately, and who holds that authority?

• People. Is this augmenting our workforce, or quietly replacing judgement that should stay human?

We believe a system whose vendor cannot answer those questions clearly does not belong in a regulated, customer facing business, however impressive the demonstration. Handled well, agentic AI is an augmentation story. It absorbs repetitive, low judgement work and returns skilled attention to customers, ranging and growth. We think boards should hold management to that framing, and to the evidence behind it.

3. Liabilities: on the balance sheet and in the boardroom

Finally, we encourage directors to look at agentic AI through the lens of liability, in both senses of the word.

On the balance sheet, it reduces liabilities the business already carries. Ageing and excess inventory is a liability dressed as an asset, and every dead stock write down and mistimed markdown is value the organisation loses today. Agents that position stock where it will sell, clear ageing product at the right moment, and catch shrinkage early attack that liability directly. This frees working capital and protects margin. This is where the sector level numbers cited above translate into a specific, defensible ambition: not a headline figure, but a measurable reduction against the retailer’s own baseline.

In the boardroom, it must not create new liabilities. Autonomous systems touch customer data, pricing and money movement, all areas governed by privacy law, consumer law, and the board’s own accountability. The controls that make the technology safe are the same ones that manage this exposure. A complete, attributable audit trail is not only a governance tool but evidence of diligence. Scoped access and single sign on reduce the data security surface. SOC 2 ready controls meet the bar auditors and enterprise customers expect. An API first, open standards approach avoids the vendor lock in that erodes a business’s position over time. Adopted this way, we believe the technology strengthens the organisation’s control environment rather than weakening it.

4. Compliance and the law

Retail is among the most heavily regulated corners of the economy, and much of that regulation bites precisely where agentic AI operates. Pricing and promotions must be honest and not misleading under consumer law. Customer data is governed by privacy obligations. Certain product information carries a legal mandate, food allergen declarations being the clearest example. Workforce rostering sits within award and workplace law constraints. Increasingly, the governance of AI and data is itself becoming a compliance question.

Here we see the technology as a genuine ally, in two ways. First, it helps the business stay compliant. Agents can enforce mandatory rules, refusing for instance to treat a food product as complete without its legally required allergen information. They can keep pricing and promotional actions consistent and defensible, roster within legal constraints, and leave a complete, timestamped record of what was decided and why. Second, we build the platform itself to be governed, with human approval on anything touching money or stock, so it strengthens the control environment rather than adding regulatory risk.

One caveat we hold firmly: technology can support compliance, but it does not discharge the obligation. Accountability remains with the board and management. Our ambition is for the platform to make compliant operation the path of least resistance, and non-compliance visible, not to substitute for legal judgement.

Closing

Agentic AI in retail is no longer a question of if, but of how well. We believe the organisations that benefit most will not be those that automate most aggressively. They will be the ones that automate most deliberately, pairing the reach of machines with the judgement of their people, and keeping the board’s oversight firmly in the loop.

That is, in the end, a familiar assignment. Govern it clearly. Weigh the risk on both sides. Hold the accountability that cannot be delegated. Understand where it changes the organisation’s liabilities. We believe directors who bring those four disciplines to agentic AI will capture its value without surrendering control, which is exactly the stewardship they are appointed to provide.