BML ManAIfesto

We wrote the first BML ManAIfesto nearly a year ago because the conversation around AI had become almost impossible to take seriously. On one side, AI was going to replace everybody, transform everything and make existing business models irrelevant by Tuesday. On the other, it was an unreliable parlour trick with no place near a serious organisation.

Neither position was particularly useful to someone responsible for running a business. Our view was much simpler: AI is the answer; what’s the question?

That question still matters. But AI can now do more than generate an answer or help create a document. It can plan work, use systems, make decisions and take action. The large model providers are moving towards public markets and the commercial pressure that comes with them. Regulators have stopped speaking in principles and started enforcing rules.

So this is the updated ManAIfesto. Ten things we believe about using AI in a real organisation, based on what we have learnt from helping complex businesses make consequential change work.

1. 01AI is part of the operating model
2. 02AI should make more of people, not simply use fewer of them
3. 03Business outcomes are the priority
4. 04Ethics must be built in
5. 05The new moat is data, trust and corporate intelligence
6. 06Humans remain in control even when autonomy has been earned
7. 07The economics stretch far beyond the pilot
8. 08Don't tie yourself to things you can't control
9. 09Accountability, accountability, accountability
10. 10There is no prize for being first

The important change is not that models have become better at writing. It is that they can now participate in the work. An agent can receive an objective, find information, use software, update a record and trigger the next step. Once it starts doing that, AI is no longer sitting alongside the operating model. It is inside it.

That changes the management question. We need to decide where AI has authority, what information it can see, when it must stop and who owns the outcome. Those are operating-model decisions, not technical settings to be left with the implementation team.

The phrase “digital workforce” makes agents sound like infinitely scalable employees who do not complain, take holidays or require much management. They are none of those things. An experienced employee knows the written process and all the reasons it does not quite work. They recognise the important exception, the awkward customer and the number that looks right but cannot possibly be. An agent has an objective, some context and whatever access we have decided to give it.

If we give it too much, too soon, that is our failure rather than the machine’s.

AI can remove repetitive work, make expertise available more widely and give people better information at the point they need it. The real opportunity is to redesign work so that human judgement, creativity, empathy and relationships are applied where they create the greatest value.

Some tasks will disappear. Some roles will change and some teams may become smaller. We should be honest about that. But treating headcount reduction as the strategy misses most of the value and risks removing the experience the organisation still needs when the process meets reality.

People hold context that was never documented, resolve conflicts the process does not acknowledge and notice when an apparently sensible answer is wrong. Remove that expertise too quickly and the business can become wonderfully efficient at repeating mistakes.

The objective is not to preserve every role exactly as it is. Nor is it to take humans out of the system at any cost. It is to stop wasting people on work that no longer needs them while investing in the judgement and capability that still does.

Do not automate the old organisation and call it the future.

Most organisations do not have an AI problem. They have customers waiting, margin leaking out of processes, people making decisions with incomplete information and expensive expertise tied up doing work that adds very little. Start there.

We are unapologetically enthusiastic about AI when it can change one of those things. We are considerably less interested in an AI strategy built around a list of products, pilots and use cases copied from everybody else’s AI strategy.

Sometimes AI will be central to the answer. Sometimes the answer will be better data, a simpler process or removing a step nobody can remember adding. The business should not care which one wins. A pilot can be exploratory, but it still needs a question: what are we trying to learn, and what decision will we take when we know it? Without that, the pilot is simply a small cost waiting to become a larger one.

Execution and outcomes will always beat strategy and slides. They also beat adoption dashboards.

An AI system can be compliant, secure and commercially attractive and still be the wrong thing to build. Ethics therefore cannot sit with a committee that reviews the finished product shortly before launch. The important decisions have already been made by then: which problem was chosen, whose data was used, what the system optimises, which errors are acceptable and who is expected to live with them.

We need to ask those questions while the work is being designed. Who benefits? Who might be excluded or disadvantaged? What behaviour will the system encourage? What happens to somebody who cannot follow the expected route? Would we be comfortable explaining the decision to the person affected by it?

The answers should shape requirements, data, testing and the decision to go live. They also need to be revisited once the system is operating. Data changes, models change and consequences appear that nobody predicted in the workshop.

Ethics is not one team’s job and it is not a final approval gate. It should sit alongside security, functionality and performance throughout the lifecycle. That is the point of EthSecDevOps: ethics becomes part of how the product is built and run, rather than an apology written after something goes wrong.

Most organisations will have access to broadly the same models. Buying the latest one before everybody else may create a temporary advantage, but it will not create a defensible business. The real value sits in what the organisation brings to the model.

Data provides the context nobody else has: customers, transactions, operations and the history of what actually happened. Trust provides the permission to use AI in decisions and customer journeys that matter. Corporate intelligence is the understanding of what good looks like, how the business actually runs, where the exceptions live and when the answer should be challenged.

Much of that corporate intelligence is currently scattered across processes, documents and experienced people. If it remains there, AI will produce competent, generic output. If it can be identified, codified and applied without flattening the nuance out of it, AI can scale something genuinely distinctive.

This is why the organisation’s advantage must live outside the model. The model should consume and apply that advantage, not become the place where it is stored. Otherwise, the business is helping to build the provider’s moat rather than its own.

We would not give a new employee unrestricted access to every system on their first morning. The same principle should apply to an agent. Give it a defined job. Limit the information and actions available to it. Watch what happens when the data is incomplete, the instruction is ambiguous and the customer does something outside the happy path. Expand its authority only when the evidence supports it.

Generative AI could give you a bad answer. Agentic AI can do something with it. It can act on the wrong source, follow a malicious instruction hidden in a document, repeat a transaction or carry a mistake from one system into another. It can also do all of that far quicker than the organisation can convene a meeting to discuss it.

We do not need a human approval click in front of everything forever. That simply creates a more expensive bottleneck with better technology around it. We need autonomy matched to consequence, with clear permissions, escalation and reversal when something goes wrong.

People do not have to touch every transaction. Someone does have to remain answerable for the system carrying them out.

The software price is the beginning of the AI business case, not the end of it. There is data to prepare, systems to integrate, security to manage, people to train and work to redesign. Agentic AI adds inference, tool use, monitoring, exception handling and the cost of correcting actions that should not have happened.

Benefits require the same discipline. Ten minutes saved is not ten minutes of value unless the organisation can explain what happens to it. More output is not automatically better output. Contact deflected from a service centre is not a success if the customer simply goes elsewhere.

AI programmes have also developed some wonderfully generous methods of counting value. Every generated document becomes a productivity saving. Every employee with a licence becomes an active user. Every completed agent action becomes proof of transformation. Very little of this necessarily reaches revenue, cost or customer experience.

Stress the economics. What happens at ten times the usage? What happens if agent execution costs more than expected? What happens if the provider doubles the price? If the business case disappears, we need to know before the process becomes dependent on it.

The best model for a piece of work today may not be the best one in six months. It may also become unavailable, uneconomic or incompatible with the risk the organisation is prepared to carry. Business logic, proprietary knowledge and critical workflows should sit outside the model wherever practical. We should be able to change provider without rebuilding the business process around it.

This matters more as the large providers approach public markets. Both have funded extraordinary growth and infrastructure, and public investors will expect extraordinary revenue and margin in return. Token prices may continue to fall, but that does not mean the total enterprise cost will. Premium reasoning, agents, tools, memory, security, data residency and minimum commitments all create room for the bill to move in the other direction.

The quickest proof of concept often uses the provider’s entire stack. Its model, tools, memory, agent framework and proprietary features work neatly together and make the demonstration look excellent. Every one of those choices can make leaving harder. There will be occasions when the proprietary capability is worth that dependency — make the decision consciously and be clear about what the business receives in return.

Lock-in discovered during procurement is an architectural failure.

Explainability is important, but it is only one part of running AI properly. The real test is accountability. AI that is part of the operating model must be observable, auditable and defensible. We need to know where models and agents are operating, what information they used, which actions they took, what those actions cost and how their performance is changing. If the system is part of the operation, it needs management information rather than occasional reassurance from the project team.

A customer needs a comprehensible explanation. An operator needs to understand the exception. Risk and audit need a reliable record. The board needs to know whether the system is creating value within the risk appetite it approved. Each audience needs a different level of detail, but the same underlying discipline: observable, auditable and defensible.

We may not always be able to explain every internal step taken by a complex model. We must still be able to defend the system and the decision. Which model was used? What information did it rely on? What was it allowed to do? How is performance being monitored? Can the outcome be challenged or reversed? Those are answerable questions, and answering them is what makes AI accountable.

If the only explanation available is that the model produced the answer, the organisation has surrendered judgement without surrendering liability.

Some organisations are ahead. Some are behind. Most are probably less advanced than their conference presentations suggest.

It is perfectly reasonable not to be at the forefront of every development. Models are improving, costs are changing, regulation is becoming clearer and yesterday’s essential platform is frequently absorbed into something the business already owns. A well-informed fast follower may get better technology, clearer standards and stronger commercial terms without paying for everybody else’s mistakes.

That does not mean doing nothing. Organisations should understand what is changing, test where there is a genuine business case and put the foundations in place around data, governance and capability. There is a difference between deliberate timing and denial.

There is no shame in being behind the hype cycle. There is considerable risk in rushing to the front of the wrong one.

This will change again

This version of the BML ManAIfesto will date. Of course it will. Models will improve, economics will shift, regulation will mature and organisations will find entirely new ways to get this right and wrong.

The test is whether the principles survive those changes. Start with the business. Put people first. Build ethics in. Be explicit about value. Earn autonomy. Keep control of the architecture and the data. Know what the system is doing. Leave accountability with people.

AI is already changing how organisations work. The winners will not be those who talk about it most or buy it fastest. They will be the ones who turn the capability into a better business.

BML helps complex and regulated organisations do exactly that, from identifying where AI can create value through to operating model, governance and delivery.