The bottleneck has moved. It may now be you.

You're reading The AI Directive — practical AI intelligence for
business leaders. Subscribe
here if someone forwarded this.

A business introduces AI into commercial planning.

The team can now combine market data, customer performance,
competitor activity and financial scenarios in hours rather than days.
It can test assumptions, expose risks and prepare a credible
recommendation before the old process would have produced its first
draft.

The result is not a faster commercial decision.

It is a larger decision pack.

More scenarios are requested because they are now cheap to produce.
More stakeholders comment because the material is easier to circulate.
Finance, operations, sales and technology each add a condition. The
decision moves upwards because nobody is quite sure who owns the
trade-off between them.

The senior team receives better information than it has ever had.

It also receives more decisions than it can absorb.

This is not a meeting-scheduling problem. Nor is it simply a case of
executives needing to answer their emails faster.

It is a structural problem.

AI has increased the organisation's capacity to prepare, analyse and
recommend. But the authority, risk model and management capacity around
those decisions were designed for a world in which good analysis was
slower and more expensive.

The bottleneck has moved.

And in many businesses, it has moved upwards.

Last week, I argued that making one task faster does not make the
whole business faster. People, process, data and technology have to
change around it.

Decision-making is where that argument becomes uncomfortable for
leaders.

The same four elements are present here. People need
the judgement and confidence to own a choice. Process
determines where authority sits, what must be consulted and when an
issue escalates. Data provides the evidence—but can
also become an excuse to defer. Technology increases
the speed and volume at which options, analysis and recommendations
arrive.

Change only the technology and the other three do not remain neutral.
The process clogs with more material, people push risk upwards and data
expands beyond what the decision genuinely requires. The organisation
has improved its ability to produce an answer without improving its
ability to act on one.

AI can reduce the effort required to assemble evidence, compare
options and draft a recommendation. That is useful. But it also removes
a natural constraint.

When analysis was expensive, organisations rationed it. A team might
prepare three scenarios because a fourth required another week of work.
An investment paper might contain the evidence that could reasonably be
assembled by the deadline. A manager would decide whether an issue
genuinely warranted escalation because escalation created work.

AI changes those economics.

Ten scenarios become possible. Every assumption can be
sensitivity-tested. Every stakeholder can request a slightly different
view. More exceptions can be identified and more recommendations can be
generated.

The supply of material increases.

The supply of executive attention does not.

Atlassian describes a version of this as the AI efficiency paradox:
output accelerates, then accumulates at reviews, approvals and other
human-judgement gates. Its 2026 research found that 89 percent of
executives believed AI had increased the speed of work, while only 6
percent were confident they could identify organisation-wide ROI.

That should not be read as an argument against AI.

It is an argument for redesigning the part of the organisation that
AI has exposed.

Most organisations do not describe themselves as slow to decide.

They say they are being thorough.

They need one more view from Finance. Legal has not formally signed
it off. Operations wants the downside modelled. The executive sponsor
would like another option. The data is not quite complete. The issue
should go through the normal governance route.

Sometimes all of that is justified.

An irreversible acquisition, a major safety decision or a material
regulatory exposure deserves more care than a reversible pricing test.
Serious leadership involves knowing when speed would be reckless.

But much of what passes for diligence is something else:

AI can make this worse by supplying almost unlimited intellectual
camouflage.

A decision can always support another scenario, another comparison or
another polished paper. Deferral starts to look evidence-led.

But there is a point at which the issue is no longer insufficient
information.

It is an unwillingness to choose.

Telling managers to make decisions faster is not a serious
answer.

People delay when the system makes delay rational.

If the consequences of a poor decision are personal but the
consequences of waiting are distributed across the business, waiting is
safer.

If authority is vague, consultation protects the individual.

If every exception has historically travelled upwards, teams learn
that judgement belongs above them.

If governance measures whether every box was ticked but not whether
value was lost through delay, the process will optimise for compliance
with the route rather than quality of the outcome.

This is why decision-making has to be designed as part of the
operating model.

For recurring decisions, the organisation should be able to
answer:

If those answers exist only in habit, hierarchy or the memory of
experienced people, AI will not repair the system.

It will feed it faster.

One of the most useful distinctions in decision-making is also one of
the most frequently ignored: can the decision be reversed?

A controlled customer trial, a limited marketing test or a temporary
workflow change can often be monitored and corrected. The cost of delay
may be greater than the cost of being imperfect.

A large capital commitment, a regulatory submission or an action with
serious safety consequences is different. The evidential standard should
be higher because correction may be difficult or impossible.

Yet many businesses apply broadly the same machinery to both.

The reversible decision inherits the controls of the irreversible
one. The low-value exception travels through the same layers as the
strategic commitment. Senior attention is consumed by matters that could
have been bounded through clear limits, live measures and an explicit
stop rule.

That is not prudent governance.

It is a failure to design governance in proportion to risk.

AI makes proportionality more important because the volume of
possible actions and recommendations will continue to increase. If every
AI-assisted recommendation enters the old approval architecture, the
centre will be overwhelmed.

The answer is not indiscriminate delegation.

It is clearer authority within better-designed boundaries.

There is a fashionable but crude argument that AI will eliminate
layers of middle management.

It confuses management with information relay.

Some managerial activity has indeed consisted of collecting updates,
chasing inputs, condensing information, reformatting it and passing it
upwards. AI can reduce much of that work.

But that is not the most valuable part of management.

Good managers create context. They frame choices. They resolve
competing priorities. They coach judgement. They notice when a local
decision has wider consequences. They allocate scarce resources and own
the outcome when the evidence is incomplete.

As routine preparation becomes cheaper, those capabilities become
more important, not less.

The role changes from moving information through the organisation to
helping the organisation act intelligently on it.

That also changes how management capacity should be assessed.

Headcount alone tells us very little. A manager with a small team may
face high ambiguity, tightly coupled processes, frequent exceptions and
substantial risk. Another with a much larger team may oversee stable,
standardised and largely reversible work.

Gartner calls this a span of complexity, rather than
simply a span of control. It is a useful distinction. Add AI agents,
faster output and more exceptions, and the number of people on an
organisation chart becomes an even poorer measure of managerial
load.

Widening spans while ignoring complexity may reduce boxes on a chart
and increase the burden at every remaining decision point.

That is cosmetic efficiency, not operating-model redesign.

This is the part many AI programmes avoid.

It is easier to ask employees to adopt new tools than to ask the
executive team which decisions it should stop owning.

It is easier to automate the preparation of a board paper than to
challenge why the paper needs twelve contributors.

It is easier to commission better management information than to
decide which measures no longer matter.

It is easier to add an AI governance group than to remove duplicated
authority from the governance already in place.

After 30 years around technology and transformation, I have seen this
pattern repeatedly. Organisations redesign the work below the management
layer while treating management itself as fixed.

The system changes around leaders, but not through them.

That leaves the most expensive and consequential constraints
untouched.

If AI is genuinely changing how quickly evidence can be assembled and
work can be prepared, senior leaders should expect changes to their own
agendas, information flows, escalation rules and decision rights.

Otherwise the organisation is not adopting a new operating model.

It is sending faster work into an old one.

Take ten recent decisions that materially affected delivery,
customers, cost, revenue or risk. Ask who owned each outcome, who was
actually allowed to decide, why it moved upwards and whether the
evidence requested changed the choice or merely delayed it.

Review each through the same four lenses as the wider operating
model: whether the people had the judgement and
confidence to decide; whether the process gave them
clear authority; whether the data was sufficient and
decision-relevant; and whether the technology helped
action or simply generated more material for the system to absorb.

The pattern matters more than the individual case. Repeated
escalation of reversible decisions, fragmented vetoes and demands for
further analysis after the trade-off is already clear are signs that the
constraint is structural.

The test is simple: can the management system convert better
information into action, or does it merely absorb more of it?

Take ten recent decisions and reconstruct the authority, evidence,
risk and escalation behind them.

If the same decisions repeatedly travel upwards, do not ask why your
managers are too cautious.

Ask what your operating model has taught them about who is allowed to
decide.

AI can prepare more work, more evidence and more recommendations than
the old management system was designed to absorb.

The answer is not simply faster approval. It is clearer authority,
proportionate risk and stronger judgement closer to the work.

But that raises a bigger question.

Why was the organisation built to gather, condense and move so much
information through so many layers in the first place?

That is where this arc goes next.

See you Tuesday.

— Toby

@TobySmithFirth on
YouTube

Subscribe to The AI Directive here: https://theaidirective.co.uk/subscribe?utm_source=beehiiv&utm_medium=email&utm_campaign=issue_019

These are tools I use personally. Affiliate links marked - I earn a small commission if you sign up, at no extra cost to you.

Some links in this issue are affiliate links. I only recommend tools I actually use.