What does it mean to invert the approval hierarchy?

In a traditional hierarchy, permission flows down: an idea must be approved at the top before anyone may try it. In the inverted model, evidence flows up: people at any level may run small AI experiments within set limits, and leadership decides which proven ones to adopt and fund. The top stops rationing attempts and starts selecting results.

Doesn't letting everyone experiment with AI create risk?

Only if there are no limits. The model depends on a short, firm set of boundaries: approved tools, what data may be used, no unattended actions that reach customers or move money, and a single official source for each key metric. Inside those lines, no permission is needed. Outside them, the normal approval applies.

How can AI help executives with approvals?

By preparing the decision. Each candidate initiative arrives as a standard one-page brief, largely generated from the experiment's own records: the problem, what was tried, measured benefit, data and systems touched, risks, reversibility, and the cost to adopt. Leaders spend their time on the sense check rather than on assembling information.

Where should an organisation start?

With one team and ninety days. Publish the limits, open a simple register, review it monthly, and promote one or two experiments at the end. Expand only after the first cycle has produced something the wider organisation can see.

Inverting the Approval Pyramid: A Governance Model for Bottom-Up AI

We wrote recently that AI isn’t the bottleneck, leadership attention is. That piece was a diagnosis. This one is the treatment we recommend.

The short version: most organisations still run an approval model built for a world in which trying something was expensive. It no longer is. The model needs turning over.

The pyramid as it stands

	Traditional model
Who may start something new	Whoever secures executive sponsorship
What travels up	Proposals and requests for permission
What travels down	Approval and budget
Capacity limit	The number of decisions a few senior people can make
What happens to good ideas at the edge	They wait, shrink, or leave with the person who had them

Middle managers sometimes create pockets of initiative, and those pockets tend to close when the manager moves on or a new corporate programme arrives. Nothing in the structure holds them open.

Why AI breaks this model

Two things change at once.

Attempts become cheap. A staff member with a company AI seat can build a working prototype of a process improvement in hours, with no budget and no vendor. Rationing attempts through an executive gate now costs more than the attempts themselves.

Volume rises. If everyone can try things, the number of candidate ideas multiplies. A gate that was already saturated cannot absorb ten times the flow. Either ideas die in the queue or leaders approve without really looking.

Giving executives AI tools helps them decide better. It does not change the arithmetic. The structure has to change.

The inverted model

Five parts. None is complicated. All five are needed.

1. A sandbox with firm edges

Publish what anyone may do without asking. We keep it to a page.

Tools: the company’s approved AI plan and connectors
Data: what may be used freely, what needs care, what is never used
Actions: nothing that contacts a customer or supplier, moves money or changes a system of record without a named person’s approval
Numbers: exploratory analysis is labelled as such, and the official figure stays official

Inside those edges, permission is not required. That single sentence is the inversion.

2. A register, not a request form

One shared list. For each experiment: who, what problem, what was tried, rough time saved or value created, status. Two minutes to fill in. It exists so that good ideas are visible and duplicated effort is caught, not so that anyone can say no.

3. A promotion path with stated criteria

An experiment is put forward for adoption when it meets any of these:

more than one person or team wants to use it
it needs to read from or write to a core system
it produces something that goes outside the organisation
it would change a documented process

Below that line it stays a personal improvement, and that is a perfectly good outcome.

4. AI-prepared approval briefs

Every promoted candidate reaches leadership in the same one-page format, drafted by AI from the register entry and the experiment’s own files, then checked by its owner:

The problem, in a sentence
What was built and how long it has been running
Measured benefit
Data and systems touched
Risks, and what happens if it fails
How easily it can be switched off
What adoption would cost in money and people

Leaders stop assembling information and start judging it.

5. A different job at the top

	Traditional	Inverted
Leaders decide	What may be tried	What gets adopted and funded
Leaders receive	Proposals	Evidence
Leaders’ main output	Permissions	Boundaries, sense checks, sponsorship
Time per decision	Long, information-gathering	Short, judgement-focused

This is a narrower role and a more valuable one. It also scales, because a one-page brief on something already working takes minutes to assess.

Decision rights, made explicit

Ambiguity about who can approve what is where these programmes stall. Write it down.

Tier	Example	Who approves
Personal	A routine that drafts my weekly report	Nobody. It is inside the sandbox
Team	A shared workflow for the whole finance team	Team lead
Cross-team or system-touching	An automation that reads the accounting system	Function head, with IT sign-off on access
External or financial	Anything sending to customers or preparing payments	Executive, with a human approval step kept in the process

What stops it working

Leaders who cannot let go. The model asks senior people to be consulted less. Boards and chief executives have to want that, say so, and then behave accordingly the first time someone acts without asking.

Punishing the fizzles. Most experiments lead nowhere, which is why they need to be cheap. If a failed experiment counts against someone, the register will be empty within a quarter.

Initiative that depends on one sponsor. The register and the promotion path exist so that a good idea survives its champion leaving.

No boundaries. Freedom without edges produces data incidents and conflicting numbers, and then a clampdown. See our guide to AI security and governance for the policy underneath the sandbox.

Start with one team

Choose a team with a willing lead.
Publish the one-page sandbox and open the register.
Run for ninety days, with a thirty-minute review each month.
At the end, promote one or two experiments using the standard brief.
Tell the rest of the organisation what happened, including what did not work.

Then widen it, one team at a time. It is a significant shift in how an organisation thinks about authority, and it lands best in small steps.

If you would like help drafting the sandbox, the register and the brief for your organisation, talk to us.