The Real AI Risk Is Not Underspending. It Is Spending Without Ownership

Most AI budgets do not fail at the model.

They fail at the part nobody wants to fund.

And more money makes that failure bigger, not smaller.

There is a strong argument circulating right now that caution has become the expensive option in AI. BCG has made it well. Restraint that looks like discipline, controlled pilots, tight licence management, and an ironclad business case before scaling increasingly reads as a competitive disadvantage. The performance evidence behind it is serious. BCG’s research reports that AI leaders deliver roughly three times the cost reduction of their peers, about 1.6 times higher EBIT margins, and around 2.7 times the return on invested capital.

I agree with the direction. Underinvestment has quietly killed more programmes than overinvestment ever did, and I have spent twenty-five years in enterprise IT operations and automation delivery watching it happen across twenty-plus countries.

But the conclusion executives draw from that argument is usually the wrong one. They hear “spend more” and reach for a benchmark percentage to justify a bigger line item.

That is not the lesson. Consider the other finding in the same body of research. Roughly 60% of companies report minimal or no material value from AI despite substantial spending. Substantial spending. If money were the binding constraint, that number would be small. It is not.

Underspending is a symptom. Ownership is the disease.

So before you argue about how much to spend, answer three questions. They take about ninety minutes with the right people in the room, and they are worth more than any benchmark.

Question one: where does your value leak?

Most AI programmes leak at the same four gates.

Promised. The business case built to win funding. Optimistic by construction.

Piloted. Something works in a controlled environment with a friendly user group and clean inputs.

Production. It survives real volume, real data quality, real exception handling, real compliance review.

Permanent. It is still running eighteen months later, without the original team babysitting it, and the savings have actually been removed from a budget line.

BCG’s research puts a hard edge on this. Companies it classifies as future-built push roughly 62% of their AI initiatives into production. Laggards manage about 12%.

That is the gap that matters. It is not a spending gap. It is a gate three gap.

The arithmetic is unforgiving. If you lose most of your initiatives between pilot and production, doubling the pilot budget doubles the waste. You will generate twice the activity, twice the vendor invoices, and twice the disappointment.

Find the gate first. Fix the gate. Then fund.

Most executives cannot name their own conversion rate from pilot to production. That is the first thing I would put on a steering committee agenda, ahead of any budget discussion. It is a number your organisation already has. It has simply never been asked for, because nobody wants the answer in the same meeting as the funding request.

Question two: what do you actually own?

Five locks. Model, data, talent, cost, exit.

Model. Can you switch providers without rewriting your applications? Is your orchestration layer yours, or is it a vendor’s abstraction that happens to work today?

Data. Do your embeddings, evaluation sets, and encoded process knowledge live somewhere you control? Or inside a platform that becomes more valuable to the vendor every quarter you stay?

Talent. Can your own people modify what is running in production? Or does every change route through a partner statement of work with a six-week lead time?

Cost. Do you know your unit economics per workflow, per token, per resolved case? Or do you know your monthly invoice and nothing underneath it?

Exit. What does it cost to leave, in months and in currency? If you cannot answer, you do not have a supplier. You have a landlord.

If a vendor holds four of those five locks, you are not building capability. You are renting a dependency and calling it strategy. The dependency reprices at renewal, which arrives precisely when your operating model has become inseparable from it.

This is not an argument against vendors. I run a tech-agnostic delivery firm. Vendors are how you move fast, and moving fast matters. It is an argument for knowing which locks you have traded away, deliberately, in exchange for what.

Worth noting that most definitions of AI spend, including BCG’s, include fees paid to third parties. A meaningful share of any AI budget leaves the building permanently. Ownership is what determines whether the remainder compounds into capability or simply funds next year’s renewal.

Question three: what are you measuring?

Three tiers. Output, outcome, impact.

Output is activity. Tickets summarised. Documents generated. Seats licensed. Agents deployed. Output is what pilots report, because output is easy to produce and almost always positive.

Outcome is a business number. Handling time down. First contact resolution up. Cycle time down. Defect rate down. Outcome requires a baseline captured before you started, which is why most programmes cannot produce one.

Impact is what survives the next budget cycle. Headcount reallocated. A contract cancelled. A margin point defended in the audited accounts. Impact is the only tier a CFO recognises without translation.

If your AI reporting stops at output, spending more will only buy you more output.

This is the mechanism behind that 60% reporting no material value. Those organisations are not lying, and their pilots were not failures. They measured the wrong tier and discovered it at the end of the year, when the savings could not be located in the accounts.

Set the ladder before you set the budget. Every funded initiative should name, in advance, the outcome metric it will move and the impact line it will eventually touch. Initiatives that cannot name either are experiments. Fund some. Just call them what they are.

The seventy percent nobody wants to fund

BCG puts the value equation at roughly 10% algorithms, 20% technology and data, and 70% people, process, and behaviour change.

That last number is the whole game.

It is also the line item that never survives a budget review, because it has no vendor logo attached and no demo to show the board. You cannot procure it in a quarter. It shows up as governance redesign, process simplification, decision rights, upskilling, and the deeply unglamorous work of changing how thousands of people do their jobs.

Do the arithmetic on your own business. It is uncomfortable.

Take a company with two billion in revenue and an AI budget set at 1.7% of that, which is around thirty-four million. Apply the split, and a serious programme looks like this. Roughly three and a half million on models. Roughly seven million on technology and data. And close to twenty-four million on governance, process redesign, upskilling, and behaviour change.

Now open your actual plan and try to find the twenty-four million.

In most plans I have reviewed, that line is a fraction of what the ratio implies, and a good part of it is training booked to satisfy a compliance objective rather than to change how work gets done. The technology and licence lines are fully funded. The change line is a rounding error. Then the programme underdelivers, and the conclusion drawn is that the technology disappointed.

The technology did not disappoint. It was deployed into an operating model that was never resourced to absorb it.

This is the most useful test I know for reading an AI budget. Not the total. The ratio. Show me the split between technology spend and change spend, and I can estimate the probability that a programme reaches gate four without knowing anything about your industry, your vendors, or your architecture.

If the ratio is inverted, more money makes things worse. You are accelerating deployment into an organisation that cannot metabolise what you have already deployed.

There is a supporting signal worth watching. In BCG’s survey work, confidence that AI will pay off drops the further you sit from the corner office, from 62% among CEOs to 48% among executives outside the C suite.

Read that as diagnostic, not as resistance. The people closer to the work often have the more accurate read on what delivery will actually require. If your executive conviction curve and your operations confidence curve are diverging, you have located your seventy percent problem. It will not close itself with a larger licence.

Fund it by taking cost out first

The most useful point in the BCG piece is the one least likely to be quoted. Fund AI transformation by removing cost elsewhere.

Traditional levers, procurement, organisational layers, external spend, service simplification, typically release between 5% and 25% within three to twelve months. That is real money, and it arrives before any AI benefit does.

This creates the cycle that actually works. Traditional savings fund the transformation. The transformation creates structural efficiency. Structural efficiency funds the next wave.

Self-funded transformation has a second advantage that rarely gets mentioned. It survives a downturn. A programme financed from released cost has already proven it can generate cash. A programme financed from a growth budget is the first thing cut when the quarter turns, usually right at gate three, which is exactly where value dies.

If your AI plan requires new money and releases no cost in the first twelve months, you do not have a transformation. You have an experiment with a run rate.

Be loose on the forecast, strict on the ladder

There is a real argument that leaders should invest before fully mapping returns, and that excessive precision in business cases becomes a constraint in fast-moving environments. That is right.

Second-order effects genuinely do not model in a spreadsheet. You cannot forecast what becomes possible when cycle time drops by half, because the new possibilities are not in the current process map.

But there is a difference between refusing to forecast and refusing to measure. One is realism. The other is negligence. In a steering committee, they look identical for about three-quarters.

The discipline that resolves it is simple. Be loose on the forecast. Be strict on the ladder. Fund the bet without a five-year net present value, then hold it to output, outcome, and impact on a fixed cadence, and stop it without ceremony when it stalls at a gate.

The pressure runs the other way at the moment. Survey evidence suggests the overwhelming majority of companies intend to keep investing in AI at current or higher levels even if the investment does not pay off in the near term, and around half of CEOs believe their own position depends on getting AI right.

When nobody can afford to appear cautious, budget discipline stops being a natural brake. The discipline has to come from measurement instead. That is a governance design problem, and it belongs on the board agenda rather than in a project office.

How to read the research

One practical note, because this shapes every conversation above.

Benchmark percentages in AI research are usually averages across a survey population, and averages move every cycle. Matching one tells you that you have caught up with the mean. It does not tell you that you have built an advantage. Before a figure enters your budget defence, trace it back to what was measured and who was sampled.

The performance gaps deserve the same care. Companies that generate strong returns from AI tend to have better governance, cleaner data estates, faster decision rights, and executives already competent at running change. Did heavy investment create those conditions, or did those conditions make heavy investment work?

Almost certainly both. But the second direction is stronger than most summaries suggest. Give a poorly run organisation a large AI budget, and you do not get a leader. You get a well-funded version of the same problems.

That is not an argument for caution. It is an argument for sequence.

What this means for each seat

If you are the CEO. Your job is not to approve a percentage. It is to decide which two or three core processes you are willing to rebuild rather than augment. Most AI value sits in core functions rather than support functions. Pick the core, fund the seventy percent, and accept that this is an operating model decision wearing a technology costume. It will cost political capital, not just budget.

If you are the CIO. Own the gates and the locks. You are the only executive positioned to know the real pilot-to-production conversion rate and the real cost of exit. Publish both. A CIO who can show the board where value leaks and what a switch would cost has moved from order taker to capital allocator. Insist that token and inference economics enter the financial model properly, because unit costs that scale with usage behave nothing like the licence costs your budgeting process was built for.

If you are the CTO. Your deliverable is optionality. Abstraction layers so the model tier stays replaceable. Evaluation harnesses so quality is measured rather than asserted. Data and process knowledge held in formats you control. Every architectural decision either buys a lock or sells one. Know which you are doing, and say it out loud in the design review.

The line

Caution is expensive. That much is settled.

But spending more is not conviction. Spending with ownership is.

The organisations that compound an advantage are not the ones that hit a benchmark percentage. They are the ones that know their leak point, hold their five locks, fund the seventy percent, and measure impact rather than output. Those organisations will beat you at a lower spend level, and some of them already are.

Cut the pilot count. Fund the operating model. Keep the exit.

Build capability you own, not a subscription you renew.

OG Approved. No BS.