The AI Advantage Is Real. It Is Just Not Where You Are Spending.

The AI Advantage Is Real. It Is Just Not Where You Are Spending. | Trazup

The gap, in three numbers

3.6x

Three-year shareholder return of the top 5% of AI adopters versus laggards

73%

Of enterprises that blew past their original AI cost projections in 2026

90 days

Mid-market time from pilot to production. Large enterprises take nine months or more

Sources: BCG (n=1,250) · FinOps Foundation · MIT Project NANDA

There are two stories being told about AI in business right now, and both of them are true. The first is that a small group of companies is pulling away from everyone else at a speed that has no recent precedent. The second is that most AI spending produces nothing you can find in a P&L. Reconciling those two facts is the whole strategy question, and it matters most to the companies with the least room for error.

Large enterprises can absorb a failed AI program. It becomes a line item, a reorganisation, a slide in next year's deck. A 60-person company cannot. If you run one, the useful question is not whether AI creates advantage. It does, measurably. The question is whether you can buy that advantage at a price that does not sink you, and the answer to that turns out to be more encouraging than the headlines suggest.

The advantage is real, and it compounds

BCG surveyed 1,250 senior leaders and sorted them into three groups. About 5% qualified as future-built, roughly 35% were scaling, and 60% were seeing almost nothing. The financial separation between the top group and the bottom is not a rounding error, and it shows up across four different measures at once.

Figure 1 · Leaders vs laggards

The top 5% are not slightly ahead

Performance of AI leaders as a multiple of laggards. A value of 1.0 means no difference.

BCG, The Widening AI Value Gap: Build for the Future (2025), n=1,250.

The return numbers point the same way. IDC's research for Microsoft puts the average at $3.70 back per dollar invested, with leading adopters near $10 and a median time to positive ROI of 13 to 14 months. That is a good return by any capital allocation standard. It is also a return that only about one in twenty organisations is actually collecting.

Almost nobody is capturing it

McKinsey surveyed nearly 2,000 executives across 105 countries and found 88% of organisations using AI in at least one function. Only 39% reported any EBIT impact at all, and roughly 6% attributed more than 5% of EBIT to AI. Deloitte's survey of 3,235 leaders across 24 countries found two thirds reporting productivity gains, a fifth reporting revenue growth, and a quarter saying initiatives hit the ROI executives had expected.

Figure 2 · The value funnel

Adoption is near universal. Value capture is in single digits.

Share of organisations at each stage, from using AI somewhere to moving the P&L in a way finance recognises.

McKinsey State of AI 2025 (n=1,993); BCG AI Radar 2026; MIT Project NANDA (2025).

The abandonment data is moving in the same direction. S&P Global found 42% of companies dropped most of their AI initiatives in 2025, up from 17% the year before. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, attributing it to escalating costs, unclear business value and weak risk controls, in that order.

Adoption is not the moat. Almost everyone has adopted. Getting one workflow into production and keeping it there is the moat.

Speed is the real advantage, and it favours the small

Buried inside the MIT Project NANDA report is the finding that should reorganise how a smaller company thinks about this. Mid-market firms moved from pilot to production in about 90 days. Large enterprises took nine months or longer. Same technology, same vendors, same talent market. The difference is the number of committees between a decision and a deployment.

Figure 3 · Cycle time

Pilot to production, by company size

Days to get a working AI capability into live operations with real users.

MIT Project NANDA, The GenAI Divide (July 2025). Enterprise figure is a floor, not an average.

In a market where model capability resets every few months, a three-month cycle beats a nine-month cycle even when the nine-month version is technically better. The large company finishes building for a world that has moved. The small one has already learned what does not work and is on its second attempt.

The same report found that buying from specialised vendors succeeded roughly 67% of the time, while internal builds succeeded about a third as often. Enterprises keep building anyway, particularly in regulated industries. A smaller company that has no engineering bench is effectively forced into the strategy with twice the success rate.

Figure 4 · Buy vs build

The constraint is doing you a favour

Share of deployments that reached production with measurable impact.

MIT Project NANDA (2025). Build figure is approximately one third of the buy rate.

The cost trap that catches everyone

Here is where smaller companies get hurt, and it is rarely the sticker price. Most AI business cases cover model spend, cloud and platform tooling, then stop. Licensing and model fees are usually a minority of what the thing costs to run.

What arrives later is data remediation, which is almost always the largest single effort line. Then integration into the systems that actually matter. Then evaluation, testing and red teaming. Then governance and monitoring, now consuming 8 to 12% of the average AI budget against 3 to 5% in 2024. Then training and process redesign. Then the maintenance nobody budgets for, because model versions get deprecated and prompts rot.

Figure 5 · Total cost of ownership

Where the money actually goes

Illustrative allocation of a production AI workflow's first-year cost. Governance share is sourced; the rest reflects typical delivery patterns.

Governance range from 2026 enterprise AI budget research. Remaining split is illustrative and will vary by data maturity.

The FinOps Foundation found 73% of enterprises exceeded their original AI cost projections in 2026. Other survey work put the average overrun near 31%, with a quarter of organisations off by half or more. And the reason is genuinely counterintuitive: unit prices are collapsing while invoices go up.

One analysis of 2.4 billion enterprise API calls showed blended cost per million tokens falling 67% year over year. Consumption grew faster than price fell. Agentic workflows burn an order of magnitude more tokens per task than a chatbot does. Uber gave 5,000 engineers access to an AI coding tool in December 2025 and burned the entire annual AI budget by April.

Figure 6 · The pricing paradox

Cheaper per unit, more expensive per month

Indexed unit price against indexed total spend. This is why seat-based budgeting fails on a usage-based cost.

Directional, based on reported 67% year over year decline in blended cost per million tokens against rising enterprise invoices.

The single most controllable variable in that equation is routing. Organisations that match workloads to the right model by task landed near $2.31 per million tokens. Organisations that send everything to the best available model paid closer to $18.40 for identical work. The spread between the cheapest production model and a frontier reasoning model runs past 4,000x.

Figure 7 · Same work, different bill

Routing is an 8x decision

Task-routed workloads$2.31 / M tokens

Everything to the frontier model$18.40 / M tokens

For a smaller company, this is often the difference between an AI capability that pays for itself and one that quietly becomes the third largest software line on the books.

Blended production rates reported across enterprise workloads, 2026.

Where smaller companies actually stand

The adoption numbers for small and mid-sized businesses depend entirely on how you ask. The US Chamber of Commerce reports 58% of small businesses using generative AI, up from 40% the year before. Census Bureau data tracking active use in production operations puts it closer to 17 to 20%. Both are correct. One measures experimentation, the other measures work that has actually changed.

Goldman Sachs found 76% of small businesses reporting AI use and 14% saying it is fully embedded in core operations. That 62-point gap is the entire opportunity, and it is not a technology gap.

Figure 8 · Barriers

What actually stops smaller firms

Reasons cited by non-adopting small businesses. Cost does not lead the list.

SBA Office of Advocacy (2025). Respondents could cite more than one barrier.

Notice what is missing from the top of that list. The dominant barrier is not price and not infrastructure. It is 77% of non-adopters not seeing a use case that applies to them. Among firms under five employees that number reaches 82%. Meanwhile the small businesses that did adopt report the outcomes you would want: Salesforce found 91% saying it lifted revenue and 86% saying it improved margins, and growing firms were far more likely to have adopted than declining ones.

The adoption gap between small and large firms narrowed from 1.8x to 1.2x in a single year. No previous technology cycle closed that fast.

How a smaller company should fund this

The practical fix is not more governance. It is refusing to run three different kinds of work under one set of expectations. Most of the reported failure rate is capability-building work being measured with cost-reduction expectations.

Tier	What it covers	How to fund it	How you judge it
Efficiency	Back office, document handling, support workflows	Full funding against a hard cost baseline	Cost per transaction down, outside spend actually removed
Growth	Customer and revenue-facing workflows	Staged funding against leading indicators	Conversion or cycle time improvement that holds two quarters
Learning	Capability building, emerging tools	Capped and time-boxed, no ROI test	A documented decision to kill or scale inside 90 days

One more discipline matters more for a small company than a large one, because a large one can survive getting it wrong. Before you fund anything, run it at three times current usage and see whether the case still holds.

The 3x stress test

If your AI spend triples and comes in 31% over plan, does the case still work? Adjust the numbers for your own situation.

Current monthly AI spend ($)

Monthly value you can defend ($)

Workload type

Model routing discipline

Annual spend today

Annual spend at 3x usage, plus 31% overrun

0.0x

Return on the stressed number

Adjust the inputs to see the verdict.

The five moves that decide it

Pick one workflow with a bill attached to it

Not a department, not a tool rollout. One process where you already know the cost per transaction, so improvement is arithmetic rather than argument. Capture the baseline before you start, never after.

Buy the commodity, partner for depth, build almost nothing

Vendor deployments convert at roughly twice the rate of internal builds. Build only where the process is genuinely proprietary, which for most companies is one or two things, not ten.

Route by task and cap the spend

Send routine extraction and classification to cheap models. Reserve frontier reasoning for the steps that need it. Set hard usage caps and alerts on day one, not after the first surprise invoice.

Redesign the work, do not decorate it

BCG's 10-20-70 rule holds that 10% of the value comes from the algorithm, 20% from data and technology, and 70% from people and process. Most budgets are allocated in almost the reverse proportion.

Bank the savings somewhere finance can see them

The most common way a good case turns into a bad outcome is counting hours saved that never leave the cost base. Decide in advance where freed capacity goes: reduced outside spend, more volume with the same team, or a role you no longer need to hire.

What to ask before you sign anything

The failure rate is not evidence that AI does not work. It is evidence that most organisations are running AI as a technology programme when the things going wrong are operational, financial and organisational. Three questions will tell a smaller company where it stands faster than any assessment.

Which line does this move, and who owns that line? If nobody's number changes, nobody will defend the budget when it gets tight.
What does this cost at three times current usage? If that number ends the conversation, you found out now instead of in month seven.
What work are we redesigning rather than decorating? A tool bolted onto an unchanged process produces a good demo and no result.

The uncomfortable truth for large enterprises is that their advantage in this cycle is smaller than they think. Capital does not buy speed, and speed is what is scarce. The opportunity for everyone else is that the winning strategy, buy rather than build, one workflow at a time, hard baselines, tight cost control, is the only strategy a smaller company could afford to run anyway.

Being constrained is not the disadvantage here. Acting like you are not is.

Sources: MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (July 2025); McKinsey State of AI 2025 (n=1,993); BCG The Widening AI Value Gap (n=1,250) and AI Radar 2026; Deloitte State of AI in the Enterprise 2026 (n=3,235); Gartner agentic AI cancellation forecast and 2026 CFO priorities survey; IDC and Microsoft, The Business Opportunity of AI; S&P Global; FinOps Foundation State of FinOps 2026; US Census Bureau BTOS (2026); US Chamber of Commerce small business technology survey; SBA Office of Advocacy (2025); Salesforce SMB Trends; Goldman Sachs 10,000 Small Businesses (2026). Figures marked illustrative are modelled, not surveyed.

Top stories this week