AI Costs a Fortune. Mismanaging AI Will Cost You More.

Lukic & Goydan / BCG — AI Costs a Fortune. Mismanaging AI Will Cost You More.

▶ Watch on YouTube · Boston Consulting Group · 19:53

Vlad Lukic, BCG’s global leader for tech and digital advantage, and Paul Goydan, global leader of BCG’s cost offer, explain why so many companies spend more on AI than they get back. What is the fix? They argue it isn’t simply cutting AI spend, but assigning clear ownership, categorizing costs correctly, and tying every dollar to a business outcome.

TL;DR

A 20-minute episode of The So What from BCG, hosted by Georgie Frost, from Boston Consulting Group. The guests are Vlad Lukic (global leader, Tech and Digital Advantage) and Paul Goydan (global leader, AI Cost Advantage). The opening claim: “AI cost has gone from the server room to the boardroom in the last two years. If we simply go after AI cost the same way we go after commodity IT spend, we don’t get the value in the business.”

The main points:

• From “tokenmaxxing” to “valuemaxxing”. Companies ran pilots they never switched off and pushed staff to use as many tokens as possible, “and that led to exorbitant bills but low value.” Lukic files the early spend under learning and development. The shift now is to being “very explicit on where is the AI adding into the process.”
• Frontier models on routine work. Goydan: “A frontier model costs 30% more than a good enough model. Yet… we see more often than not very basic rudimentary tasks are being handed over to frontier models. It’s like giving your teenage driver a brand-new Ferrari.” He calls model selection “a whole new capability”, along with managing context-window size.
• Waste and underspend at the same time. Lukic finds “20 to 30% of the dollars being wasted” in the tech stack, because procurement has not caught up with the shift from licences to consumption to tokens. Recovering it (“sweating the stack”) funds new AI. Meanwhile strong lean operators underspend, aiming for “one to two, 3% improvement” when “with AI, you can be thinking 30, 40, 50%.”
• A cost split that moves ownership to the business. Lukic sorts AI spend into CapEx (infrastructure built), OpEx (tools for daily work) and COGS (AI inside products and services sold). That split makes it “fairly easy to then link it into the business ownership.” Goydan: “The business owner needs to be accountable for the cost of deploying AI… the ultimate ROI sits with the business, just like any other investment.” One client’s marketing AI moved into the CMO’s budget, measured on engagement and revenue. BCG’s rule of thumb: for each dollar of value, about 20 cents more goes on the tech stack.
• Measure the outcome, not the task. A client cut a task from 10 days to one, but the customer still waited 10 days, because the process around the task never changed: “there was no forcing function.” Switching the metric to customer response time and cost led them to remove committees and steps and reach one day end-to-end.
• Low-value targets. A manufacturer wanted AI to cut waste that was “less than 1%” of cost of goods sold: “you couldn’t pay for the AI.”
• FOMO as fear of obsolescence. Boards ask “are you consuming enough AI?” Goydan reframes FOMO as “fear of market obsolescence”; Lukic calls it “the right driver”, to be channelled, not cured. He cites BCG figures that firms which have “figured it out” grow 1.5–2x faster, with 3–4x the total shareholder return and 1.5–2x the EBIT.
• Shadow AI can’t be banned. Unapproved tools leak data, widen the attack surface and produce inconsistent data. The answer is to “acknowledge it, educate the employees… and then create effective alternatives.”

Dynamic-capabilities reading

• digital-seizing/balancing-digital-portfolios — the episode is about where to cut and where to spend: shut down dead pilots, recover the 20–30% of stack waste, and move the money to core workflows where lean operators underinvest.
• digital-transforming/redesigning-internal-structures — moving AI cost ownership from the CIO to business owners, with CapEx/OpEx/COGS categories, is a change in who is accountable for what. The 10-days-to-one example shows the process redesign that an outcome metric forced.
• contextual/external-triggers — board pressure and “fear of market obsolescence” are named as the drivers of much current spending.

How it connects

• Two weeks earlier, the same series ran Ellmer & Dhar on why AI rollouts fail, which comes at the same value gap from behaviour change; this episode comes at it from cost ownership. Both belong to enterprise-ai-adoption.
• The frontier-versus-good-enough point is the buyer-side version of the routing argument in Belcak et al. (NVIDIA); see small-language-models.
• The 10-day example (the task gets faster, the outcome does not) is the pattern micro-productivity-trap describes.

What was actually ingested

The full 19:53 episode from the creator-uploaded English captions, used as fetched. YouTube chapter markers are in the raw file.

Linked entities and concepts

• Entities: Boston Consulting Group
• Concepts: enterprise-ai-adoption, small-language-models, micro-productivity-trap
• Dangling (single-source mention, deferred): Vlad Lukic, Paul Goydan, Georgie Frost

Scope and reliability

A consultancy promoting a new practice. Goydan says outright that BCG has “a whole new practice emerging… that didn’t exist six or eight months ago.” The figures (30% frontier premium, 20–30% waste, 1.5–2x growth, 3–4x TSR, 20 cents per dollar) are BCG’s own, given without method, sample or date. The 30% premium in particular is far below the price gap between frontier and small models in most published price lists, so read it as a round number, not a measurement. The value of the episode is the cost split and the ownership argument, not the numbers.