What is the McKinsey 60-70% figure actually saying?

McKinsey's 'Economic potential of generative AI' analysis estimates that 60-70% of the addressable economic value from LLMs sits in a small number of knowledge-work functions (customer ops, sales support, software, marketing). That is a top-down estimate of total addressable economy-wide value, not a guarantee any individual business will capture 60-70% productivity uplift across its operation. The two are routinely conflated by AI marketing.

What is the realistic AI productivity uplift for an Australian SMB?

Typical realised productivity uplifts are 10-30% on the specific workflows AI fits. On a well-rolled-out Claude implementation, that translates to a meaningful operating margin lift across the business - usually 5-15% on EBITDA terms once the workflows compound. Larger numbers (30-50%) are realistic on specific high-leverage workflows but not on every workflow simultaneously.

Why does the gap between potential and realised value exist?

Three reasons. Workflow selection - businesses pick AI workflows that do not fit AI's strengths, or pick the wrong starting point. Governance - tools rolled out without policy, oversight or training underperform. Change management - the 90-day adoption sprint that makes AI rollout stick is the work most often skipped. Close all three gaps and realised value approaches the potential. Skip them and the headline numbers stay theoretical.

Are the bigger productivity claims fake?

Not fake, but cherry-picked. Vendors and AI consultancies often quote the peak workflow uplift (a 5-10x speedup on the single best-fit task) as if it were the average. The peak is real; the average across an SMB's operations is much lower. Honest reporting separates the peak workflow from the average across operations.

How should an SMB owner think about AI ROI?

Three rules. Calibrate to your stage of adoption (Reactive tier expects 3-8% individual uplift; Embedded tier expects 10-30% across the business). Measure wage-equivalent savings on specific workflows, not handwave productivity. Plan for the gap between potential and realised value to close over 2-3 years of compounding rollout, not in a quarter.

← All insights

AI Strategy·8 min read

The AI productivity myth: where the McKinsey numbers do and do not apply

The McKinsey figure is everywhere. “Generative AI could add $2.6 to $4.4 trillion annually” - or that 60-70% of LLM-addressable value sits in a few knowledge-work functions. It is genuinely correct. It is also routinely misapplied as if any individual business will capture those numbers next quarter. Here is the honest read on what AI productivity uplifts look like in real Australian SMBs.

What the McKinsey number actually says

The headline figure - 60-70% of LLM-addressable economic value - comes from McKinsey’s 2023 “Economic potential of generative AI” analysis. The full sentence: 60-70% of the total addressable economy-wide LLM value concentrates in a small number of knowledge-work functions (customer ops, sales, software, marketing).

That is a top-down estimate of the maximum economy-wide value ceiling, based on what activities LLMs are capable of accelerating in principle. It is not a forecast of what any specific business will capture, by when, on its current adoption trajectory.

How it gets misread

The number gets quoted in three misleading ways:

“Your business can expect 60-70% productivity uplift from AI.” No. The 60-70% is the concentration of the addressable opportunity, not the realised uplift any business will see.
“McKinsey says AI will add 60% to GDP.” No. The figure is about LLM-addressable value within the AI subset, not GDP.
“Your team will be 70% more productive after rollout.” No. The number says nothing about adoption rates, change-management costs or workflow fit.

The number is real and the underlying analysis is robust. The misapplications are mostly vendor marketing.

What Australian SMBs actually see

From XLev’s engagements across the first half of 2026, realised productivity uplifts typically land:

Reactive tier (28% of SMBs): 3-8% individual uplift on the workflows people happen to use AI for. No team-level gain.
Exploring tier (33%): 5-15% individual uplift on the workflows AI fits; small team-level gain on the function that has been partially rolled out.
Operational tier (24%): 10-25% across the function with a proper rollout; measurable EBITDA-quality improvement.
Embedded tier (12%): 15-35% across multiple functions; compounding quarter-over-quarter; meaningful operating-margin uplift.
Leading tier (3%): 20-45% across the operation with sustained discipline; structural competitive advantage; multiple-point EBITDA-margin lift.

None of those numbers are 60-70%. None of them are theoretical. The gap between the McKinsey ceiling and the realised numbers is mostly governance, workflow selection and change management.

The cherry-picked peak vs the operation average

Vendors and AI consultancies often quote the peak workflow speedup as if it were the average. A 5-10x speedup on the single best-fit task (drafting a proposal, summarising a transcript) is real - but it is the peak, not the average. The average across an SMB’s actual mix of work is much smaller because:

Most of the work being done is not on the peak workflow
Adoption is uneven across the team
Some workflows AI fits poorly and produces negative net value if forced
Governance and review overhead reduces gross uplift

Honest reporting separates the peak workflow from the operation average. Marketing rarely does.

Closing the gap

The gap between McKinsey’s ceiling and the realised number closes when three things are in place:

Right workflows. Pick the workflows where AI compounds - high volume, unstructured input, recoverable error cost. Skip the workflows where it does not.
Real governance. One-page policy, named owners, training before access, breach reporting. Cheap to install; expensive to skip.
Real change management. The 90-day adoption sprint. Without it, licences are bought and the productivity gap stays open.

What to tell the board

When you brief the board on AI adoption, anchor the discussion on three numbers honestly:

Your current tier and the realised uplift consistent with that tier (calibrate against the percentages above)
The tier you are targeting in 12 months and the implied uplift
The work required to get there - which is mostly governance and change management, not tooling spend

Boards respond well to honest calibration. Boards respond poorly to the McKinsey headline numbers being quoted as forecasts. The gap between the two is where consultant credibility is built or lost.

Frequently asked questions

What is the McKinsey 60-70% figure actually saying?: McKinsey's 'Economic potential of generative AI' analysis estimates that 60-70% of the addressable economic value from LLMs sits in a small number of knowledge-work functions (customer ops, sales support, software, marketing). That is a top-down estimate of total addressable economy-wide value, not a guarantee any individual business will capture 60-70% productivity uplift across its operation. The two are routinely conflated by AI marketing.
What is the realistic AI productivity uplift for an Australian SMB?: Typical realised productivity uplifts are 10-30% on the specific workflows AI fits. On a well-rolled-out Claude implementation, that translates to a meaningful operating margin lift across the business - usually 5-15% on EBITDA terms once the workflows compound. Larger numbers (30-50%) are realistic on specific high-leverage workflows but not on every workflow simultaneously.
Why does the gap between potential and realised value exist?: Three reasons. Workflow selection - businesses pick AI workflows that do not fit AI's strengths, or pick the wrong starting point. Governance - tools rolled out without policy, oversight or training underperform. Change management - the 90-day adoption sprint that makes AI rollout stick is the work most often skipped. Close all three gaps and realised value approaches the potential. Skip them and the headline numbers stay theoretical.
Are the bigger productivity claims fake?: Not fake, but cherry-picked. Vendors and AI consultancies often quote the peak workflow uplift (a 5-10x speedup on the single best-fit task) as if it were the average. The peak is real; the average across an SMB's operations is much lower. Honest reporting separates the peak workflow from the average across operations.
How should an SMB owner think about AI ROI?: Three rules. Calibrate to your stage of adoption (Reactive tier expects 3-8% individual uplift; Embedded tier expects 10-30% across the business). Measure wage-equivalent savings on specific workflows, not handwave productivity. Plan for the gap between potential and realised value to close over 2-3 years of compounding rollout, not in a quarter.

Where this fits

Consulting

$599/hr operator advisory or monthly retainer. For a single decision, an unblocking session or ongoing AI orbit.

See Consulting →Book a free 30-min discovery call