AI Cost Benchmarks: Why Yours Are the Only Ones That Matter
Everyone wants the number other companies are spending. It is the least useful number available, and here is the one to build instead.
Erin Moore
Fractional Chief AI Officer
The benchmark worth having is your own cost per unit of work, tracked over time. Published industry averages — percent of revenue, spend per employee, average deal size — are close to unusable for planning, and this article deliberately does not quote any. Here is why, and what to build instead.
Why published benchmarks mislead
They average incompatible things. "AI spend" in one survey means model API costs; in another it includes the data platform, the engineers, and a machine learning team that predates generative AI entirely. Comparing your number to that average compares two different definitions.
The denominator moves. Percent-of-revenue benchmarks are dominated by industry margin structure. A software company and a distributor with the same absolute AI spend land in completely different percentiles for reasons that have nothing to do with AI.
Survivorship. Benchmarks are compiled from companies willing to report, which skews toward those with programmes they are pleased with. Spend that was quietly written off is not in the dataset.
They are stale on arrival. Unit pricing in this category has moved substantially and repeatedly. A figure gathered last year, published this year, describes a market that no longer exists. Stanford HAI's AI Index is the most careful public tracking of these trends and is worth reading for direction of travel — but direction of travel is not a budget.
The benchmark to build instead
One number per use case: fully loaded cost per unit of work completed.
Not per seat, not per month. Per ticket resolved, per document processed, per proposal drafted, per reconciliation completed. Fully loaded means the model or licence cost, plus the platform and integration share, plus the internal time spent reviewing and correcting output — that last component is routinely the largest and routinely uncounted.
Three properties make this the useful number. It is comparable across vendors regardless of how they price. It exposes usage-based pricing before renewal, because you can project it at triple volume. And it lets you compare against the honest alternative, which is the fully loaded cost of the same unit of work done the way you do it now.
The components are broken down in AI total cost of ownership, and the budgeting envelope around it in AI budget planning.
How to build it in a week
Pick your three largest AI line items. For each, establish the volume of work actually completed last month — not licences issued, work completed. Add up the direct cost. Then estimate review and correction time by asking the people doing it, and cost it at loaded salary rates.
Divide. You now have three numbers that are more useful than any external benchmark, because they describe your processes, your volumes and your quality bar.
Re-measure quarterly. The trend is the point: cost per unit falling as adoption matures is the signal that the programme is working. Cost per unit flat while total spend rises means you are buying volume, not leverage.
The two comparisons worth making
Against your own prior period. The only genuinely like-for-like comparison available.
Against the manual baseline. Fully loaded, honestly counted, including the rework and the queue time. Many AI initiatives look expensive against an imagined manual cost and reasonable against the real one.
What this does not answer is whether the work should be done at all. A process automated at a third of the cost is still a loss if the process was unnecessary — which is why the rejection list from an AI opportunity assessment is worth as much as the shortlist.
What to do when the board asks for an industry comparison
Give them the trend in your cost per unit, the manual baseline, and the vendor concentration picture. Then say plainly that published AI benchmarks mix incompatible definitions and go stale within a year, so you are not going to manage against one. That is a more defensible position than a borrowed number, and it survives the follow-up question — which is always how the number was derived.
Frequently asked questions
What is a good benchmark for AI spending? Your own fully loaded cost per unit of work completed, tracked quarterly. Published industry averages mix incompatible definitions of AI spend and age badly, which makes them poor planning inputs.
Why not use percent of revenue? Because the denominator is dominated by your industry's margin structure. Two companies spending the same absolute amount land in very different percentiles for reasons unrelated to AI.
What is usually missing from an internal AI cost figure? The internal time spent reviewing and correcting output. It is frequently the largest component and almost never counted.
What should the trend look like? Cost per unit of work falling as adoption matures. Flat cost per unit with rising total spend means you are buying more volume rather than more leverage.
Further reading
Tagged with:
Ready to Automate Your Business?
Let's discuss how AI automation can deliver measurable ROI for your organization in 90 days or sooner.