The cheapest AI contract is rarely the cheapest AI bill

One in every five dollars (21%) of enterprise localization budgets now goes to rework, inconsistency and fixing AI-generated content before it can be deployed. That figure comes from a survey of 200 enterprise leaders by RWS, written up by its CEO Benjamin Faes as a Forbes Technology Council post on 29 July 2026. Faes sells the layer his piece recommends, and the survey is RWS's own, so treat the exact percentage as directional rather than audited. The direction is the useful part.

It points at a number most buyers do not price when they compare AI vendors: the share of generated output that still needs a human rewrite before it ships. On the quote you see licence cost and, if you push, token cost. You do not see the rewrite line, because the vendor cannot see it either. It depends on your writers, your systems and the standard your audience holds you to.

Contract price is the smallest of three numbers

Three numbers decide what AI content actually costs a team.

The first is the one on the invoice: licence, seats, token spend. Buyers compare it line by line.

The second is the rewrite rate: the percentage of generated drafts an operator still has to repair before publishing. If the rate is high, a cheaper tool is more expensive, because each piece absorbs human time that scales linearly with volume. The RWS survey puts a crude price tag on this at enterprise scale, in one specific context: multi-market localization, where every language and region multiplies the fix work.

The third is the integration tax: how well the output drops into the systems you already run, versus how much stitching, reformatting and governance you have to add. The same survey reports that only 14% of enterprise leaders have centralized content management in place. The rest are scaling AI output on top of disconnected systems. Faes frames the choice bluntly: fix the content once before it ships, or keep paying the rework tax market by market.

Contract price is almost always the smallest of the three. In a procurement comparison it is also the only one on the spreadsheet.

The small-team version of the problem

The RWS figure is specific to large enterprises doing multi-market deployment. A ten-person technology team is not that. But the structure of the problem is the same, and in some ways sharper.

A small team cannot absorb a high rewrite rate. If each AI draft needs forty minutes of operator time before it ships, and the team publishes ten pieces a week, that is a role. If the next vendor cuts the rewrite rate from forty minutes to ten, the licence can cost three times as much and still be the cheaper contract. That calculation almost never appears in a vendor shortlist.

The advantage small teams have over the enterprises in the survey is the 14% number. Centralised content management is a years-long programme for a global firm. For a team of ten, it is a decision this quarter. You can impose structure before you scale the output: one source of record for positioning, one place where approved claims live, one review step every draft passes through. Do that, and a cheaper AI tool becomes viable because the surrounding system catches what the model gets wrong.

Skip it, and the opposite happens. The faster the tool, the faster the inconsistency compounds, and the rework bill moves from an occasional annoyance to the dominant cost line.

What to actually do before signing

A practical test, before the contract:

• Pick three real pieces the team would otherwise write this month. Not demo prompts; actual work with your constraints.
• Run each one through the vendor end-to-end, from brief to publishable draft.
• Have the operator who would normally own the piece bring it to shipping standard, and time the rewrite.
• Multiply the average rewrite time by expected monthly volume, and price it at a realistic internal rate.

Add that number to the licence quote. That is the contract price you are actually agreeing to for the next twelve months. Compare vendors on the sum, not on the licence line.

The test also reveals the integration tax, because the operator doing the rewrite will tell you, unprompted, where the output breaks: tone, structure, factual anchoring, the three claims it keeps hallucinating because the model has no access to your positioning. That list is the real brief for whichever vendor wins.

Where this leaves procurement

The common buying move right now is to run a bake-off, pick the lowest licence cost that clears a quality bar, and sort out operations later. The RWS piece is self-interested, but it describes what "later" tends to look like at scale: a rework line that absorbs a fifth of the budget it was meant to free up, and a governance problem that only gets harder as output grows.

For a small technology team, cost-effective AI is not the cheapest contract. It is the one whose output clears your bar often enough that the human time saved is real, inside a content system disciplined enough to catch what the model does not. The vendor matters less than that system. Choose the system first, and the vendor comparison becomes a short meeting instead of a long argument.

At Matrex Analytica we built Matrex Terminal around exactly this problem: content infrastructure that holds the positioning, claims and review discipline a thought-leadership programme needs, so the AI layer sitting inside it produces drafts that ship without a human rewriting the second half. If you are comparing vendors this month, run the test above first. The number it gives you is more useful than any brochure.

Source: Benjamin Faes, Fast To Create, Expensive To Fix: The Hidden Cost Of AI Content At Scale, Forbes Technology Council, 29 July 2026.