How to compare AI content operations pricing models before you sign

What to take away

Which pricing model fits your content volume?

Start with volume, not features. A team publishing 20 articles a month behaves differently from one producing 400 localised pages. Peak month matters more than average, because caps and overage rates bite then.

Seat pricing charges per named user, so costs stay flat as output rises. Usage pricing charges per token, generation or API call: quiet months cost less, busy months more.

Credit pricing bundles units into packs, hiding the unit rate until credits are divided by real output. Hybrid models add a platform fee, so part of the bill stays predictable. The costs and budget guide for England maps the surrounding line items.

How do usage and credit models behave at scale?

Usage models are transparent until volume spikes. If a campaign doubles your output, the invoice can double in the same month. Ask for a written CAP and a notification threshold, not a verbal promise.

Credit models look cheaper per unit because packs are sold in bulk, but unused credits expire, turning a discount into waste. A team buying 10,000 credits at £0.04 each pays £400, and loses £120 if a quarter of the pack expires.

Those are illustrative figures, not quoted rates. Token pricing adds a second variable, because long documents and repeated context cost more than short prompts.

What do seat and managed models cost in practice?

Seat models are the easiest to forecast. For example, a ten-person team on £30 per seat per month pays £360 a month, an illustrative figure. The risk is shelfware: seats bought for occasional users. Check the minimum term too, because a twelve-month commitment changes the effective monthly rate.

Managed models price a service, often a monthly retainer plus a per-item rate. They cost more per article but include briefing, editing and quality control. Output-based pricing is similar but ties payment to accepted items, which puts more pressure on your acceptance criteria.

Which due diligence questions should you ask?

Compliance sits alongside price. If a vendor processes personal data, its risk assessment should be checked against the ICO guidance on AI and data protection. Ask how training data is sourced and whether the contract includes an indemnity.

Content that screens or ranks people falls under the Equality Act 2010 in England, Wales and Scotland, so bias testing belongs in the contract. Public-sector buyers may also align roadmaps with the National AI Strategy action plan.

How should you model the return?

Build a simple model: subscription plus credits plus internal review hours, divided by accepted outputs. Then compare that against the value of the work. The return on investment method shows how to turn those inputs into a defensible figure before renewal.

If the cost per accepted item rises as volume grows, the model is wrong for you, however attractive the headline rate looks. Re-run the numbers at your peak month, not just the average.

Pricing model checklist

Common questions

Is seat pricing always cheaper than usage pricing?

No. Seat pricing wins when usage is high and predictable. Usage pricing wins when output is low or seasonal. Compare both at your peak month.

What is a realistic credit expiry period?

Vendors set their own terms, so treat expiry as a commercial risk. Ask for 12 months or a refund, then model the waste.

Should I bundle tools and services in one contract?

Bundling can simplify procurement but weakens your negotiating position. Split tool and service pricing so you can renegotiate each part separately.

How often should pricing be reviewed?

Review at renewal and after any volume change above 20 per cent. A quarterly check of cost per accepted item catches drift earlier.