Adding AI features to an existing SaaS product: the patterns that work, what they cost to run, and how to test them.

The AI features that stick are the ones that remove a step users already take. Look at where people copy text between screens, read a document to fill in fields, sort a queue by hand, or write the same kind of message every day. Your support tickets and product analytics usually point straight at them. A feature that starts from one of those jobs has a measurable before and after; a feature that starts from wanting to have AI rarely does.

Write the job down in one sentence, with the user, the input, the output and how you would know it worked. That sentence becomes the test, the cost estimate and the release note.

Nearly every AI feature in a SaaS product is one of four patterns, and they differ most in how predictable their cost and their failures are. Start with the cheapest pattern that does the job.

Model providers charge per token, input and output separately. At Anthropic’s September 2026 list prices Haiku 4.5 is $1 per million input tokens and $5 output, and Sonnet 5 is $2 and $10; batch processing halves the price for work that can wait, and prompt caching cuts the cost of repeated context. For classification and extraction the small model is usually enough, which is why that pattern costs least.

The number that matters is cost per active customer per month, set against what that customer pays. Attribute every model call to a customer and a feature, set a monthly ceiling per customer and trigger it once on purpose before release, and decide how the cost is recovered: an allowance included in each plan, or usage metered into billing through Stripe. In our own products, moving one high-volume feature from the frontier model to the small one made it several times cheaper and faster with no drop in success.

An AI feature needs a test set before it needs a launch date. Collect real examples from your own product, with personal data handled under your existing rules, write down what a correct output is for each, and agree a threshold the feature must pass. Run it, read the failures, and change the prompt, the model or the pattern until it passes. Then keep the set, because every later change of model or prompt has to pass it again.

In most SaaS products you process your customers’ data on their behalf, so adding a model provider means adding a sub-processor. Update your sub-processor list and notify customers as your contract requires, use the provider’s business or API terms under which inputs are not used for training, and confirm where the data is processed. Tell end users plainly where AI is used, and keep a DPIA where the feature affects individuals in a significant way, such as screening or scoring people.

For search over customer data, filter by customer before retrieval, not after, so a question from one customer can never surface another’s documents. Keep prompts and outputs only as long as you need them for support and evaluation, and say so in your retention policy.

Release an AI feature the way you would release any risky change, only more slowly. Switch it on for one or two customers first, with their agreement, and watch both the quality and the cost before widening it. Give users a quick way to say an answer was wrong, log enough to understand why, and keep a switch that turns the feature off for one customer or for everyone without a deployment.

We run AI features in both of our own products. In GoRizzume, a careers platform for UK students and graduates, AI scores a CV against a specific job advert, which is the classification pattern. In Onrolo, an applicant tracking system for UK recruitment agencies, AI helps agencies screen the applications they receive, with recruiters making the decisions. Both follow the rules above: every call attributed to a customer and a feature, ceilings tested, the small model by default, and new features switched on per organisation first.

When we add AI to someone else’s product, the work is priced as a fixed-price phase: £8,000 to £25,000 for a custom AI system, or as part of a SaaS phase at £8,000 to £30,000, with the evaluation set, the cost attribution and the ceilings included rather than sold as extras.