Industries/ Startups/ AI inside your product
AI features in your SaaS that do one job customers do by hand.
We build search, summaries, recommendations or document reading into your own product, scoped to one task and measured by whether customers actually use it. The cost per customer is worked out before the feature goes anywhere near your pricing page.

- Used by
- Your customers, your product team and whoever sets pricing
- Takes the place of
- Copy and paste into a chatbot, manual tagging, keyword search
- Connects to
- Your product database, a model provider and usage billing
- Built with
- Our custom ai features service
Where it hurts today
A normal week, before it exists
Your customers already do this job by hand inside your product. They read long documents and type a summary into a notes field. They tag records one by one because search only matches exact words. Some have started pasting your data into a general chatbot in another tab, which tells you the need is real and also that your customers' data is going somewhere you cannot see.
The pressure to add AI comes from several directions at once. A competitor's release notes mention it, an investor asks about your AI plans, and a co-founder has a demo working in an afternoon. The demo is impressive on three hand-picked examples. Nobody has tried it on the messy records real customers create, or asked what happens when the model provider is slow on a busy afternoon.
Then there is the bill. Most AI models are charged per request, so a popular feature can cost more to run than the plan it sits in earns. Early SaaS pricing is rarely built with that in mind. A founder finds out at the end of the month, when the provider's invoice arrives and nobody can say which customers or which feature drove it.
What it does
AI inside your product, feature by feature
Ordinary product engineering, applied to a component that is probabilistic, sometimes slow and billed by the use.
- 01
One AI feature, chosen on evidence
We look at support tickets, session recordings and the steps where customers give up, and pick one repetitive job to take off their hands. It is usually small and specific, like summarising an uploaded contract or suggesting tags, rather than the ambitious idea at the top of the roadmap.
- 02
LLM cost per customer, modelled first
We work out what each request costs, how often a typical account will make one and what that means for each plan. Smaller models handle the easy requests, and usage limits sit inside each plan, so a heavy user cannot quietly erase the margin on their own subscription.
- 03
Grounded in each customer's own data
The feature only reads records belonging to the account using it, enforced in the data layer rather than in the prompt. Answers point back to the document or record they came from, so a customer can check a summary against the source in one click.
- 04
A fallback for bad afternoons
When the model provider is slow, rate-limited or down, the feature falls back to a cached result, plain keyword search or a clear message, and the rest of your product carries on. A feature that takes your product down with it is a liability, not a selling point.
- 05
Measured against doing nothing
We keep a set of real examples with known good answers and test every change against it. In production, the feature runs behind a flag for some accounts first, and we compare what those customers do with everyone else, not how clever the output looks.
- 06
The model as a replaceable part
The model sits behind an internal interface in your codebase, so switching to a cheaper or better one is a configuration change and a test run, not a rewrite. This layer changes quickly, and your product should be able to take advantage rather than be stuck.
A day with it
How the day runs instead
A normal week with one AI feature live for some of your accounts.
- MondayCustomer
A customer uploads a twenty-page supplier contract. Moments later the product shows its term, renewal date and notice period, each line linked to the clause it came from. The customer corrects the notice period, and the correction is saved against that record.
- TuesdayProduct lead
The product lead opens the feature monitor. Corrections cluster on one document type: scanned contracts with handwritten amendments. That type is added to the evaluation set, and the feature now flags scanned files for a person to check instead of guessing.
- WednesdayFounder
Cost per active account on the Starter plan has crept above the ceiling agreed at launch. The founder sees it on the costs screen and decides with the team to route short documents to a smaller model rather than raise the price.
- Thursday afternoonEngineer
The model provider has a slow hour. Requests time out, the fallback takes over, and customers see the document with a note that its summary will follow. An alert reaches the engineer, who watches it recover rather than firefighting.
- FridayFounder
With the evaluation set passing and costs back under the ceiling, the founder widens the rollout to all paid accounts using the flag. No release is needed, and the switch to turn it off stays exactly where it was.
What it connects to
Works with what you already run
An AI feature lives inside your product, so most of its connections are to things you already run.
Your product database and file storage
The feature reads only the records and documents the current account is allowed to see, through your existing permission checks, and writes results back as ordinary data your product can search, filter and export.
A model provider
Language or vision models reached through an API, for example from OpenAI, Anthropic or Google, chosen per task on quality, speed and cost, and set up under your company's own account and data terms.
Usage tracking and billing
Requests counted per account and per plan, so limits are enforced in the product, and usage can be passed to your payment provider if you decide to charge for it separately.
Monitoring and tracing
Every request logged with its cost, response time and outcome, so slow answers, failures and drifting quality show up on a dashboard and alert a person rather than waiting for a support email.
The first version
Start small, then grow it
A first version should be useful on day one and small enough to get right. This is where we would usually draw the line, and what tends to follow once people are using it.
In the first version
- One AI-assisted task, built into an existing screen of your product
- A cost model per plan, with usage limits enforced in the product
- An evaluation set of real examples with known good answers
- A fallback for when the model is slow or unavailable
- A feature flag to release it to some accounts first
Once it is in use
- A second task once the first has shown it changes behaviour
- Search across a customer's records that matches meaning as well as keywords
- A smaller model of your own, once you have enough labelled examples
- Charging for heavy AI usage as an add-on, if the numbers support it
Buyer's guide
Before you commission one
Four decisions worth making before any AI feature is built.
- 01
Buy a tool, call an API or train a model
If the need is internal, such as your own team summarising support tickets, an AI feature in a tool you already pay for is usually the better buy. Inside your product, calling a model provider's API is the normal starting point. Training or fine-tuning a model of your own only makes sense once you have plenty of labelled examples and a clear reason.
- 02
What to prepare first
A few dozen real examples of the job, anonymised where needed, with what a good result looks like for each. Your terms of service and privacy notice, and any customer contracts that mention sub-processors. And a rough idea of how often a typical account would use the feature, because that is what drives the cost model.
- 03
Who should own it
A product lead or founder should own the feature: what it does, which accounts get it and when it is switched off. Someone on the finance side should own cost per account against each plan. Whoever answers security questionnaires should own the record of providers and data flows, because enterprise customers will ask.
- 04
How you will know it is working
Watch whether customers use it more than once, whether they accept results or correct them, whether support tickets about that task change, and whether cost per account stays inside the ceiling you set. If the flagged group behaves no differently from everyone else, the feature is not earning its place yet.
Terms worth knowing
The words you will meet
- Retrieval-augmented generation
- Often shortened to RAG. Instead of relying on what the model learned in training, the product first finds the relevant records or documents, then passes them to the model with the question. Answers stay grounded in the customer's own data, and that retrieval step is where tenant isolation is enforced.
- Prompt injection
- An attack in which instructions hidden in a user's input, or in a document the model reads, try to make it ignore its rules. OWASP ranks it first in its 2025 Top 10 for LLM applications. It is why access control belongs in your code and data layer, not in the prompt.
- Evaluation set
- A fixed collection of real examples with agreed good answers, used to test the feature every time the prompt, the model or the data changes. It turns quality from opinion into something you can check, and it is what makes switching to a cheaper model a safe decision.
Handled with care
The details that matter here
AI features fail in quieter ways than ordinary ones. These are the details we watch closely.
What your terms allow
Before anything reads customer data, we check what your terms of service, customer contracts and privacy notice permit, and what the model provider's terms say about retention and training. We design within that with your data protection lead, and flag anything that needs your terms updating first.
Isolation in the data, not the prompt
Instructions in a prompt can be ignored or manipulated. So the separation between customers is enforced before anything reaches the model: the feature can only retrieve records the account is allowed to see. We test for leaks between accounts as part of every release.
Clear about what is generated
Generated content is labelled in the interface, linked to its source and always editable. In the EU, the AI Act's transparency duties have applied since 2 August 2026, including telling people when they are interacting with an AI system. We work out with your compliance lead which duties apply to your feature.
Ready for security questionnaires
Larger customers will ask where their data goes and who processes it, and some will ask for a SOC 2 report. We keep a plain record of the model providers in use, what is sent to them, how long it is kept and where it is processed, so your answers are ready before the questionnaire arrives.
Guides
Worth reading first
Questions
What buyers ask first
Should we just add a chatbot to our SaaS?
Sometimes that is the right first step. More often, customers get more from AI built into the screen they already use: a summary next to the document, suggested tags on the record, search that understands a vague description. That removes a step from their day rather than adding a new place to ask questions. We help you work out which shape fits your product and your customers.
Do we need our own data to add AI features?
Usually not at first. Anything language-shaped, such as summarising, extracting or drafting, can work with a general model and your customers' own records. Training a model of your own needs a good number of labelled examples and only makes sense once you have them. We tell you early which case you are in, so you do not pay for a training project you cannot use yet.
Will the AI provider train on our customers' data?
Check the specific terms of the provider you choose. OpenAI and Anthropic, for example, both say their APIs do not use customer data to train models by default. Retention periods differ, and some providers offer shorter retention by agreement. We set the feature up under your company's own account, filter records by account before anything is sent, and give you a plain description of the data flow for your privacy notice.
What happens when the AI gets it wrong?
Sometimes it will, so we design for it. Answers link to their source so a customer can check them. Anything with consequences is confirmed by a person rather than applied automatically. Corrections are recorded and added to the evaluation set, so the same mistake is caught next time. And where the feature is not good enough to help, we would rather switch it off than leave it annoying your customers.
Should AI features be included in our plans or sold as an add-on?
That is your pricing decision, and the cost model helps you make it. If the feature is cheap per request and most customers will use it lightly, including it with a fair-use limit is simpler. If use varies widely between accounts, an add-on or a usage allowance protects your margin. We build the metering either way, so you can change your mind once real usage is visible.
Got a job customers keep doing by hand?
Tell us what your customers repeat inside your product every day. We will tell you whether AI is the right tool for it and what it would take to run, and we reply within 48 hours.
More for startups
What else we build for startups
Pitch-ready brand identity
A mark, colours, type and a pitch-deck template sized for where you are now, with room to extend later. Investors, first hires and early customers judge credibility on thin evidence, so give them a company that looks as considered as the product.
A narrow first release
One workflow, real accounts and real billing, built on foundations that will not need replacing. Small enough to ship early and put in front of paying customers, so what you build next is decided by what they actually do.
Finishing a stalled build
An honest assessment of the half-finished product you were left with, then the parts that matter stabilised, tested and completed. Rebuilding by reflex is expensive, and often the existing code needs finishing rather than replacing. We will tell you which.
Metrics your board trusts
Signups, activation, retention, revenue and churn pulled straight from your product and billing system into one live view, so investor updates and board packs stop costing a founder a weekend of spreadsheet work.
Phone apps and store launch
When your product needs an app, we build for iPhone and Android together and handle the store side in parallel: developer accounts, testing tracks, privacy declarations and reviewer notes, so submission never holds up a launch date you have announced.
