Skip to content

Service · AI

AI Development

Commission App Clone Labs to take an artificial intelligence (AI) idea from a bounded workflow to a released feature: we pick the lightest technique that solves the task, prove it on examples drawn from your own business, and build the product, review screens and controls around it. You keep the code, the test set and the operating notes.

Smallest technique that passes your test set

Acceptance cases agreed before any build

Code, prompts and evaluation data handed over

Deliverables

What an AI development engagement delivers.

01Delivered

A scoped AI capability inside a working product

The feature itself, whether a drafting assistant, a document extractor, a triage step or a search experience, delivered as part of screens and APIs people actually use, not as a notebook demo that someone else must productionise.

02Delivered

An evaluation set built from your cases

A few hundred representative inputs, including messy and borderline ones, each paired with an expected outcome or a grading rubric your reviewers signed off. It becomes the yardstick for every later prompt, model or data change.

03Delivered

Guardrails matched to the risk

Input checks, output validation against a schema, confidence thresholds that route doubtful results to a person, and refusal behaviour for requests the feature should not handle at all.

04Delivered

Review and correction screens

Interfaces where staff can approve, edit or reject what the system produced, with those corrections stored so they can feed the next round of evaluation rather than vanishing into an inbox.

05Delivered

Cost, latency and quality telemetry

Per-request logging of tokens, response time, model version and outcome, aggregated on a dashboard so finance and product can see what a useful result costs and whether quality is holding.

Early decisions

Technical decisions we settle early in AI development.

01Decided

Rules, classic ML or a language model

Plenty of tasks sold as AI are solved by a lookup table, a regular expression or a small trained classifier. We test the cheapest credible approach first and only reach for a large model when the cheaper ones fail on your examples.

02Decided

Hosted API or a model you run

Managed APIs are quick to adopt; open-weight models on your own infrastructure give control over data residency and unit cost at volume. The choice depends on your data terms, traffic profile and in-house operations capacity.

03Decided

What counts as a correct answer

Exact match, a rubric scored by a reviewer, a model-graded score checked against human samples, or a downstream business signal such as a ticket not being reopened. We fix the measure before tuning anything, otherwise improvement cannot be shown.

04Decided

Where a person stays in the loop

Some outputs can go straight to customers; others need sign-off every time; many sit in between with sampling. We set that boundary per output type, together with the people who carry the risk.

05Decided

Data allowed into prompts

Personal data, contract terms and regulated records each carry processor and retention conditions. We list which fields may be sent to which provider, and redact or keep the rest on your side.

06Decided

How the feature degrades

Provider outages, rate limits and slow responses will happen. We decide in advance whether the product falls back to a simpler model, a queued job or the manual path it replaced.

Engagement

How an AI development engagement runs.

  1. 01Phase 1

    Frame the task and collect cases

    We sit with the people doing the job now, write down inputs, decisions and failure costs, and gather real examples under your data rules to form the first evaluation set.

  2. 02Phase 2

    Prototype against the test set

    Two or three candidate approaches are scored on the same examples, so the choice between them rests on numbers and cost per result rather than a persuasive demo.

  3. 03Phase 3

    Build the product around the winner

    The chosen approach is wrapped in production code: APIs, interfaces, review flows, logging, fallbacks and access control, released behind a flag to a limited group of users.

  4. 04Phase 4

    Widen the rollout and hand over

    With live results tracked against the acceptance threshold, usage expands in steps; your engineers then receive repositories, the evaluation harness and a runbook for changing prompts or models safely.

Fit

When a commissioned AI build is the right choice.

01Good fit

The workflow is narrow enough to test

You can describe the input, the expected output and what a bad result costs, and you can supply real examples with permission to use them.

02Good fit

Probabilistic output is tolerable

The task survives an occasional wrong answer because a reviewer, an undo or a fallback catches it, rather than a single error causing irreversible harm.

03Good fit

You want one accountable team

You prefer a single delivery team covering model work, application code and release, instead of stitching together a data scientist, a backend agency and a front-end freelancer.

Start with the smallest thing that could work

Teams often arrive asking for a large language model when the task is classifying emails into seven buckets. A fine-tuned small classifier or even a set of keyword rules might clear the bar at a fraction of the running cost and with far more predictable behaviour. We run the cheap options first, on your examples, and keep the result on record.

When a language model is genuinely needed, the same habit continues: one well-structured call with a schema beats a chain of five, and a mid-sized model that passes the test set beats the flagship one that passes it by a hair while costing several times more.

Your examples are the specification

A written requirement such as “summaries should be accurate and concise” cannot be tested. A set of real documents with the summary your best reviewer would accept can. We treat the evaluation set as the specification: it is versioned, it grows whenever a new failure is found in production, and no change ships unless the numbers hold.

Building that set is also where hidden disagreements surface. Two reviewers often grade the same output differently, and resolving that early saves weeks of tuning towards a target nobody actually agreed.

The model is a small part of the product

In a finished AI feature, the model call is a few lines. Around it sit authentication, data access rules, queueing, retries, the interface a person uses to check the result, audit logs and the fallback for when the provider is down. Most of the engineering effort and most of the risk live in that surrounding code, which is why we build it with the same care as any other production system.

Designing for wrong answers

Every AI feature will sometimes be wrong. The design question is what happens next. Confidence scores can route uncertain cases to a person; schema validation can reject malformed output before it reaches a database; undo and edit controls let users repair mistakes cheaply. We decide these behaviours with the people who would otherwise answer the complaints.

Keeping it working after launch

Providers update models, your data shifts and users find inputs nobody imagined. The telemetry and evaluation harness we hand over let your team spot a decline and test a fix before it reaches customers. If you later run several models in production and need formal release gates, that operational layer is what our MLOps work covers.

FAQ

Questions to resolve before the build.

01What drives the cost of an AI development project?

Mostly the state of your data and the strictness of the acceptance bar. Clean, permissioned examples and a forgiving review step keep effort down; scattered data, regulated content, several integrations and a demand for near-perfect output push it up. Model usage fees are a separate running cost we estimate from your expected traffic.

02How long does it take before something is live?

It depends on how quickly representative examples can be gathered and approved, how many systems the feature touches, and how much review your risk owners require before widening access. We give a phase-by-phase plan once the first evaluation set exists, because that is when the hard part becomes visible.

03What exactly do we receive at the end?

Application and service source code, prompts and configuration under version control, the evaluation set with its grading scripts, infrastructure definitions, dashboards, and a runbook describing how to swap a model or edit a prompt without breaking the acceptance threshold.

04Who owns the code, prompts and evaluation data?

You do, once the agreed fees are paid. Our contracts assign the deliverables and intellectual property to you. Third-party models remain under their providers’ licences, and we record which licence applies to each component.

05How do we decide the feature is accepted?

Before building, we agree a numeric threshold on the evaluation set and a list of failure types that must never occur. Acceptance means passing that threshold, surviving a limited live release, and your reviewers signing off a sample of real outputs.

06What is outside the scope of an AI development engagement?

Training a foundation model from scratch, labelling large datasets by hand at volume, and making legal determinations about your use of AI are not included. We flag where specialist labelling vendors or counsel are needed and can work alongside them.

Primary sources

References behind this page

Dated official documentation, standards, and research that support the factual claims on this page.

  1. 01
    NIST AI Risk Management Framework

    Guidance on mapping, measuring and managing risks in AI systems.

  2. 02
    OWASP Top 10 for Large Language Model Applications

    Common security weaknesses in applications built on language models.

How to read this page

Scope, cost and timeline figures are planning guidance until confirmed in a written proposal. Diagrams and screens are illustrations, not client deliverables, and source-code rights and handover follow the signed agreement. See the delivery process and the full disclaimer.

Next decision

Turn the brief into an accepted product scope.

Define outcomes, constraints, evidence, rights and handover before delivery begins.

Commercial rights, repositories, environments, documentation, acceptance and handover remain contract-defined.