# AI systems development for startups, tested on your data

URL: https://pixelactstudio.com/services/ai-systems

> AI systems development for startups: RAG search, document processing, and support tools with human review, proven in a fixed-price pilot on accuracy and cost.

AI systems development for startups is building software where a language model does one defined job, such as answering questions from company knowledge, processing documents, classifying requests, drafting support replies for human review, or automating a workflow. Pixelact starts every AI system with a fixed-price pilot that measures accuracy and cost per task on real data.

- Price: Scoped on a call
- Model: Fixed-price pilot, then priced by milestone
- Timeline: Pilot first, then build milestones set in the proposal

## Who it is for

- Our team answers the same questions every day from docs, tickets, and Slack threads that nobody can search well.
- Someone on our ops team retypes fields from PDFs into our system every week, and the errors only show up at month end.
- Support volume is growing faster than the team, and we want AI drafts without a bot answering customers unchecked.
- Every request lands in one inbox, and a person spends hours each week sorting and routing it by hand.
- We shipped an AI feature quickly to prove the idea, and now we need to know how often it is wrong and what it costs at real volume.

## What you get

- **Pilot readout and eval set:** Real examples with accepted answers, kept in your repository, plus the measured accuracy, failure modes, and cost per task from the pilot.
- **The AI system in production:** The search, document workflow, classifier, support tool, or automation agreed in the proposal, running in cloud and model provider accounts in your name.
- **Human review queue:** A screen where people approve, edit, or reject the outputs that need a check, with each decision logged and corrections added as new eval cases.
- **Written data boundary:** What the model may read, what it never receives, which provider processes the data, and how user permissions carry into retrieval.
- **Monitoring and cost tracking:** A trace for each request, alerts on error rates and spend, and a cost-per-task report so the unit economics stay visible after launch.
- **Handover pack:** The repository, a runbook, the eval commands, a process for prompt and model changes, and a walkthrough with whoever takes over, so your engineers can change the system safely.

## How it works

1. **Discovery call.** A 45-minute call about the job you want the system to do, how it is done today, and what a wrong output costs. No slides, and no solution before we understand the problem.
2. **Pilot proposal walkthrough.** Within 48 hours of the call, we walk you through the pilot proposal live. It sets the question, pass criteria, the data we need, exclusions, payment plans, and the AI limits for your case.
3. **Pilot sprint.** Once the deposit arrives, we build the riskiest piece, run it on real examples, and measure accuracy, failure modes, cost, and latency. The readout recommends build, change direction, or stop.
4. **Build in milestones.** We add integrations, the review queue, permissions, and monitoring in working slices on a preview environment, and rerun the evals before each milestone ships.
5. **Launch with review on.** The system goes live with human review on every output that needs it. Automatic handling widens only where the review data supports it.
6. **Handover.** Your team gets the eval set, the runbook, and a walkthrough. The accounts are yours from the start, code and final files transfer once the final invoice is paid, and a retainer is offered once the system has proved its value.

## Not included

- Training a foundation model from scratch. We build on hosted or open-weight models, and fine-tune only when evals show retrieval and prompts cannot reach the target.
- Autonomous agents that move money, delete data, or contact customers without a person approving the action.
- A promise that the system is never wrong. The proposal states measured accuracy and known failure modes instead.
- Model, embedding, and vector database usage, which providers bill to accounts in your name with no markup from us.
- Legal or regulatory sign-off on your AI use. We document data flows and controls for your counsel or advisor to review.

## What does AI systems development for startups cover?

AI systems development for startups covers five kinds of work, each a job your team does by hand today. The model handles the language part, and plain code handles the rest: permissions, validation, retries, and logging.

- **AI search over company knowledge:** Answers drawn from your docs, tickets, and contracts, each citing its source. This is retrieval-augmented generation (RAG), and our RAG development services spend most of the effort on retrieval: chunking, hybrid search, permissions, and ranking.
- **AI document processing:** Fields pulled from invoices, contracts, or forms into structured records, checked against your business rules, with uncertain fields sent to a person.
- **Classification and routing:** Emails, tickets, or requests labeled by type, urgency, or owner and sent to the right queue.
- **AI customer support with human review:** Draft replies built from your help content and the customer's account, approved or edited by an agent before sending.
- **Workflow automation:** Multi-step processes run as background jobs with retries and logs. Workflow automation development is mostly plain code, with a model called only where a rule cannot do the job.

Our guide to [AI search over company knowledge](https://pixelactstudio.com/blog/ai-search-over-company-knowledge) covers the first kind in depth. Work that is mostly screens with a small AI step fits [internal tools development](https://pixelactstudio.com/services/internal-tools).

## How does an AI project run from pilot to production?

Every AI system starts with a fixed-price [pilot sprint](https://pixelactstudio.com/services/pilot-sprints), because nobody knows how a model performs on your data until it is measured there. The pilot ends with a recommendation: build, change direction, or stop.

**What the pilot measures**

| Measure | Why it matters for the build |
| --- | --- |
| Accuracy on the eval set | Decides whether the system clears the pass bar agreed before the pilot |
| Failure modes | Shapes the validation rules, review routing, and fallbacks |
| Cost per task | Shows whether the unit economics work at your volume |
| Latency | Decides between a live response and a background job |
| Review rate | Shows how much staff time the system saves |

If the pilot meets its pass criteria, the build proposal uses these numbers to set milestones and the review a launch needs. If it misses, you have spent the pilot fee, not the build budget.

## How do you test RAG and other AI systems before launch?

We test AI systems with evals: a fixed set of real inputs with the answers you would accept, scored automatically on every change, next to the tests our [engineering standards](https://pixelactstudio.com/about/engineering-standards) already require. The set includes awkward cases, such as scanned PDFs, and grows with corrections from the review queue.

### Testing retrieval on its own

In a RAG system, a fluent answer built on the wrong document is still wrong. So we score retrieval first, then faithfulness: whether each claim in the answer is supported by the passages it cites.

```typescript
type EvalCase = { question: string; expectedSourceIds: string[] };
type Retrieve = (question: string) => Promise<string[]>;

export async function hitRateAtK(cases: EvalCase[], retrieve: Retrieve, k: number): Promise<number> {
  if (cases.length === 0) throw new Error("The eval set is empty");
  let hits = 0;
  for (const evalCase of cases) {
    const topIds = (await retrieve(evalCase.question)).slice(0, k);
    if (evalCase.expectedSourceIds.some((id) => topIds.includes(id))) hits += 1;
  }
  return hits / cases.length;
}
```

## How does human review work in AI customer support and document processing?

Human review means a person approves, edits, or rejects an output before it takes effect. We route by the cost of a mistake, not a confidence score alone, because a model can be confidently wrong.

**Default review routing at launch**

| Output | Default route |
| --- | --- |
| Support reply drafts | An agent approves or edits each one before it is sent |
| Extracted document fields | Fields that pass schema and business-rule checks go through, and the rest go to review one field at a time |
| Request classification | Routed automatically, with uncertain items held and a weekly sample reviewed |
| Actions that move money, delete data, or contact customers | Always wait for a person |

The review screen shows the evidence next to each output, such as the cited passages or the source document, so a reviewer checks the output instead of redoing the task.

## What data can the AI model see?

An AI model sees only the data its task needs, and the proposal writes that boundary down.

- **Permissions carry into retrieval.** Search filters by the user's tenant and role inside the retrieval query, so restricted text never reaches the model. Postgres row-level security on a pgvector table adds a second check at the tenant boundary.
- **Sensitive fields stay out.** Code removes or masks data the task does not need, such as payment details or dates of birth, before the model call.
- **Provider terms are checked.** The proposal names each model provider, where it processes data, and the retention and training settings for your account.
- **Untrusted content stays untrusted.** Documents and emails can hide instructions aimed at the model, a risk the 2025 [OWASP Top 10 for LLM applications](https://genai.owasp.org/llmrisk/llm01-prompt-injection/) lists first as prompt injection. Tools get the narrowest permissions the job allows, and the server, not the model, authorizes every action.

## What AI limits go into every proposal?

Every AI proposal has a written AI limits section, agreed with the rest of the scope.

- Data boundaries: what the system reads, never sees, and where data is processed.
- Measured accuracy and the failure modes found in the pilot.
- Which outputs need human review at launch.
- What happens when the model is wrong, slow, or unavailable.
- Expected cost per task at your forecast volume.
- Which changes, such as a model upgrade, require the evals to pass again.

> **No AI system is right every time.** Instead of a promise, you get how often the system was wrong on your data, what those failures looked like, and what catches them before they reach a customer.

## How is AI systems development priced?

AI work is priced in two steps: a fixed-price pilot sprint from $4,500 and then a build priced by milestone from what the pilot measured. Larger systems are split into phases, each with its own milestones.

- **Running costs are separate.** The proposal lists each expected provider with its estimated monthly cost, and you pay providers directly.
- **Live AI features can be reviewed first.** A [software audit](https://pixelactstudio.com/services/software-audits) from $2,500 checks prompts, data flow, cost, and failure handling. If it leads to a build within 60 days, 50% of the audit fee is credited against the build.
- **After launch,** [engineering retainers](https://pixelactstudio.com/services/engineering-retainers) buy fixed monthly capacity to keep evals current as models, prices, and your data change.

## How should you prepare for an AI project?

Bring one job and evidence of how it is done today. The [AI feature readiness checklist](https://pixelactstudio.com/resources/ai-feature-readiness-checklist) goes further.

- The job in one sentence, and what a wrong output costs.
- Real examples with the answers you would accept, ideally 50 or more.
- Sandbox access to the systems involved, such as a helpdesk or CRM.
- Any data that must never reach a model provider.
- The person who reviews pilot outputs and signs off the result.

## How does a custom AI system compare with off-the-shelf AI tools?

Use the AI feature in a tool you already pay for when it does the job. Build when the job spans your own data and systems.

**Ways to add AI to your operations**

| Option | Good fit when | Trade-off |
| --- | --- | --- |
| AI features in your helpdesk or workspace tools | The job lives inside one tool | Little control over sources, review, and evals |
| Low-code automation with an AI step, such as Zapier or n8n | A simple, low-volume flow where mistakes are cheap | Hard to test and monitor as flows multiply |
| Hiring an AI or ML engineer | AI is core to the product for years | Slow to hire, and knowledge sits with one person |
| Custom build with a studio | The job spans your data and systems | Higher upfront cost, and someone must own evals after handover |

## Frequently asked questions

### How long does it take to build an AI system?

Building an AI system starts with a short pilot whose end date is fixed in the proposal. The build timeline follows from what the pilot measured: how many sources and integrations are involved, how much review tooling is needed, and how close accuracy already is to the target. The build proposal then sets dates milestone by milestone, based on those measurements rather than a guess.

### Which AI models do you use?

We pick the model for each task from eval results and cost per task, not brand preference. That can be a hosted model from a provider such as OpenAI, Anthropic, or Google, or an open-weight model you host when data cannot leave your infrastructure. Model calls go through one small interface in the code, so switching providers means rerunning the evals, not rewriting the system.

### Will our data be used to train someone else's model?

That depends on the provider, the plan, and its settings, so we check before any of your data is sent. The pilot proposal names each provider, its data use and retention terms, and the settings we will use. The accounts are in your name, so you can see and change those settings. If no hosted option fits your data boundary, we use an open-weight model in your own cloud.

### What happens when the AI gets something wrong?

Every AI system we build expects wrong outputs. Validation rules catch outputs that break known constraints, uncertain or high-risk outputs go to a person, and every request is logged so a bad answer can be traced to its input, sources, and model version. Corrections are added to the eval set, so the same failure is tested on every later change.

### Can you add AI to our existing product?

Yes. Many AI systems sit inside a product or internal tool that already exists, reading from its database and APIs. If we have not seen the codebase, a software audit first shows where an AI feature can attach safely, with a build path and rough costs. The pilot then runs in your repository, so the code that works can carry into the build.

### How much does an AI system cost to run each month?

The monthly running cost of an AI system depends on volume, the model, and how much text each task sends and receives. The pilot measures cost per task on real examples, and the proposal multiplies it by your forecast volume. Providers bill you directly, with no markup from us. We set spending alerts at launch, and prompt caching or a smaller model for simple steps can lower the cost further.

### Do you build AI agents?

Pixelact builds AI agents for bounded jobs. Here an agent is a workflow where the model chooses among a few tools, such as searching a knowledge base or drafting a ticket update, and each tool can do only what the job needs. A refund, a deletion, or a message to a customer waits for a person to approve it. We do not build open-ended agents with broad access to production systems.

### Should we use RAG or fine-tuning?

For questions over company knowledge that changes, retrieval comes first: sources update without retraining, and answers can cite where they came from. Fine-tuning can help with a fixed output format, a house style, or a narrow classification task at high volume. We decide from the pilot's eval results, and many systems never need fine-tuning at all.