Frequently Asked Questions

What agentic AI and RAG actually mean, what the work costs, how long it takes, and how we keep it defensible in regulated industries.

What we build

The two things clients ask us to build most often, and what the words behind them actually mean.

What is agentic AI and how does it differ from traditional AI?

Traditional AI answers one question at a time. Agentic AI plans a sequence of steps, carries them out, checks the result, and adjusts. The practical difference is that you hand it an outcome rather than a question: instead of asking a model to summarize a claim, you ask a system to work the claim, and it decides which records to pull, which rules apply, and when a human needs to look. We build these as multi-agent architectures, where each agent owns one narrow job and a supervising agent coordinates them inside defined guardrails. Every agent action is logged, which is what makes the approach usable in a regulated environment at all.

What is a RAG system and why does my business need one?

RAG, retrieval augmented generation, is how you get a language model to answer from your documents instead of from its training data. The system searches your policies, contracts, claims history or knowledge base, hands the model the passages that matter, and asks the question against those passages. Two things follow. Answers cite a source you can open and check, and when the underlying document changes, the answer changes with it, with no retraining. If your business runs on documents that have to be right, this is usually the first thing worth building.

What is the difference between RAG and fine-tuning an LLM?

RAG supplies knowledge at the moment of the question. Fine-tuning changes the model itself. Think of RAG as giving someone the right file before they answer, and fine-tuning as sending them on a training course. For most business questions RAG wins, because your documents change weekly and retraining does not keep up, because the answer can cite its source, and because your data stays where it is. Fine-tuning earns its place when the problem is style or vocabulary rather than facts, for example getting output into a specific regulatory format every time. The two combine well, and we usually start with RAG and only fine-tune once there is evidence it is needed.

What is multi-agent orchestration and when do I need it?

It is several narrow agents working one process, with a supervisor coordinating them. In underwriting that might be one agent pulling data out of the application, one checking it against risk models, one testing it against compliance rules, and a supervisor assembling the recommendation. You need it when one prompt can no longer hold the whole process, which usually shows up as an agent that works in a demo and behaves unpredictably on real cases. Splitting the work makes each piece testable, and a step that fails can be found and fixed without rebuilding the whole thing.

Working with us

How an engagement starts, what it costs, and how long things take.

What does a fractional AI CTO do?

Senior technical leadership, part time, usually two to four days a week. In practice that means owning the architecture, choosing and testing vendors instead of taking their word, setting the governance that keeps AI work auditable, mentoring the engineers you already have, and keeping the roadmap tied to outcomes the business actually wants. It suits companies that need the judgment of a senior AI leader without committing to a full time hire at the salary that role now commands.

How much does a custom RAG system cost?

Custom RAG systems typically range from $5,000 to $20,000. What moves the number is how many sources have to be connected, how clean the documents are, and what has to be proven afterwards. A single-source knowledge agent starts around $5,000. A multi-source system with compliance logging, evaluation and production deployment sits in the $15,000 to $20,000 range. Document cleanup is the line item that surprises people most often, so we scope it explicitly rather than discovering it halfway through.

How long does it take to build and deploy an AI agent?

A working proof of concept takes two to four weeks. A production agent, with compliance logging, monitoring and real integration into your systems, takes six to twelve weeks. Fractional CTO engagements run continuously with a three month minimum. We demo every two weeks against real data rather than a prepared example, because the gap between those two is where most AI projects quietly fail.

How do I get started with Goatechs?

Book a free 30 minute call. We go through the problem you are trying to solve, what your data currently looks like, and what compliance has to be satisfied. You get a written proposal with scope, timeline and price. Most engagements start within one to two weeks of signing.

Do you offer ongoing support after deployment?

Yes. Every engagement ends with documentation and a handover built so your team can run the system without us. Beyond that we offer support covering monitoring, retraining, performance work and new features. A fractional CTO engagement includes this by its nature, since we never actually leave.

How is this different from hiring an AI consultancy?

We ship code rather than recommendations. The deliverable is a running system, its tests, its documentation and a team that understands it, not a slide deck proposing one. We work inside your repositories, review pull requests, and sit in your stand-ups. That shortens the distance between a decision and a working change, and it means the knowledge stays with you when the engagement ends.

Compliance and regulated industries

Most of our work is in places where being right matters more than being fast, and where someone will eventually ask how the system reached its answer.

How do you ensure compliance in healthcare AI projects?

We design for HIPAA from the first architecture diagram rather than auditing for it at the end. That means encryption in transit and at rest, role-based access, audit logging of every AI decision with the inputs that produced it, and data minimization so the system only ever sees what it needs. We build in explainability so a clinical decision can be reviewed after the fact, and we keep the documentation a regulator will ask for as the work happens. Our architectures support Business Associate Agreement obligations.

How do you handle data privacy in insurance AI systems?

Insurance systems hold personal data, health records, financial detail and claims history, so the design question is always how little the system can be given rather than how much. We apply data minimization, encryption in transit and at rest, role-based access, anonymisation of training data, and audit logging throughout. Systems are built against state insurance regulation and NAIC model law, and against GDPR where it reaches. Retention policies and deletion requests are handled in the design, not bolted on when someone asks.

What AI regulations should fintech companies know about in 2026?

Four areas matter most. The EU AI Act treats credit scoring and fraud detection as high risk, which brings documentation and testing obligations. The SEC has been explicit about disclosure where AI informs trading. The OCC expects AI models to sit inside existing model risk management rather than beside it. And state privacy law, CCPA and CPRA in particular, governs the data underneath all of it. The common thread is that each one asks you to show your work, so we build model documentation, bias testing, explainability reporting and audit trails as part of the system rather than as a later project.

How do AI agents handle errors and edge cases in production?

By assuming they will happen. Agents score their own confidence and route anything uncertain to a person. Fallback chains mean a failed model call is retried elsewhere rather than returning nothing. Circuit breakers stop one failing dependency from taking the process down. Everything is logged with enough context to reproduce the failure, and error rates are alerted on rather than discovered. For decisions that carry regulatory weight, a human stays in the loop by design, not as a temporary safeguard.

Can AI replace my entire customer service team?

No, and we would not build it that way. The pattern that works is automating the routine contacts, password resets, policy questions, claim status, while routing anything unusual to a person who receives the full context rather than starting cold. Your team stops doing the repetitive portion and keeps the part that needs judgment. In regulated industries human oversight is not optional anyway, so a design that tries to remove it creates a compliance problem rather than a saving.

Technology and results

What we build with, who we build for, and how success gets measured.

What industries do you specialize in?

Regulated industries, where accuracy and auditability decide whether a system can be used at all: insurance, including MGAs, carriers and brokers; healthcare and life sciences; fintech and banking; and federal agencies. We also work with growing companies that would rather build on an architecture that will survive their first audit than rebuild after it. Our principals' background is enterprise delivery, including Adobe as a Senior Technical Consultant, InterContinental Hotels Group, and six years in a regulated financial services environment at U.S. Bank.

What AI models and technologies do you use?

We are model agnostic, and we test rather than assume. Current work spans the Anthropic Claude, OpenAI GPT and Google Gemini families, plus open models such as Llama and Mistral where cost or data residency points that way. For retrieval we use vector stores including pgvector, Pinecone and Weaviate, choosing pgvector when a client already runs Postgres and would rather not add a system to operate. Orchestration is LangGraph, CrewAI or purpose-built, depending on how much control the process needs.

Can you integrate AI with my existing tech stack?

Yes, and that is most of the work. We integrate with CRMs including Salesforce and HubSpot, ERP systems, data warehouses, Adobe Experience Platform and internal tools that were never designed to be integrated with. The approach is to put an API layer between your systems and the AI so that neither one has to be rebuilt to accommodate the other, and so the AI can be changed later without touching the systems of record.

How do you measure the success of an AI implementation?

We agree the measure before any code is written, and we take a baseline, because a percentage improvement is meaningless without the number it improved on. Typical measures are cycle time on a named process, accuracy against a human-reviewed sample, the proportion of cases closed without human touch, hours returned to a team each week, and customer satisfaction where the process reaches the customer. These are instrumented and put on a dashboard you can read without us, so the result is something you measure rather than something we claim.

What is the ROI of implementing AI in insurance operations?

It depends almost entirely on your starting point, which is why we insist on a baseline before quoting a number. The gains come from a short list: claims that no longer wait in a queue for manual data entry, fraud signals surfaced while a claim is still open rather than in a later review, and underwriting research that arrives assembled instead of being gathered by hand. Whether that is worth a little or a lot depends on volume, how much of the process is manual today, and how clean your data is. We scope the first engagement so the measurement comes out of it either way, and we would rather show you a real number from your own process than an industry average from someone else's.

Question not answered here?

Book a free 30 minute call. Bring the problem rather than the technology, and we will tell you whether AI is the right answer to it.