Generative AI FAQ
What generative AI actually is, what it is not, and what has to be true before you put it on a real workflow.
Generative AI is software that produces new text, images, code, or audio from a prompt, instead of only classifying or scoring an input. In a company it is useful only when retrieval, evaluation, permissions, and a human fallback sit around the model — otherwise you have a demo that improvises.
What generative AI actually is
If you only need one distinction: classifiers score; generative models produce. Production is everything you wrap around that production so it cannot improvise on your customers.
What is generative AI?
Generative AI is a class of models that produce new content — text, code, images, audio — from a prompt, instead of only classifying or scoring an input. The model predicts the next token (or pixel) given context. In enterprise use that generation is constrained by retrieved documents, tool results, schemas, and a refusal when evidence is missing. Without those constraints it will still generate; it just will not be trustworthy.
How is generative AI different from traditional AI?
Traditional machine learning maps an example to a label, a rank, or a number — fraud score, next-best offer, churn. Generative AI samples a continuation: an email, a SQL query, a summary. Discriminative models are usually easier to evaluate and cheaper to run. Generative models earn their place when the output is language or structure that cannot be templated, and they need evals because “looks right” is not a metric.
How does generative AI work?
A transformer (or similar) model encodes the prompt as tokens and repeatedly samples the next token from a probability distribution. At work, the prompt is assembled from a system policy, user text, retrieved passages, and tool results. The raw model has no live view of your wiki; anything current has to be fetched. Sampling is why two runs can differ, and why temperature, schemas, and evals are production controls, not research trivia.
What can generative AI actually do for a company?
Draft and classify tickets, answer from a permissioned knowledge base, extract fields from documents, write SQL against a warehouse, summarise threads, and drive tools in a scoped agent. The high-ROI first uses are high-volume, well-bounded, and checkable against a source of truth. Open-ended “strategy partner” chat is a poor first buy: you cannot define success, so you cannot tell if it failed.
Is ChatGPT the same thing as generative AI?
No. ChatGPT is a consumer and business product on top of generative models. Generative AI is the category: many model families, APIs, and open-weight checkpoints, plus the retrieval, tools, and evals you add. A company can use ChatGPT, an API, a self-hosted model, or all three. ReinforcedX is not a ChatGPT reseller; we implement systems on the models you already pay for, with no token markup.
What are the risks of using generative AI at work?
Hallucinated facts, leaked documents via a too-wide retriever, prompt injection through untrusted text, irreversible tool calls, and silent quality drops when a provider updates a model. None of those are solved by a better prompt alone. The controls are permissions at retrieval and tool time, refuse-when-empty, traces, an eval suite, and a human queue for low-confidence and high-stakes cases.
Do we need generative AI, or is regular automation enough?
If the workflow is deterministic — same fields, same clicks, same branch — a script or RPA bot is cheaper and more reliable. Generative AI is for the long tail of language, messy documents, and decisions that do not fit a flowchart. We will say so in the first conversation if the job is a script. Mixing the two is common: the model proposes, a deterministic path commits.
How a first system gets built
The build is a workflow, a source of truth, a golden set, and an agreement on what “good” means — not a model bake-off.
How long does a first generative AI system take to ship?
Four weeks is the standard implementation: discovery in week one, environments in week two, a shadow-mode pilot in week three, handover in week four. A playground demo can exist in days. Production is the eval suite, permissions, traces, and fallback, which is what the four weeks are for. Heavier regulated reviews run longer because the governance layer is heavier, not because the model is slower.
What do you need from our team?
One process owner who knows the workflow, one engineer with access to the systems being integrated, and a weekly 45-minute review. We train your engineers while we build rather than dumping documentation at the end. If nobody can say what a correct answer looks like, we stop before week two — that is a policy problem, not a model problem.
Do we have to pick one model vendor?
No. We are model-agnostic: Anthropic, OpenAI, Google, Mistral, open-source, and your fine-tunes run through one interface. Pin a version, route easy cases to a smaller model, fall back when a provider degrades. You pay the provider directly. Switching models is an eval-gated config change, not a rewrite, if traces and prompts were built that way from day one.
Can generative AI run on our own data without training a model?
Yes. Retrieval-augmented generation fetches permissioned passages at question time and asks the model to answer from those passages. Fine-tuning is for a skill, a format, or a voice that prompting cannot hold. Most knowledge problems are RAG problems. Training is the exception, not the default, and it still needs evals or you will not know if it helped.
How do you measure whether generative AI is working?
A golden set of real cases, rubric scores kept separate (retrieval, faithfulness, tool-use, tone), and a CI gate that fails the release on a drop. Online monitors sample live traffic so a silent model update shows up on a dashboard. “The demo looked good” is not a metric. If we cannot define success, we say so before taking the work.
Who owns the system after you leave?
You do. Fine-tuned weights, datasets, evaluation suites, and runbooks are yours at handover. There is no lock-in that requires us to keep it running. Thirty days of on-call cover is included; a retainer after that is optional. Your engineers have been in the weekly review the whole time, so handover is not a document dump.
How is a generative AI implementation priced?
A platform subscription plus a fixed-scope implementation fee, quoted in writing before work starts. Implementation is priced by engagement, not by the hour. Inference is paid by you to your provider; we never resell tokens with a markup. A slower week costs you nothing extra. Scope changes are written down, not absorbed into a vague retainer.
Running generative AI in production
Production is permissions, traces, refusal, and the ability to tell whether a change made things better.
Where does generative AI run — your cloud or ours?
Yours. Deployment happens inside your cloud perimeter, against your data stores and your identity provider. We do not copy the corpus onto ReinforcedX infrastructure, and it is not used to train shared models. For regulated work this is the default: VPC, on-prem, or your existing Kubernetes cluster, with provider zero-retention settings where the API path is still required.
What happens when generative AI is wrong?
Low-confidence and high-stakes cases route to a human queue by design. Every failure is captured with its trace and becomes a new eval case, so the same mistake is caught automatically next time. That loop is how the system improves after handover. We do not promise zero errors; we promise a defined behaviour when errors happen.
Can we put company data into generative AI?
Yes, if the path is an API or a self-hosted model with a written data-use policy, not a consumer chat window. Consumer ChatGPT can train on prompts depending on settings; that is the wrong product for source code and customer records. Production traffic stays in your tenancy, with PII redaction in logs, and retrieval filtered to what the asking user can already open.
Does generative AI still hallucinate in production?
It can, if retrieval is empty, the prompt allows ungrounded claims, or a tool result is ignored. Production systems refuse when evidence is missing, require citations on knowledge answers, and score faithfulness on a golden set. You reduce hallucinations; you do not delete the sampling process. Teams that skip the refuse path find out from a customer, not from an eval.
What does a four-week implementation actually cover?
Week one: workflow, success metric, sources, and risks. Week two: environments, connectors, traces. Week three: shadow mode on real traffic against the golden set. Week four: permissions, runbook, handover, and 30 days on-call. A demo is not week four. If the workflow cannot be bounded in week one, we do not pretend the calendar will invent a metric.
Who is responsible when the output is bad?
Your operators, under the runbook: the agent proposes, the queue owns exceptions, and the eval suite owns regressions. We stay on-call for 30 days after handover for defects in what we shipped. The model provider is not a party you can page about a wrong refund. That is why traces, version pins, and a human fallback are in the design, not in a slide.
When to use it — and when not to
The architecture travels. What changes is the workflow, the definition of a wrong answer, and whether a person has to confirm.
When should we not use generative AI?
When the path is fully deterministic, when you cannot define a success metric, when a wrong action is irreversible and you have no review path, or when the source of truth is still in someone’s head. Also when the job is a classifier that a smaller supervised model already does well. We will decline the work if we cannot measure success.
How does generative AI show up in customer support?
Grounded answers on high-volume intents, drafts for agents, and scoped actions in the helpdesk or order system — with escalation on identity, billing disputes, and weak retrieval. The customer-support FAQ covers the workflow; this page is the model layer. Inventing a policy that is not in the retrieved source is a production bug, not a personality.
Can regulated industries use generative AI?
Yes, inside the perimeter, with reconstructable traces (inputs, retrieved policy, tool calls, model version, output) and human checkpoints on anything that moves money, PHI, or a customer record. Banks and health systems typically need a longer governed pilot than the four-week default because review is heavier. We do not invent a certification you cannot audit.
Generative AI vs agents vs RAG — what do we actually buy?
Generative AI is the model capability. RAG is how you ground it on documents that change. An agent is how you let it take actions with tools. Most first production systems are a generative model plus RAG on one workflow; tools open after shadow mode holds. Buying “an AI” without choosing among those three is how programmes stall in committee.
What usually fails in the first 90 days?
No golden set, a single service account over the corpus, write tools on day one, and no refuse-when-empty path. Silent provider updates then land on customers because nobody sampled live traffic. The antidotes are in the RAG, evaluation, and implementation FAQs. We would rather kill a bad first workflow in week one than rescue it in month three.
Generative AI is the class of models that generate content rather than only labels. Traditional machine learning maps an example to a score; a generative model samples a continuation. Enterprise systems pair a model with your data (usually via RAG), scoped tools, an eval suite, and a human queue. ReinforcedX designs and hands over those systems inside your cloud in a four-week implementation. You own the weights, datasets, evals, and runbooks. Hallucination, permission leaks, and unmeasured quality are the failure modes that kill rollouts, not a missing brand of model.
Keep reading
Ready to refine
your workflow?
Share your current process. We’ll help you identify what can be automated — and where efficiency can be reclaimed.