The choices behind every AI decision, side by side.
Concepts, not vendors. Each comparison starts with the one-sentence answer, then the table, then when to choose which.
This page compares fine-tuning with retrieval and prompting, small fine-tuned models with frontier APIs, on-premise with private cloud and managed deployment, voice agents with IVR and outsourced call centres, agentic automation with RPA, and AI-native services with consultancies and software.
Fine-tuning vs retrieval (RAG) vs prompting
In short: Prompting changes what a model is asked, retrieval changes what it can see, and fine-tuning changes what it is. Most production systems use two of the three, and the mistake is using the expensive one for a problem the cheap one solves.
| Prompting | Retrieval (RAG) | Fine-tuning | |
|---|---|---|---|
| What changes | PromptingThe instructions and examples in the request | Retrieval (RAG)The documents the model reads before answering | Fine-tuningThe model's weights |
| Best for | PromptingBehaviour, tone, format, simple rules | Retrieval (RAG)Facts that change, private documents, citing sources | Fine-tuningStyle, format compliance, narrow tasks, speed and cost |
| Data needed | PromptingA handful of examples | Retrieval (RAG)The documents, cleaned and chunked | Fine-tuningHundreds to thousands of examples, or verifiers |
| Cost to change | PromptingMinutes | Retrieval (RAG)Hours: re-index | Fine-tuningDays: retrain and re-evaluate |
| Typical failure | PromptingLong prompts drift and cost tokens every call | Retrieval (RAG)Wrong chunk retrieved, answer confidently wrong | Fine-tuningOver-fit to the examples, forgets the general case |
Choose Prompting
when the task is simple, the volume is low and the facts do not change.
Choose Retrieval
when the answer lives in documents that change, or must be cited.
Choose Fine-tuning
when one narrow task runs at volume and latency, cost or format compliance matter.
A small fine-tuned model vs a frontier model over an API
In short: A frontier model over an API has the highest ceiling and the fastest start; a small model fine-tuned for one job wins on latency, cost per call, data residency and stability. For a bounded job at volume, the small model is usually the production answer and the frontier model is the exception path.
| Frontier model over an API | Small fine-tuned model on your hardware | |
|---|---|---|
| Latency | Frontier model over an APIDepends on the provider and the queue | Small fine-tuned model on your hardwareSet by your hardware, predictable |
| Cost per call | Frontier model over an APIPer token, for ever | Small fine-tuned model on your hardwareFixed hardware, near-zero marginal |
| Data residency | Frontier model over an APIThe provider's region | Small fine-tuned model on your hardwareYour building, your rules |
| Drift | Frontier model over an APIProvider updates change behaviour | Small fine-tuned model on your hardwareNothing changes unless you change it |
| Ceiling | Frontier model over an APIHighest available | Small fine-tuned model on your hardwareHigh on the job, low elsewhere |
| Time to first result | Frontier model over an APIHours | Small fine-tuned model on your hardwareWeeks, including the data work |
Choose Frontier API
when the task needs open-ended reasoning, the volume is low, or you are still proving the job is worth doing.
Choose Small model
when one job runs thousands of times a day, or the data cannot leave.
Choose Both, with a router
when the small model handles the common cases and escalates the uncertain ones.
On-premise vs private cloud vs managed deployment
In short: On-premise puts the model inside the client's own building, private cloud puts it in a tenancy the client owns in a region it chooses, and managed deployment puts it in Zingaro AI's environment. All three can be sovereign in practice; the difference is who holds the keys and how updates arrive.
| On-premise (including air-gapped) | Private cloud | Managed by Zingaro AI | |
|---|---|---|---|
| Where inference runs | On-premise (including air-gapped)Your data centre, your hardware | Private cloudYour cloud account, your region | Managed by Zingaro AIZingaro AI's environment |
| Who holds the keys | On-premise (including air-gapped)You | Private cloudYou | Managed by Zingaro AIZingaro AI, under your contract |
| How updates arrive | On-premise (including air-gapped)Signed bundles, carried in and verified | Private cloudSigned images pushed to your account | Managed by Zingaro AIContinuous |
| Outbound connection | On-premise (including air-gapped)Optional; none when air-gapped | Private cloudYes, under your controls | Managed by Zingaro AIYes |
| Fits when | On-premise (including air-gapped)Regulation names the building, or the data cannot cross a boundary | Private cloudResidency rules are about the country, not the building | Managed by Zingaro AISpeed matters more than sovereignty |
Choose On-premise
when a regulator, a contract or a threat model says the data cannot leave the network.
Choose Private cloud
when the rule is about the country or the account, and your cloud team can operate it.
Choose Managed
when you want the outcome without running anything, and the data is allowed to leave.
Voice agent vs IVR vs an outsourced call centre
In short: An IVR routes callers through a menu, an outsourced call centre puts people on every call, and a voice agent holds a real conversation for the bounded tasks and hands the rest to a person. The right answer for most operations is a voice agent with people on the exceptions.
| IVR menu | Outsourced call centre | Voice agent with hand-off | |
|---|---|---|---|
| What it handles | IVR menuRouting and simple self-service by keypad | Outsourced call centreEverything, at the cost of a person per call | Voice agent with hand-offBounded tasks end to end; hands off the rest |
| Conversation | IVR menuNone | Outsourced call centreNatural | Voice agent with hand-offNatural for the tasks in scope, in dialect |
| Scales with | IVR menuNothing; fixed | Outsourced call centreHeadcount | Voice agent with hand-offVolume |
| Exceptions | IVR menuDropped or transferred blind | Outsourced call centreHandled, quality varies | Voice agent with hand-offTransferred warm, with a summary |
| Record | IVR menuKeypresses | Outsourced call centreNotes, if written | Voice agent with hand-offTranscript, extracted fields, outcome in your system |
Choose IVR
when you only need routing and the call volume is small.
Choose Outsourced call centre
when every call needs judgement and the volume is low enough to staff.
Choose Voice agent
when a large share of calls follow a flow and the rest must reach a person quickly.
Agentic automation vs RPA
In short: RPA replays a fixed sequence of clicks and breaks when the screen changes; agentic automation gives a model a goal, tools and permissions and lets it handle variation, with a person on the exceptions. RPA is right for a stable, high-volume sequence; agents are right for work with judgement in it.
| RPA (robotic process automation) | Agentic automation | |
|---|---|---|
| How it works | RPA (robotic process automation)A scripted sequence of UI actions | Agentic automationA model plans, calls tools and checks its own work |
| Handles variation | RPA (robotic process automation)Poorly; every new case is a new script | Agentic automationYes, within the permissions it is given |
| When the software changes | RPA (robotic process automation)Breaks | Agentic automationUsually adapts; screen agents re-read the page |
| Judgement | RPA (robotic process automation)None | Agentic automationBounded, with confidence thresholds and a review queue |
| Audit | RPA (robotic process automation)Logs of actions | Agentic automationLogs of actions, reasoning and who reviewed what |
Choose RPA
when the process is identical every time and the systems never change.
Choose Agentic automation
when inputs vary, exceptions are common, or a person makes a small decision on every item today.
AI-native services vs a consultancy vs buying software
In short: A consultancy advises and leaves, software gives you a tool and a training session, and an AI-native services company does the work and delivers the outcome into your systems. The difference is who is accountable for the number on Friday.
| Consultancy | Software (SaaS) | AI-native services | |
|---|---|---|---|
| What you get | ConsultancyA recommendation and a plan | Software (SaaS)A login | AI-native servicesThe work done |
| Who runs it | ConsultancyYour team, after they leave | Software (SaaS)Your team | AI-native servicesAgents on a stack the provider owns, with the provider's people on the queue |
| Pricing | ConsultancyTime | Software (SaaS)Seats | AI-native servicesOutcome or volume |
| When it breaks | ConsultancyA new engagement | Software (SaaS)A support ticket | AI-native servicesFixed by the provider that week |
| Data | ConsultancyShared for the engagement | Software (SaaS)In the vendor's cloud | AI-native servicesWhere you say it stays, including your own servers |
Choose Consultancy
when you need a strategy and have a team to execute it.
Choose Software
when the job is generic and your team will operate the tool.
Choose AI-native services
when you want the outcome and would rather not build or run the system yourself.
Not sure which column you are in?
Bring the job. Twenty minutes is enough to say which of these fits, and whether we can take it.
A pilot starts within 5 working days of agreed scope · Nothing upfront · No seat licences