Foundations
How generative AI differs from earlier AI, what foundation models and prompts really are, the business tasks it handles well, its known weaknesses, and how to deploy it safely.
8 minute read · Written by the Fresh Mango AI team
Generative AI is a class of artificial intelligence that creates new content — text, images, code, audio or video — in response to a prompt, rather than only classifying or predicting from existing data. It is powered by large foundation models trained on very large datasets, and it is the technology behind Microsoft Copilot, ChatGPT, Claude and Google Gemini.
Earlier commercial AI was narrow and discriminative. A model was trained for one job — approve or decline, fraud or not fraud, defect or no defect — and it output a score or a label. Generative AI inverts this. A single foundation model, trained on a very broad corpus, can produce fluent original output across an enormous range of tasks without being retrained for each one. You steer it with instructions in natural language instead of by writing code.
That generality is what made AI a mainstream workplace technology rather than a specialist data science capability. It is also what makes governance harder: the same tool that helps a marketing executive rewrite a landing page can be used by a finance manager to analyse a payroll file, and neither use was individually procured or risk-assessed. Organisations that succeed with generative AI put a usage framework in place early, because the technology arrives everywhere at once.
A very large neural network trained on broad data — GPT, Claude and Gemini families are examples. It provides general language and reasoning capability that applications are built on top of.
The instruction you give the model. Effective prompts state the role, the task, the audience, the constraints and the desired format rather than asking an open question.
The amount of material the model can consider at once. Modern models accept the equivalent of hundreds of pages, which is what makes document analysis practical.
Retrieval-augmented generation: relevant documents are fetched from your systems and passed to the model so answers reflect your real content and can be cited.
Additional training on your own examples to shape style or specialised behaviour. Useful, but far less often necessary than vendors imply — grounding solves most business problems more cheaply.
The strongest generative AI use cases share a shape: high volume, language-heavy, currently done from scratch each time, and reviewed by a competent human before it leaves the organisation. Drafting client correspondence, proposals and reports. Summarising long documents, meetings and email threads. Rewriting technical content for a non-technical audience. Producing structured data from unstructured input. Generating test cases, formulas and code. Translating and localising. Answering internal policy and process questions from a grounded knowledge base.
We consistently see the largest, most reliable gains where the alternative is a blank page. A first draft produced in seconds and edited in minutes replaces an hour of starting, stalling and restarting. Across a professional services firm that compounds quickly, which is why our Department-Specific AI Productivity Pack concentrates on embedding a small number of high-frequency drafting and summarising patterns rather than teaching everything a tool can do.
Models can produce confident, fluent and wholly incorrect statements — invented statistics, non-existent case law, misattributed quotes. Controls: ground answers in retrieved source documents, require citations, and mandate human verification of any external or regulated output.
Staff pasting confidential material into consumer tools that may retain or train on it. Controls: provide a sanctioned enterprise tool, block or discourage unsanctioned ones, and state clearly in policy what may never be entered into any AI system.
Generated content may resemble training material, and ownership of AI output is still legally unsettled in several jurisdictions. Controls: keep humans in the authorship loop, avoid generating brand-critical assets purely from prompts, and record where AI contributed.
A sound sequence is: choose a small number of high-value use cases; confirm the data those use cases touch is correctly permissioned; select an enterprise-grade platform with contractual data protections; write a short, readable usage policy; train the people who will actually use it against their real work; measure before and after; and review quarterly. Skipping the permission and policy steps is the most common cause of a stalled or withdrawn rollout.
For most organisations the platform decision is less consequential than it feels. Microsoft Copilot, ChatGPT Enterprise, Claude and Gemini are all capable; the difference in outcome between them is far smaller than the difference between a trained, governed deployment and an untrained, ungoverned one.
Stanford Institute for Human-Centered AI (CRFM)
Meta AI research (arXiv)
UK Information Commissioner's Office
Book a free 30-minute discovery session with an AI consultant, or request information and we will send the detail you need within one business day.
Recommended · Free · 30 minutes
Opens our live booking calendar in a new tab — pick any slot that suits you.
Prefer not to book yet?
Send a short enquiry and we will come back with the detail you need.
FAQs
No. ChatGPT is one product built on generative AI models. Generative AI is the underlying category, which also includes Microsoft Copilot, Claude, Gemini, image generators and code assistants.
Because it generates statistically plausible language rather than retrieving verified records. If the correct information is not in its context, it will still produce fluent text. Grounding answers in your own documents and requiring citations dramatically reduces the problem.
Usually not. Retrieval-based grounding gives the model access to your current documents without the cost, delay and maintenance burden of fine-tuning, and it respects existing permissions. Fine-tuning is worth considering for highly specialised style or classification tasks.
We advise against it. Any content that leaves the organisation, informs a client decision or carries regulatory weight should be reviewed by a competent person who takes responsibility for it. This should be written explicitly into your AI usage policy.
Provide a sanctioned alternative first — prohibition without provision simply drives usage underground. Then combine clear policy, targeted training and technical controls such as data loss prevention and conditional access.
Related articles
Explore the connected topics in the Fresh Mango AI Knowledge Centre.
Foundations
Artificial intelligence (AI) is the field of computer science concerned with building systems that perform tasks normally requiring human intelligence…
Read this guidePlatforms
ChatGPT is a conversational AI assistant developed by OpenAI, built on its GPT family of large language models. Launched publicly in November 2022, it…
Read this guidePlatforms
Microsoft Copilot is Microsoft's family of generative AI assistants. Microsoft 365 Copilot is the version embedded in Word, Excel, PowerPoint, Outlook…
Read this guideBrowse the full AI Knowledge Centre, see sector guidance in industries, or review measured results in our case studies.
Decision brief
Straight answers to the five questions that decide whether an AI project is worth starting.
Fresh Mango AI is the artificial intelligence practice of Fresh Mango Technologies, an IT, cyber security and cloud provider that has supported businesses since 2004 from offices in Ripon, Leeds, Skipton and Tortola in the British Virgin Islands. The same engineers who secure your identity, data and Microsoft 365 tenant advise on your AI adoption, so recommendations are grounded in what your estate can actually support rather than in vendor marketing.
Clients typically release several hours per person per week on drafting, summarising, searching and reporting, shorten document and approval cycle times, and remove manual re-keying between systems. Every engagement starts by baselining the work involved so that the benefit is measured in hours released and cycle time reduced, not in licences purchased.
Doing nothing is not a neutral position. Staff adopt consumer AI tools on their own, so company and client data leaves your control without record; competitors compress the cost of proposals, reporting and service delivery; and permission sprawl inside your file estate remains unaddressed, which becomes an incident the moment AI search is switched on. Delay also compounds the UK GDPR and EU AI Act governance work that will eventually be required of you anyway.
You book a free 30-minute discovery session. We ask about your objectives, systems and constraints, tell you honestly whether AI is the right answer, and set out a recommended first step — usually an AI Productivity & Readiness Assessment or an AI Governance & Security Project. You receive a written summary and a proposal only if there is a clear case for one.
Book a free 30-minute discovery session with an AI consultant using our live booking calendar, or request information and we will reply within one business day. You can also call the UK office on +44 (0) 1765 606700 or the BVI office on +1 (284) 340 0466.
Recommended next step
Free, 30 minutes, no obligation — speak directly with an AI consultant about your objectives and the fastest safe route to results.
Book a Free 30-Minute Discovery SessionRequest InformationTalk it through with an AI consultant, or request written information — whichever suits you.
Your next step
Whether you're exploring AI for the first time or looking to scale existing AI initiatives, our specialists can help you identify practical opportunities and avoid common pitfalls.
