Your accounting package has been using AI for years. Every time Xero or QuickBooks suggests an account code for a transaction, a machine-learning model trained on your bookkeeping history is making a prediction. Nobody called it AI until ChatGPT arrived. Now every vendor calls everything AI, and the label has stopped telling you anything useful. The two main kinds do very different jobs, carry different risks, and deserve different questions in a demo. Knowing which one you’re looking at is worth ten minutes of your time.
What’s the actual difference between traditional AI and generative AI?
Traditional AI makes decisions from structured data. It scores, classifies, ranks and routes. Is this payment fraudulent? Which lead is warmest? Where should this ticket go? Generative AI produces new content on demand, drafts, summaries, images and code, by predicting what plausibly comes next. One answers scored questions with a measurable right or wrong. The other writes.
The training explains the behaviour. A traditional model learns from labelled history, past invoices marked legitimate or fraudulent, past leads marked converted or lost, and then repeats that judgement at scale on new cases. Its performance is measurable. You can say the fraud model catches 95 in 100 and tune it from there. Stripe Radar works this way, scoring every card payment for fraud risk in real time and blocking the worst.
A generative model learns from internet-scale text and produces the plausible next word, sentence or paragraph. ChatGPT, Microsoft Copilot and Google Gemini all work this way. That design makes them fluent at drafting, summarising and rephrasing, and it also makes them capable of stating things that sound right and are wrong. The industry calls this hallucination. The model has no internal test for truth, only for plausibility.
The two sit side by side the way a spreadsheet and a word processor do, different tools for different jobs.
Why does the difference matter for your business?
Because the two fail in different ways, and the failure you can’t see coming costs more. A traditional model that misjudges a lead score wastes a phone call. A generative model that invents a legal citation, a price, or a policy clause creates a liability that reads perfectly until someone checks it. Where each type sits in your business should follow its failure mode.
The generative failure mode is well documented. A 2023 Stanford Law analysis showed GPT-4 producing credible but fabricated legal citations, complete with plausible case names. The same year, Samsung restricted staff use of ChatGPT after engineers pasted confidential source code into the tool. In both cases a fluent tool was trusted one step further than it deserved.
Traditional AI fails differently, and at scale. The clearest UK example is the 2020 A-level grading algorithm, abandoned within days after it systematically downgraded students based on their school’s history. An opaque model made high-impact decisions about people, and nobody had built in a way to challenge them.
Set against that, the upside is real. McKinsey’s 2023 State of AI survey found 55% of organisations reporting AI adoption in at least one business function, with the biggest reported value in sales, marketing and operations. Microsoft’s early research on 365 Copilot found users completing certain drafting, search and summarising tasks around 29% faster. The gains show up where the failure modes are managed, drafts reviewed, decisions overseen.
Where will you actually meet each one?
Traditional AI arrives as features inside software you already pay for. Auto-coding in your accounting package, fraud scoring on card payments, lead scoring in the CRM, spam filtering in your inbox, ticket routing in the helpdesk. Generative AI arrives as assistants, Copilot drafting emails and summarising meetings, support tools proposing replies, chat interfaces turning scattered notes into documents.
For an owner-managed service business, the traditional side is usually already working for you. Xero and QuickBooks suggest account codes from your bookkeeping history. Your card processor scores payments for fraud. HubSpot and Salesforce rank leads on past conversion data. These features earn their keep on well-defined questions with clean historical data, predicting a no-show, flagging an overdue account, routing a ticket by urgency.
The generative side is where the new spend usually goes. First drafts of proposals and engagement letters. Meeting summaries and action points from transcripts. Turning ten years of scattered notes and emails into a usable set of SOPs. Support tools like Zendesk AI and Freshworks Freddy drafting replies for an agent to approve. In every case the pattern is the same, the machine drafts and a person decides.
And sometimes neither is needed. If the rule fits on a sticky note, send invoices over £5,000 to the finance lead for approval, a plain if-this-then-that automation is cheaper, faster and safer than either kind of AI.
When should you ask which type it is, and when can you ignore the label?
Ask when the tool makes or shapes decisions about people or money. Credit checks, hiring screens, pricing, anything that profiles a customer. Ignore the label when the tool drafts and a person reviews before anything leaves the building. The regulators take the same view. The ICO and NCSC care what the system does with data rather than which architecture sits underneath.
The ICO’s position is that AI processing personal data carries full UK GDPR duties, fairness, transparency, data minimisation and the rest, and that using AI to profile customers or make decisions about individuals can trigger their right to a human review. The NCSC advises treating AI services like any other SaaS with added supply-chain risk, which means vendor due diligence, access controls, and a firm rule against pasting confidential material into public tools.
The EU AI Act goes further, classifying credit scoring and some recruitment systems as high-risk with governance duties attached. It matters to you if you serve EU clients or process EU residents’ data. Closer to home, the CMA has been examining the foundation-model market, and ICAEW has warned accountants about confidentiality and the need for professional scepticism when reviewing AI output.
What does that mean on Monday? Three things are worth an hour of your time:
- A one-page policy naming which tools are allowed and what must never be pasted into them.
- A check of your key vendors’ terms on training data and data location.
- A named human reviewer for anything the machine produces that touches money, clients or staff.
Which related terms are worth knowing?
Machine learning is the training method behind almost all of this, software that learns patterns from examples rather than following hand-written rules. A large language model is the engine inside generative tools like ChatGPT and Copilot. Hallucination is the generative failure mode, fluent output that happens to be false. And automated decision-making is the ICO’s term for the traditional risk that carries legal duties.
You’ll also hear traditional AI called predictive AI or narrow AI. Same thing, three names. Machine learning covers both camps in the technical sense, but in a vendor demo it usually signals the traditional kind.
One honest caveat. The line between the two is blurring. Vendors increasingly build classification features on top of large language models, and generative tools are becoming more factually reliable. In a few years the traditional-versus-generative question may matter less than it does today. The question that will keep mattering is the one underneath it, does this tool draft something a person will review, or does it decide something on its own? Ask that in every demo and you’ll be ahead of the vendor’s own salesperson more often than you’d expect.
