Business insight

Every AI Model Has a Job. The Trap Is Using the Wrong One.

Ten families of AI model in plain English, and how to tell which one suits the job before somebody else chooses for you. With the four questions that expose a recommendation built on nothing.

18 Nov 2026 9 min readAdvanced Analytics & AI

Summarised from the Go2Cab business insight series

The 5G problem

Most of us now carry a phone that says 5G on the screen. Few could explain precisely what 5G is, how it differs from 4G, or where its limits sit, and there is no particular reason we should. It is a technical detail the network quietly takes care of. The phone either works in the lift or it does not, and judging it by that is perfectly sensible.

AI is being discussed in exactly the same way, and there the analogy stops being harmless. “We are doing AI”, “we are using a large language model”, “we are moving to an agentic platform” is the new 5G. The vocabulary travels much faster than the understanding behind it, and the decisions being made on the strength of that vocabulary are considerably harder to unwind than a phone contract.

Why the comparison breaks down

One network generation replaces another and, for most purposes, they are interchangeable. AI models* are not. The choice between them decides four things at once: how accurate the answer is, what it costs to produce at volume, where your data lives, and what happens when the output is wrong.

The advice reaching the executive floor is, in our experience, well intentioned. Vendors, consultants, internal champions and capable team members recommend what they have seen work, on the problems they happened to face. The field moves faster than any shared definition of it, so a model that was excellent for one job gets carried into another where it was never suited. Nobody behaved carelessly, and the recommendation is still wrong.

The cost rarely arrives as an invoice line. It arrives eighteen months later as a workflow nobody trusts, a renewal nobody can justify, or an assistant quietly switched off after it produced confident nonsense for the third time.

Ten families of model, in plain English

The technical literature names far more architectures than a leader needs. Ten categories cover what is commercially relevant today. Read them as tools on a bench rather than as a ranking: the newest is not automatically the right one, and the cheapest is not automatically the wrong one.

For each, the useful question is the same. What is it, where does it earn its place, and what does it cost you in return?

Language and reasoning

  • Large language model (LLM). A digital reader and writer trained on an enormous amount of text. Best at first-line support replies, drafting internal communications and summarising dense contracts. The trade-off: it can state an invented fact with total composure, and every answer carries a running cost. GPT-4-class models and Meta’s Llama are the usual reference points.
  • Small language model (SLM). The same idea, shrunk to run on a laptop, a phone or a device on the shop floor. Best at on-device feedback sorting, offline field reports, and anywhere the data should not leave the building. The trade-off: it reasons less well and handles less context, so complex multi-step work is out of reach. Microsoft’s Phi models and TinyLlama sit here.
  • Reasoning language model (RLM). A model trained to work the problem through in steps before answering, instead of reaching for the next likely word. Best at code diagnosis, financial modelling and multi-step compliance analysis. The trade-off: noticeably slower and several times more expensive per answer, which is a poor trade for routine work. OpenAI’s o-series and DeepSeek-R1 are examples.

Seeing, reading and acting

  • Vision-language model (VLM). A model that reads and looks at the same time. Best at inspection, damage assessment from photographs, and pulling numbers out of a scanned invoice. The trade-off: very sensitive to lighting, angle and image quality, so it needs a designed test on your own images rather than a vendor demo. CLIP and Flamingo are the research reference points.
  • Multimodal large model (MLLM). One model that takes text, high-resolution images, audio and video together. Best where several formats have to be understood at once, such as a service call where the voice, the screen and the transcript all matter. The trade-off: the heaviest infrastructure of the ten, and the integration effort is almost always underestimated. Gemini is the best-known example.
  • Vision-language-action model (VLA). It looks, understands the instruction and moves something. Best in warehousing and on production equipment, and for now only inside controlled or simulated environments. The trade-off: a mistake is physical, so safety engineering has to come before capability. Google’s robotics models and SmolVLA are early examples.
  • Large action model (LAM). A planner that clicks, fills and submits across applications rather than writing a paragraph. Best at automations spanning systems with no modern interface, for example onboarding someone across an ageing HR system and a newer finance package. The trade-off: it breaks when a screen layout changes, and an agent with permission to press buttons deserves serious access control. Rabbit’s original Large Action Model put the term on the map, and the newer browser-use agents extend it.

How the engines are built

  • Mixture of experts (MoE). A network of specialist sub-models where only the relevant few are activated for each request. Best for high-volume routing between legal, technical and billing knowledge. The trade-off: economical to run, difficult and expensive to build, and it needs specialised deployment skills. Mistral’s Mixtral is a well-known example, and most frontier systems now use the technique internally.
  • State space model (SSM). An alternative to the transformer that reads very long records quickly: multi-year ledgers, continuous sensor streams, hour-long recordings. Best where the answer depends on something said far earlier in a very long document. The trade-off: a younger ecosystem with fewer ready-made parts. Mamba and RetNet are the names to know.
  • Foundation model. The broad base layer everything else is adapted from. Best when you intend to build a proprietary assistant on your own material and fine-tune it deliberately. The trade-off: it arrives carrying whatever bias is in its training data, and fine-tuning is a project with governance attached, not a toggle.

The one-page comparison

Short enough to take into a steering meeting. If a proposal cannot be placed in one of these rows, ask why not.

LLM

Core strength

Versatile writing and summarising

Typical business use

Support replies, document drafting

Main trade-off

Invites invented facts

SLM

Core strength

Cheap, private, runs locally

Typical business use

On-device analysis, offline reports

Main trade-off

Weaker complex reasoning

RLM

Core strength

Multi-step logic and maths

Typical business use

Code diagnosis, financial modelling

Main trade-off

Slow and costly per answer

VLM

Core strength

Reads and interprets images

Typical business use

Inspection, scanned documents

Main trade-off

Fragile to image quality

MLLM

Core strength

Text, image, audio and video together

Typical business use

Rich multimedia analysis

Main trade-off

Heavy integration effort

VLA

Core strength

Controls physical equipment

Typical business use

Automated warehousing, robotics

Main trade-off

Physical safety limits

LAM

Core strength

Operates software across systems

Typical business use

Multi-app workflow automation

Main trade-off

Breaks when interfaces change

MoE

Core strength

Efficient scaling to very large models

Typical business use

High-volume specialist routing

Main trade-off

Hard to build and run

SSM

Core strength

Very long context at speed

Typical business use

Ledgers, sensor and audio archives

Main trade-off

Immature tooling

Foundation

Core strength

Customisable base intelligence

Typical business use

Proprietary assistants, fine-tuning

Main trade-off

Upfront cost and inherited bias

Four questions that sort a recommendation

  • Which decision does this improve, and who owns it? If nobody can name the decision, the measure and the person accountable, the model has no job and cannot be evaluated afterwards.
  • What does a wrong answer actually cost? A drafted email is cheap to correct. A three-way match exception, a compliance finding or a safety call is not, and the acceptable error rate differs enormously between them.
  • What is the cost per thousand answers, not per seat? Model economics scale with usage, not with licences. A per-seat quote for something that runs on every transaction is a different number entirely.
  • Who tested it against your data? A published benchmark measures someone else’s problem. Insist on a measured trial on your own process, with the result written down before the purchase order.

How we choose here

We start where the evidence starts. Lean Six Sigma defines the measurable objective and the baseline, so the model is being asked a specific question instead of a general one. Statistical software then proves whether the result holds: process capability, designed experiments, and the calibration that shows a model fits your operation rather than a benchmark.

Only then does the model choice matter, and it usually gets smaller rather than larger. The right answer is the least capable model that meets the accuracy the decision demands at a cost the volume justifies, whether that lands in Microsoft Dynamics 365, Business Central, Copilot or something built around them.

So if a recommendation is in front of you and it is not clear which of these ten it is, or which decision it serves, that silence is the answer. Ask us and we will tell you in plain English what it does, what it cannot do, and what it should cost. No obligation, and no interest in selling you the largest model in the room.

*All product names, logos, brands, and AI models mentioned in this article are the property of their respective owners and are used for identification purposes only, without implying endorsement or affiliation with Go2Cab.

The takeaway

Ask which model, for which decision, tested against whose data. If the answer you are given is simply “AI”, the answer is not finished.

Advanced AnalyticsAIModel SelectionGovernanceOperational Excellence
All business insights