1) AI in one sentence

AI is software that spots patterns in data and uses them to predict, recommend, generate, or automate.

If you want a second sentence: it works best when your data isn’t a dumpster fire.

 

2) The AI family tree (what’s what)

Artificial Intelligence (AI)

Umbrella term. Everything from rule-based automation to modern generative models gets called “AI”, because sales.

Machine Learning (ML)

AI that learns from examples instead of being explicitly programmed for every rule.

Deep Learning (DL)

A type of ML that uses neural networks with many layers. Great at language, images, and anything that’s messy and high-volume.

Generative AI (GenAI)

AI that creates text, images, audio, video, or code.

Large Language Model (LLM)

A GenAI model trained on huge amounts of text, designed to predict the next token (word-ish chunk) in a sequence.

Foundation model

A big, general-purpose model you adapt to your use case with prompting, tools, or fine-tuning.

 

 

3) The acronyms you keep nodding at

LLM — Large Language Model
    • What it is: A model that predicts the next token.
    • What it does: Drafts, summarises, answers questions, writes code, classifies text.
    • What it doesn’t do: Know your business unless you give it your context.
GPT — Generative Pre-trained Transformer
    • Pre-trained: Learns general patterns from lots of public text.
    • Transformer: The architecture that makes the whole thing fast and good at context.
NLP — Natural Language Processing

Anything that helps computers work with human language: search, sentiment, summarisation, translation, intent detection.

RAG — Retrieval-Augmented Generation
    • What it is: Your LLM plus a retrieval layer that fetches relevant company info (docs, FAQs, policies) at the moment of answering.
    • Why you care: Reduces hallucinations, keeps answers grounded, uses your latest content.
    • Translation: “Stop the model making things up by letting it read your stuff.”
Vector database (or “embeddings store”)

A database designed to store embeddings (numeric fingerprints) so you can search by meaning, not just keywords.

Embeddings
    • What they are: Vectors (lists of numbers) representing meaning.
    • What they’re for: Semantic search, clustering, deduplication, recommendations.
Tokens

Chunks of text an LLM processes. More tokens = more cost and more time.

Context window

How much text the model can consider at once. If you dump a 200-page policy into a tiny context window, it will do what humans do: skim and guess.

Fine-tuning

Training the model further on your data to behave in a more specific way.

    • Good for: Consistent style, structured outputs, specialised classification.
    • Not great for: “Teaching” it your entire knowledge base (use RAG for that).
Prompt engineering

Writing instructions that reduce nonsense.

    • Better name: “Being clear.”
Agent

An AI that can plan steps and use tools (search, CRM, calendar, tickets) to complete tasks.

    • Think: “LLM with access and guardrails.”
Guardrails

Rules and checks that prevent risky outputs: policy filters, approved sources, role-based access, human review.

PII — Personally Identifiable Information

Data that identifies people. If you don’t know your PII posture, you’re not “doing AI”. You’re doing “future headlines”.

ROI — Return on Investment

The only acronym that matters after month one.

 

 

4) How AI works (the short version)

Traditional automation
    1. You define rules.
    2. It follows rules.
    3. It breaks when life happens.
Machine learning
    1. You give examples (data).
    2. It learns patterns.
    3. It predicts outcomes on new inputs.
Generative AI / LLM
    1. It’s trained to predict the next token.
    2. It becomes good at language by learning billions of patterns.
    3. It outputs plausible text.
    4. It can be useful when you provide:
      • clear instructions (prompts)
      • business context (RAG)
      • tools (agents)
      • constraints (guardrails)

Important: LLMs are not truth engines. They are fluency engines. Your job is to bolt on accuracy.

 

5) What AI is brilliant at (and what it’s rubbish at)

Brilliant at
    • summarising documents and calls
    • drafting first versions (emails, proposals, job ads, content outlines)
    • classifying and routing enquiries
    • extracting data from messy text
    • customer support on known topics
    • internal knowledge search
    • spotting trends and anomalies in data
    • personalisation and recommendations
Rubbish at
    • being correct without sources
    • reading your mind
    • understanding your business politics
    • making strategic decisions (it will happily hallucinate confidence)
    • handling edge cases without guardrails

6) Best business use cases (the ones that actually pay)

A) Customer service without the chaos

AI helpdesk copilot for agents:

    • suggests replies
    • summarises the conversation
    • surfaces relevant policies
    • logs the ticket

Benefit: faster handling time, better consistency, less training burden.

B) Sales acceleration (no, not spam)

Sales enablement copilot:

    • drafts outbound based on CRM notes
    • summarises calls into next steps
    • generates proposal first drafts
    • finds cross-sell signals in support tickets

Benefit: more selling, less admin, fewer “just circling back” emails.

C) Marketing efficiency (the sane version)
    • content briefs and outlines
    • ad copy variations
    • landing page iteration
    • SEO clustering and FAQ creation
    • competitor and market summarisation

Benefit: higher output with the same team. Not “replace the team”, unless you enjoy chaos.

D) Back-office automation
    • invoice processing
    • document extraction
    • contract review support
    • HR: policy Q&A, onboarding assistants

Benefit: fewer manual steps, fewer errors, quicker throughput.

E) Analytics and forecasting
    • demand forecasting
    • churn prediction
    • lead scoring
    • anomaly detection (fraud, downtime, cost spikes)

Benefit: earlier interventions, better allocation of time and budget.

F) Operations and maintenance
    • predictive maintenance
    • route optimisation
    • inventory optimisation

Benefit: fewer outages, lower fuel/time waste, better margins.

7) Corporate examples (what big companies did, and why it worked)

These aren’t fairy tales. They’re patterns you can copy at a sensible scale.

Netflix — Recommendations
    • Use: ML-driven recommendations.
    • Benefit: higher engagement and retention.
    • What to copy: use behaviour data to personalise journeys (content, products, offers).
Amazon — Recommendations + logistics
    • Use: recommendations, demand forecasting, warehouse optimisation.
    • Benefit: more basket value, fewer stock issues, faster fulfilment.
    • What to copy: combine transactional data with forecasting to reduce waste.
UPS — Route optimisation (ORION)
    • Use: optimisation to reduce miles, fuel, time.
    • Benefit: lower operational costs and improved delivery efficiency.
    • What to copy: AI that saves minutes per job scales into real money.
JPMorgan Chase — Document review automation (COiN)
    • Use: automated extraction and review of contract data.
    • Benefit: faster processing, reduced manual review.
    • What to copy: start with high-volume documents where accuracy can be verified.
Unilever — Hiring support
    • Use: AI-assisted screening and assessment.
    • Benefit: faster hiring process and more consistent evaluation.
    • What to copy: use AI to standardise early-stage admin, keep humans for final decisions.
Microsoft / GitHub — Copilots
    • Use: AI copilots for productivity and coding.
    • Benefit: faster drafting, reduced context switching, productivity uplift.
    • What to copy: deploy copilots where staff spend time in email, docs, spreadsheets, tickets.

Note: the “benefit” only shows up when workflows change. Buying tools and hoping isn’t strategy.

 

8) The decision-maker’s checklist (no fluff)

 

Step 1 — Pick the right problem

Good AI problems:

    • high-volume
    • repetitive
    • measurable
    • painful
    • has data or documents

Bad AI problems:

    • vague (“make us innovative”)
    • unmeasurable (“improve vibes”)
    • politically sensitive with no ownership
Step 2 — Decide the approach
    • Automation rules if the logic is clear.
    • ML model if you have labelled data and need predictions.
    • LLM + RAG if you need language + your company knowledge.
    • Agent if it must do multi-step tasks across systems.
Step 3 — Define success metrics

Pick 2–3, such as:

    • time saved per task
    • cost per ticket
    • conversion rate
    • response time
    • error rate
    • churn rate
    • average order value
Step 4 — Confirm data readiness
    • Where is the data?
    • Who owns it?
    • Is it accurate?
    • Can the AI access it securely?
    • Are your policies and FAQs up to date?
Step 5 — Control risk
    • role-based access
    • logging and monitoring
    • human review for high-stakes outputs
    • approved sources (RAG)
    • PII handling rules
Step 6 — Pilot, then scale

Run a pilot for one workflow. If it works, expand.

9) Common traps (how AI projects die)

  • “We’ll just use ChatGPT.” For what, exactly?
  • No data ownership. Everyone wants outcomes, nobody wants responsibility.
  • No workflow change. Staff keep working the old way, so benefits never appear.
  • No guardrails. Then someone pastes customer data into the wrong place.
  • Trying to automate judgement. AI supports decisions; it shouldn’t replace accountability.

 

 

10) Practical starting points (what I’d do first)

If you’re a typical business with limited time and unlimited ambition:

  1. Internal knowledge assistant (LLM + RAG): policies, product info, onboarding, SOPs.
  2. Customer service copilot: faster replies, better consistency, better handoffs.
  3. Sales call + meeting summaries: action items, CRM updates, proposal drafts.
  4. Marketing production line: briefs → drafts → QA → publish. Humans still own the final output.
  5. Document extraction: invoices, POs, contracts, applications.

These are boring. That’s why they work.

 

 

11) “Should we implement AI?”

If you answer yes to most of these, you’re ready:

  • We have repeatable processes.
  • We can measure outcomes.
  • We have usable data or documents.
  • We can assign an owner.
  • We can manage access and compliance.

If you answer no, don’t panic. Fix the basics first. AI will still be there, waiting to be overhyped.

 

 

12) Final note from Matthew

AI won’t replace your business.

But a competitor using AI properly might replace your margins.

If you want help choosing the first workflow to pilot, SOOM® can map it, build it, and keep it grounded in reality. You’re welcome.