Reference
M&A and AI glossary
Plain-English definitions of the terms you'll hear most when preparing a business for sale and when shipping AI into one. Operator-written, no advisor spin. Each definition is the citation-ready short version; the "in practice" paragraph is the operator answer.
Buying AI services
Rules of thumb
Building blocks
Agents & workflows
Vendors & products
Australian context
M&A terms
The vocabulary of lower-middle-market deals - the words your advisor, your buyer, and your buyer's QofE provider will all use.
Valuation & price
- EBITDA
- EBITDA - Earnings Before Interest, Taxes, Depreciation and Amortization - is a measure of a business's operating cash generation, used as the base on which a buyer applies a multiple to derive enterprise value.
- In practice - Lower-middle-market deals are almost always quoted as a multiple of trailing-twelve-month EBITDA, normalised for one-off and owner-discretionary items. The number a buyer pays a multiple of is rarely the same number on the income statement.
Related reading
- EBITDA multiple
- An EBITDA multiple is the ratio of enterprise value to EBITDA - the headline shorthand for what a buyer is paying for a business. A 5× multiple on $5M EBITDA implies $25M enterprise value.
- In practice - GF Data's Q1 2026 report puts the U.S. middle-market median at ~7.4× across the $10M - $250M EV range; below $10M EV, multiples typically compress to 4-6×. Industry, recurring revenue, owner independence and customer concentration explain most of the spread.
Related reading
- EBITDA normalization
- EBITDA normalization is the process of adjusting reported EBITDA to remove one-off, non-recurring, and owner-discretionary items, producing the run-rate EBITDA a buyer applies a multiple to. Done by the seller's accountant pre-process and re-tested by the buyer's QofE provider.
- In practice - Common adjustments: above-market owner compensation, personal expenses run through the business, one-off legal or restructuring costs, and pandemic-era distortions. The credibility of the normalisation set is what determines whether the buyer accepts the headline EBITDA at all.
Related reading
- Recurring revenue
- Recurring revenue is contractually predictable revenue that renews automatically - subscriptions, retainers, multi-year service contracts. It earns higher multiples than transactional revenue because the cash flow is forecastable and the customer-acquisition cost has already been incurred.
- In practice - John Warrillow's Hierarchy of Recurring Revenue ranks revenue types by predictability - long-term contracts and auto-renewing subscriptions earn premium multiples; consumable resupply and standing orders sit in the middle; one-off transactional sits at the bottom.
Related reading
Deal mechanics
- Letter of Intent (LOI)
- A Letter of Intent (LOI) is a non-binding written offer from a buyer setting out the proposed price, deal structure, exclusivity period and conditions to closing. It is the gate between marketing and diligence.
- In practice - Most lower-middle-market LOIs include 30-90 days of buyer exclusivity and a no-shop. The headline multiple in the LOI is rarely the multiple at close - diligence findings, normalisation disputes and undisclosed concentration commonly compress price 5-15%.
Related reading
- Earn-out
- An earn-out is a portion of the purchase price paid contingent on the business hitting agreed performance targets after close. It transfers execution risk from the buyer back to the seller and typically appears when the buyer wants the headline multiple to look bigger than the cash-at-close justifies.
- In practice - Earn-outs commonly run 12-36 months post-close, tied to revenue, EBITDA or customer-retention targets. In strategic deals with synergy stories, integration risk often shows up as a wider earn-out band. Sale-readiness work materially compresses the earn-out portion.
Related reading
- Strategic buyer
- A strategic buyer is an operating company in the same or adjacent industry, acquiring another business for synergies - combined cost base, cross-sell, geographic expansion, talent or technology - rather than for stand-alone returns.
- In practice - Strategics often pay a higher headline multiple than financial buyers when the synergy story is real, but tighten earn-outs on integration risk. Sellers should emphasise the assets that compound inside the strategic's platform: customer base, IP, regulatory licences, talent.
Related reading
- Financial buyer
- A financial buyer - most commonly a private-equity firm, family office or search fund - acquires a business as a stand-alone investment, holds for 3-7 years, and exits to another buyer. They value cash-flow predictability and growth optionality, not synergies.
- In practice - Financial buyers typically pay closer to a clean stand-alone multiple, with cleaner deal structure, faster close, and meaningful management-equity rollover. Founders frequently stay on as CEO post-close.
Related reading
Diligence & risk
- Quality of Earnings (QofE)
- A Quality of Earnings (QofE) report is an independent accounting analysis commissioned by a buyer to verify that reported EBITDA is sustainable, normalised, and free of one-off or owner-discretionary items. It is the single most important diligence document in lower-middle-market M&A.
- In practice - QofE providers test 24 months of monthly accounts, reconcile to bank statements, and adjust for non-recurring items. Productivity and AI claims that aren't logged with a measurable baseline get adjusted out. A clean QofE is what stops a 5× LOI from closing at 4×.
Related reading
- Owner dependency
- Owner dependency is the degree to which a business's revenue, customers, decisions and operations route through the founder personally. It is the most reliably-applied discount in lower-middle-market M&A - typically 20-40% versus comparable professionally-managed peers.
- In practice - Buyers price owner dependency as risk because the EBITDA they're paying a multiple of may not persist after the founder leaves. The discount is removed by replacing founder-only behaviours with documented systems, a real management layer, and a system-of-record other than the founder's memory.
Related reading
- Customer concentration
- Customer concentration measures how much of a business's revenue depends on its largest customers. The standard buyer threshold is the top-10 share - above 50% it materially compresses the multiple; above 25% with a single customer it usually triggers earn-outs.
- In practice - Concentration is one of the four operational signals strategic buyers most reliably price. Reducing it pre-sale is structural work - it requires real go-to-market diversification, not just contract restructuring.
Related reading
- Sale-readiness
- Sale-readiness is the operational state in which a business can survive a buyer's quality-of-earnings, IT and commercial diligence without material findings that compress the multiple or shift cash to earn-outs. It typically requires 12-36 months of structured work.
- In practice - Friction-level readiness (clean accounts, contracts, IP register) takes 6 months. Structural readiness - owner independence, recurring-revenue mix, management depth, four-quarter operational evidence - takes 24-36 months. Both move the close; only the structural work moves the multiple.
Related reading
More
- AI governance
- AI governance is the set of policies, controls, logs, evaluation frameworks and escalation rules that make a business's AI usage controlled and auditable. The standard reference frameworks are NIST AI RMF (U.S.) and ISO/IEC 42001 (international).
- In practice - Buyers price uncontrolled AI usage as risk - hallucinations, customer-facing errors, regulatory exposure, IP leakage. The minimum viable governance is logged calls, weekly evals, written escalation rules, and a one-page AI policy aligned to one of the named frameworks.
Related reading
AI terms
The AI vocabulary that shows up inside operator-built systems, and inside a buyer's technical diligence on those systems.
Buying AI services
- AI consultant
- An AI consultant helps a business work out where AI creates value, then plans and installs the tools, workflows and training to capture it. For an SMB, the useful ones ship working systems - not just a strategy deck.
- In practice - In Australia, 'AI consultant' spans a wide range - from advisory-only strategists to hands-on operators who roll out Claude, build custom software and automate workflows end to end. Day rates run from roughly $200 to $600+/hour; the deciding question is whether the consultant stays through adoption or leaves you with a plan and a login.
Related reading
- AI agency
- An AI agency designs and builds AI systems for other businesses - chatbots, custom AI products, automations and full team rollouts - as a packaged service. It sits between a traditional digital agency and a software consultancy.
- In practice - For an Australian SMB, the practical test of an AI agency is its delivery model: does it hand over a working system your team actually adopts, or a prototype that stalls after the invoice? The strongest agencies pair strategy with build, automation and training under one roof rather than reselling a thin layer over someone else's tool.
Related reading
- AI automation agency
- An AI automation agency builds automated workflows - powered by AI and tools like n8n, Zapier or Make - that replace repetitive manual work inside a business. The output is reliable systems that run tasks with no person in the loop.
- In practice - Many 'AI automation agencies' resell a thin no-code layer and demoware. The ones worth hiring scope automations against real wage-equivalent savings, build for reliability and error-handling, and connect to your actual stack - CRM, billing, scheduling - rather than a happy-path demo.
Related reading
Rules of thumb
- The 10-20-70 rule for AI
- The 10-20-70 rule, popularised by BCG, holds that the value of an AI transformation comes roughly 10% from the algorithms, 20% from technology and data, and 70% from people and process change. The model is the easy part; adoption is the work.
- In practice - It is the single best argument against the 'buy some licences and hope' approach to AI. The 70% - training, workflow redesign, governance and change management - is exactly where most SMB AI rollouts fail, and exactly where XLev concentrates its effort.
Related reading
- The 30% rule for AI
- The 30% rule is a rule of thumb for the human-AI division of labour: let AI handle roughly the routine, predictable share of a workflow while people keep judgment, oversight and high-context decisions. It is a heuristic, not a precise formula, and is sometimes quoted with the proportions reversed.
- In practice - Read it as a discipline rather than a number: automate the repeatable parts of a process, and keep a human accountable for the exceptions, the edge cases and anything customer- or compliance-facing. For an SMB, that boundary is where AI ROI is real and the downside risk stays contained.
Related reading
Foundations
- LLM (Large Language Model)
- A Large Language Model is a neural network trained on broad text data to predict the next token in a sequence. In production it powers chat, drafting, classification, extraction and reasoning over documents - the substrate most operator AI workloads run on.
- In practice - Lower-middle-market businesses don't train LLMs; they call hosted ones (Claude, GPT, Gemini) via API. The work that matters is prompt design, evaluation, governance and the wrap of tools and memory around the model - not the model itself.
- Foundation model
- A foundation model is a large pretrained model - text, image, audio or multimodal - that downstream systems specialise via prompting, retrieval or fine-tuning. "LLM" is the text-only subset of foundation models.
- In practice - Treat foundation-model choice as a procurement decision, not a religion: pin a default for cost/latency, swap models behind your prompts when a better one ships, and keep evals so you can prove the swap was net-positive.
- Context window
- The context window is the maximum amount of text (measured in tokens) a model can read in a single call - input plus its own output. Modern frontier models support 200K - 2M tokens; older or cheaper models cap around 8K - 32K.
- In practice - Bigger windows replace some - but not all - RAG. They're useful for whole-document analysis and long agent traces, but cost and latency scale with what you put in. Don't paste the whole company wiki into every call.
- Token
- A token is the atomic unit a language model reads and emits - typically 3-4 characters of English. Token counts drive both cost (priced per million) and the context-window ceiling.
- In practice - Rule of thumb: 1,000 tokens ≈ 750 English words. Cost models on tokens-in + tokens-out, not on requests, so a verbose system prompt sent on every call is the most common preventable cost line.
- Prompt / system prompt
- A prompt is the natural-language instruction passed to a model. The system prompt is the persistent instruction that frames a model's role, tone, allowed tools and refusal rules; the user prompt is the per-turn request.
- In practice - Production system prompts are versioned source code, not chat-window experiments. Treat them with the same review and changelog discipline as application code - and test every change against an eval set before shipping.
Building blocks
- RAG (Retrieval-Augmented Generation)
- RAG is a pattern that retrieves relevant passages from a knowledge base at query time and passes them to the model as context, so answers are grounded in your data instead of the model's training set.
- In practice - RAG is how operator-built AI tools answer "what does our SOP say about X" without hallucinating. It needs three boring things done well: clean source documents, a good chunking + embedding strategy, and citations rendered back in the answer so users can verify.
Related reading
- Embeddings
- An embedding is a numeric vector that represents the meaning of a piece of text, image or other content. Two items with similar meaning produce vectors close together in space - the foundation of semantic search and RAG.
- In practice - Embeddings are cheap, fast and durable: generate once at ingest, store, then re-use for every search. Re-embed when you change models - silently mixing embedding versions is a classic source of "search got worse" bugs.
- Vector database
- A vector database stores embeddings and serves nearest-neighbour search at scale. Examples: Pinecone, Weaviate, Qdrant, pgvector. It is the storage layer that makes RAG fast enough to use in real product flows.
- In practice - For SMB workloads, pgvector inside an existing Postgres is usually enough - adding a separate vector DB is the kind of architectural complexity buyers' technical diligence will flag without a clear scale reason.
- Fine-tuning
- Fine-tuning is the process of further-training a foundation model on your own labelled data so it specialises in a domain, format or style. It is distinct from prompting and from RAG, and far more expensive to maintain.
- In practice - Fine-tune last, not first. In practice, 90% of operator AI use cases are solved by better prompts, retrieval and tool design. Fine-tuning is justified when you need a fixed output format at scale, or domain accuracy that prompting can't reach.
- MCP (Model Context Protocol)
- Model Context Protocol is an open standard for connecting language models to external tools, data sources and actions over a uniform interface. It standardises what was previously bespoke per-vendor plumbing.
- In practice - MCP matters operationally because it makes "swap the model" and "add a new tool" both routine instead of a rebuild. For a buyer, MCP-shaped integrations look like normal software - they're inspectable, swappable, and cheap to maintain.
Agents & workflows
- AI agent
- An AI agent is a system that uses a language model in a loop - observe, plan, call a tool, observe the result, decide the next step - to complete a multi-step task without per-step human input.
- In practice - The bar for shipping an agent into a real workflow is not "does the demo work" but "does it fail safely when the world changes." Production agents need bounded permissions, rate limits, full call logs, and an obvious off-switch.
- Agentic workflow
- An agentic workflow is a business process in which one or more AI agents handle the routine path end-to-end and only escalate to a human on defined edge cases. It is the operating model that delivers the wage-equivalent savings buyers price.
- In practice - The replicable pattern: pick a process with high volume + clear inputs + reversible outputs (inbox triage, quote drafting, lead qualification), build the agent against an eval set, instrument every run, and measure the human-minutes-saved as a baseline-vs-current chart for diligence.
- n8n / workflow automation
- n8n is an open-source workflow automation platform that wires APIs, databases, AI calls and business systems together via a visual node-based editor. It (and peers like Make and Zapier) is the connective tissue most operator AI lives on.
- In practice - Workflow automation earns its keep when each workflow is named, owned, version-controlled and observable. "A folder of n8n flows nobody can describe" is technical debt; "15 named workflows with logs, owners and a runbook" is an operational asset.
- Human-in-the-loop
- Human-in-the-loop (HITL) is a design pattern in which an AI system proposes an action and a human approves, edits or rejects it before it takes effect. It is the standard way to ship AI into high-stakes flows without taking on uncontrolled risk.
- In practice - Most operator AI deployments should ship HITL first, measure the human override rate, and only move to full automation on the slices where overrides drop below a defined threshold. The override rate is itself a great diligence artefact.
Vendors & products
- ChatGPT (OpenAI)
- ChatGPT is OpenAI's consumer and business AI product, available in Free, Plus, Team and Enterprise tiers. It runs on the GPT family of models and powers Custom GPTs, Code Interpreter and a growing agent surface.
- In practice - For an Australian SMB, ChatGPT Team is the closest competitor to Claude.ai Teams. The choice usually pivots on writing quality (Claude generally leads on long-form), ecosystem (ChatGPT has broader plugin/app coverage) and the specific workflows in scope.
Related reading
- Google Gemini
- Gemini is Google's family of frontier AI models, distributed via the Gemini app, the Workspace add-on and the Google Cloud Vertex AI platform. It is the default AI surface for businesses running on Google Workspace.
- In practice - For SMBs already on Workspace, Gemini is the path of least resistance for in-document AI. It is weaker than Claude on long-form drafting and harder to govern than Claude.ai Teams for line-of-business teams, but stronger on multimodal (image, video) tasks for some workflows.
Related reading
- Microsoft 365 Copilot
- Microsoft 365 Copilot is the AI assistant bundled into Microsoft 365 - Word, Excel, Outlook, Teams, PowerPoint - powered by OpenAI's GPT models and Microsoft's own orchestration. Licensed per user on top of an existing M365 subscription.
- In practice - For SMBs deeply on Microsoft 365, Copilot for M365 is the in-app AI surface of least friction. It is governance-friendly (Microsoft data boundary, EU/AU data residency options) but weaker than Claude on long-form reasoning and Project-style scoped workspaces.
Related reading
- Anthropic API
- The Anthropic API is the developer-facing endpoint that exposes Claude models for programmatic use - chat completions, tool use, streaming, vision and the Agent SDK. It is the substrate for custom AI software built on Claude.
- In practice - For an SMB, the Anthropic API is where custom builds happen - internal knowledge assistants, customer-facing AI features, structured-extraction pipelines. The API also lets you mix-and-match Claude models per task (Haiku for cheap classification, Sonnet for general work, Opus for the hard reasoning).
Related reading
- Claude Projects
- Claude Projects are scoped Claude workspaces with curated context - documents, instructions and prior chats - that the model uses for every conversation in that workspace. They are the unit of organisation in Claude.ai Teams and Enterprise.
- In practice - For an SMB, the right rollout pattern is one Project per business function (operations, sales, marketing, finance, support), each loaded with the SOPs, frameworks and templates that function uses. Cross-function 'Company Brain' Projects almost always fail.
Related reading
- Claude Code
- Claude Code is Anthropic's command-line coding assistant. It plugs into a developer's terminal, reads the codebase, edits files, runs commands and ships work alongside the engineer instead of inside a chat window.
- In practice - For an SMB engineering team, Claude Code is the highest-leverage way to get Claude doing real work. Set up with the team's conventions, codebase context and review rules baked in, it shortens cycle time on routine work (tests, refactors, migrations) and frees senior engineers for harder problems.
Related reading
- Claude Skills
- Claude Skills are custom capabilities you teach a Claude Project: a named instruction pack with example prompts, voice rules, frameworks or jargon that the model should apply when invoked. They sit between a system prompt and a tool.
- In practice - For an SMB, Skills are how Claude learns the way your business writes a proposal, runs a stand-up or evaluates a candidate. Each Skill is a named recipe stored in a Project so anyone on the team can invoke it with one line.
Related reading
- GitHub Copilot
- GitHub Copilot is Microsoft and GitHub's AI coding assistant. It sits inside the IDE (VS Code, JetBrains, Neovim) as an inline autocomplete plus an in-IDE chat panel, with optional agent features in Visual Studio Code.
- In practice - Copilot is the safe default if your team already runs on GitHub and Visual Studio Code, with enterprise SSO and policy controls available. It is weaker than Claude Code on long-running agentic work and weaker than Cursor on whole-codebase navigation, but unmatched on inline autocomplete inside an existing IDE.
Related reading
- Cursor
- Cursor is an AI-first fork of Visual Studio Code from Anysphere. Its differentiator is whole-codebase context and agent-style multi-file edits driven by Claude or other frontier models.
- In practice - Cursor is the strongest day-to-day AI coding environment when whole-codebase reasoning matters. It ships with Claude Sonnet and Opus baked in, supports MCP, and is the typical choice for SMB engineering teams that want the agentic experience inside their IDE rather than at the terminal.
Related reading
- Lovable
- Lovable is a browser-based AI app builder. You describe an app in natural language; Lovable generates a working React + Supabase application that you can iterate on, edit and deploy.
- In practice - For SMB-grade internal tools, Lovable shortens the path from idea to working software dramatically. XLev's founder ranked in the top 1% of Lovable builders globally in 2025. It is best paired with a more conventional handoff to Claude Code or a developer team once the app graduates from prototype to production.
Related reading
Australian context
- Australian Privacy Principles (APPs)
- The Australian Privacy Principles are the 13 principles in the Privacy Act 1988 (Cth) that govern how Australian organisations collect, use, disclose and store personal information. The Office of the Australian Information Commissioner (OAIC) enforces them.
- In practice - For SMBs using AI, the relevant APPs are APP 6 (use and disclosure), APP 8 (cross-border disclosure) and APP 11 (security of personal information). Claude.ai Teams and Enterprise tiers publish data handling commitments that allow APP-compliant use; consumer tiers sit outside the governance perimeter and should be restricted for work data.
Related reading
- AHPRA
- The Australian Health Practitioner Regulation Agency (AHPRA) is the national agency that registers and regulates 15 health professions in Australia. Its codes of conduct apply to any AI used inside a registered practitioner's clinical work.
- In practice - For Australian medical and allied health SMBs using AI, AHPRA's expectations are: a registered practitioner remains accountable for every output, AI-generated clinical content is reviewed before use, and the use of AI is disclosed to patients where it materially affects their care. Build governance around these from day one.
Related reading
- R&D Tax Incentive
- The Research and Development Tax Incentive (RDTI) is the Australian Federal Government's primary R&D support program, jointly administered by AusIndustry and the ATO. It offers a refundable or non-refundable tax offset on eligible R&D expenditure.
- In practice - Custom AI builds that involve genuine technical experimentation (novel architectures, new evaluation methods, original integrations) can qualify for the RDTI. Off-the-shelf AI subscription costs do not. Eligibility decisions should always be sense-checked by an RDTI-experienced advisor before claiming.
Related reading
Operations & risk
- Hallucination
- A hallucination is an output a language model presents with confidence that is factually wrong, fabricated or unsupported by its sources. It is a structural property of how LLMs work, not a bug to be patched away.
- In practice - Production systems mitigate, not eliminate, hallucinations: ground answers in retrieved sources, render citations, constrain outputs to a schema, run evals against a fact set, and require human review on high-stakes flows. A buyer's diligence team will ask exactly this.
Related reading
- Eval
- An eval is an automated test for an AI system: a set of inputs, expected behaviours and a scoring function (exact match, model-judged, human-graded). Evals are how you know a prompt or model change made the system better, not just different.
- In practice - If you don't have evals, you don't have a system - you have a demo. The cheapest useful eval is 30 hand-written input/output pairs, run every time the prompt or model changes, with results logged in a CSV the team reviews weekly.
- AI governance
- AI governance is the set of policies, controls, logs, evaluation frameworks and escalation rules that make a business's AI usage controlled and auditable. The standard reference frameworks are NIST AI RMF (U.S.) and ISO/IEC 42001 (international).
- In practice - Buyers price uncontrolled AI usage as risk - hallucinations, customer-facing errors, regulatory exposure, IP leakage. The minimum viable governance is logged calls, weekly evals, written escalation rules, and a one-page AI policy aligned to one of the named frameworks.
Related reading
- AI evidence pack
- An AI evidence pack (XLev term) is the bundle of artefacts a buyer's diligence team needs to verify that AI claims are real: workflow inventory, eval results, governance policy, before/after process metrics, and wage-equivalent savings calculations with sources.
- In practice - Without the pack, AI claims get adjusted out of QofE. With it, they survive - because the evidence shape mirrors how diligence already verifies operational claims. We assemble this as a standard deliverable inside the Readiness Program.
Related reading
- Wage-equivalent savings
- Wage-equivalent savings (XLev term) is the recurring labour cost displaced by an AI or automation workflow, expressed at fully-loaded wage rates and netted of subscription and oversight cost. It's the EBITDA-impact view of AI ROI.
- In practice - Quoted as $/year, with a baseline (hours/month before), a current state (hours/month after), the wage rate used, and the net of platform cost. A buyer's QofE provider can verify each input - which is exactly why the format matters.
Related reading
Want the playbook, not just the definitions?
Start with a Diagnose. Two weeks. Written report. Honest fit assessment.
Send an enquiry