Core sections
Sage Platinum Club Winner
20+ Years Experience
UK&I + North America
First UK Sage X3 Partner
Finance has always had to absorb technological change. What is different about AI and specifically about the generative AI wave of the past two to three years is the pace at which it has moved from experimentation to operational expectation. In 2023, most board conversations about AI were speculative. By 2025, mid-market CFOs are being asked not whether to adopt AI, but to explain their governance position on AI that is already in use across their organisations.
The challenge is that AI has arrived faster than the vocabulary, the governance models, and the regulatory frameworks that would normally accompany a technology of this significance. Finance professionals are expected to evaluate vendor claims, challenge consultants, and brief non-executive directors, often without a reliable foundation to work from.
This page is designed to close that gap. It covers the concepts, the risks, the distinctions that matter commercially, and the governance frameworks that turn AI from a liability into a controlled asset. It is written for CFOs, Finance Directors, and finance teams, not for technologists.
Mid-market boards are now requiring CFOs to present AI strategy, not explore it.
These three terms are used interchangeably in vendor marketing and board conversations. They are not the same thing. Understanding the difference is not an academic exercise — it determines the risk profile, governance requirements, and realistic expectations of what any given system can and cannot do.
| Automation | Machine Learning | Generative AI / LLMs | |
|---|---|---|---|
| How it works | Follows explicit, pre-programmed rules | Learns statistical patterns from training data | Predicts and generates text or content based on training |
| Output type | Deterministic — same input always produces same output | Statistical — outputs are probabilistic within a trained range | Probabilistic — outputs vary; no guaranteed correctness |
| Finance example | Invoice routed to approver when over £10,000 | Anomaly detected in supplier payment pattern | Period-end commentary drafted from financial data |
| Risk profile | Low — failure modes are predictable and defined | Medium — performance degrades if training data is stale or biased | Higher — hallucination risk; requires governance and verification |
| When it fails | Rule not defined for edge case | Training data gap or distribution shift | Model generates confident but incorrect output |
| Governance requirement | Workflow design and exception mapping | Model monitoring and data quality controls | Human-in-the-loop, audit trail, tiered approval |
Large language models are trained on enormous quantities of text — books, articles, websites, code repositories, and structured documents. Through that training process, the model learns to predict: given this sequence of words and context, what should come next? That prediction capability, applied at scale with billions of parameters, produces outputs that can appear remarkably coherent, knowledgeable, and creative.
It is important to understand what this is and is not. An LLM does not “know” things the way a database knows things. A database retrieves a stored fact. An LLM generates a plausible continuation of a prompt, informed by statistical patterns learned during training. The distinction matters because it means LLMs can — and regularly do — generate text that is fluent, confident, and wrong.
The model’s working memory is its context window: the text it can process at one time. Modern LLMs have large context windows — enough to hold a substantial financial report or a full contract — but they are not unlimited. When input exceeds the context window, earlier content is dropped silently. For long documents, this can mean the model answers based on partial information without flagging the gap.
Training data has a cutoff date. An LLM trained on data through a given point does not have access to anything published after that date unless it is provided within the prompt or via a retrieval mechanism. For finance professionals evaluating current regulatory guidance or market conditions, this matters.
“The right question is not ‘is this AI accurate?’ It is ‘what happens when it is not, and who is accountable?'”
| ✕ Myth | ✓ Reality |
|---|---|
| LLMs are a database — they store and retrieve facts accurately | LLMs generate plausible text based on statistical patterns. They do not verify facts against a source of truth. |
| AI is always improving with use | Standard LLMs do not learn from your queries. Their knowledge is fixed at training cutoff unless updated or augmented. |
| Confident output means correct output | LLMs express uncertainty poorly. A hallucinated figure can appear with the same fluency as a correct one. |
| AI understands your documents | AI processes text tokens. It does not have contextual understanding of your business, your chart of accounts, or your regulatory obligations unless that context is provided explicitly. |
| AI can replace human judgement on material decisions | AI can surface options, draft outputs, and process volume. Material decisions — ones that affect financial statements, compliance, or stakeholder commitments — require human authority. |
A large language model trained on general data does not know your chart of accounts, your supplier base, your intercompany structure, or your latest board pack. It can generate fluent, plausible-sounding responses about these things, but they will be fabricated.
Retrieval-Augmented Generation solves this problem. Before generating a response, a RAG system retrieves relevant, current information from a specified source — your ERP, your document management system, your knowledge base — and provides that information to the model as context. The model then generates its response grounded in your actual data.
The practical implication: a RAG-enabled AI assistant querying your Sage Intacct data answers questions about your accounts, not about accounts in general. It can surface your specific aged debtors, your supplier payment terms, your actual cash position — because it is working from retrieved facts, not trained approximations.
Why it matters for finance: RAG is the mechanism that makes AI answers trustworthy enough to use in operational finance contexts. Without it, AI is a general knowledge tool. With it, AI can act as an informed assistant with access to your specific financial data. Verify whether any AI system you evaluate uses RAG — and ask what data sources it retrieves from and whether retrieval is auditable.
Model Context Protocol is an open standard, developed by Anthropic and adopted across the AI industry, that defines how AI models connect to external tools, data sources, and systems. It is, in essence, the governance infrastructure for AI integration.
Before MCP, AI integrations were bespoke — each connection between an AI model and an enterprise system required custom engineering with inconsistent security and permission models. MCP standardises those connections: it defines what data the AI can access, what actions it can take, what permissions are required, and how those boundaries are enforced.
For finance and IT leaders, MCP matters because it makes AI integrations auditable and governable. An AI system operating through MCP connections has defined, inspectable permissions — you can see exactly what it can and cannot access. An AI system operating outside a standards-based protocol is harder to audit and harder to govern.
Why it matters for finance: When evaluating AI systems for finance use, ask whether the integration with your ERP and financial data uses a standards-based protocol. Bespoke, undocumented integrations create governance and audit risks that are difficult to quantify. MCP-compliant integrations can be inspected, restricted, and documented as part of your AI governance framework.
An AI agent is an AI system capable of executing a sequence of actions to complete a defined goal — not just generating a single response. Where a conversational AI answers one question at a time, an AI agent can plan a workflow, execute multiple steps, call external tools, evaluate results, and continue until the goal is achieved.
In a finance context, an agentic AP workflow might: retrieve open purchase orders, match against incoming invoices, identify discrepancies, calculate payment proposals within defined approval thresholds, route exceptions to the appropriate reviewer, and post approved transactions to the general ledger — as a single governed, auditable process.
The capabilities are significant. So are the risks. An AI agent that can take actions — post transactions, send communications, approve payments — requires a governance framework that is qualitatively different from one applied to a conversational AI that generates text.
Why it matters for finance: Agentic AI represents the next material step in finance automation. At Mysoft, this is Stage 4 of our AI maturity framework — and we do not deploy agentic capabilities until a client has the governance foundation, data quality, and operational discipline that Stages 1 to 3 establish. The technology is ready; the governance readiness of most finance functions is not yet. That gap is where implementation risk lives.
The term “hallucination” in AI refers to a specific failure mode: an AI model generates output that is factually incorrect, invented, or unsupported — and presents it with apparent confidence. There is no visible uncertainty, no caveat, no signal that the output should be verified. The model simply predicts the next plausible token, and sometimes that prediction is wrong.
Hallucinations arise from the fundamental architecture of large language models. Because LLMs are statistical prediction engines rather than fact retrieval systems, they do not “know” when they do not know something. A model asked about a regulatory requirement it has not seen in its training data will not say “I don’t have that information.” It will generate a plausible-sounding response — and that response may bear no relationship to the actual requirement.
For finance professionals, the implications are not abstract. Finance involves material figures, regulatory obligations, contractual commitments, and statements that carry legal weight. An AI-generated figure in a board report, a fabricated regulatory citation in a compliance document, or an incorrect intercompany balance in a reconciliation are not theoretical risks — they are the specific failure modes that arise when AI is deployed without verification structures.
The problem is compounded by two behaviours that research consistently identifies in human-AI interaction. The first is over-reliance: users defer to AI output without checking it, particularly when the AI presents its answer with apparent confidence. The second is automation bias: humans given an AI-generated recommendation are less likely to identify errors than humans working without AI assistance. Both behaviours are well-documented and both are design problems, not user problems.
Risk: Model generates a confident summary based on general training data. Specific exemptions applicable to the entity are not referenced. Tax position filed incorrectly.
Risk: A supplier bank account has been updated fraudulently. The agent has no mechanism to flag account changes. Payment authorised and processed.
Human-in-the-loop by design. Define which AI actions require human review before execution. Tiered approval models calibrate oversight to risk level — not every action requires CFO sign-off, but every material action requires a human checkpoint.
Audit trail by default. Every AI-executed action should generate a timestamped, tamper-evident record. The audit trail should capture what the AI did, on what data, at what time, and what human action followed. This is the minimum requirement for regulatory defensibility.
Verification checkpoints. AI-generated figures in financial statements, board reports, or regulatory filings should be subject to the same verification process as any other input. The AI is an input mechanism, not a verification mechanism.
Exception escalation. Governance frameworks should define what constitutes an exception — transactions outside normal parameters, anomalous patterns, confidence thresholds — and route exceptions to human review automatically. AI handles volume; humans govern exceptions.
The AI landscape available to a finance team today includes both general-purpose AI tools and AI capabilities embedded within finance systems. These are not interchangeable. Understanding the difference is a prerequisite for making defensible decisions about what AI your organisation should use, where, and under what governance.
This is not a case for or against any specific tool. General-purpose AI has legitimate uses in a finance context — research, drafting, learning, ideation. The risk arises when general-purpose tools are used for tasks that require access to verified financial data, auditability, or compliance alignment — tasks they were not designed to perform.
| Dimension | Generic AI Tools (e.g. ChatGPT, Copilot) | Embedded Finance AI (e.g. Sage Intacct) |
|---|---|---|
| Data access | Works with data you paste or upload to the prompt. No direct ERP integration. Data leaves your controlled environment. | Integrated within your ERP. Works with your live financial data within the system's permission and security framework. |
| Data security | Data submitted to general-purpose AI may be used to train future models (depending on provider and plan). Review data processing agreements carefully. | Data remains within the ERP environment. Governed by the same data handling policies as your financial system. |
| Auditability | Conversations are not part of your financial audit trail. AI-generated outputs used in financial processes are not traceable within your systems of record. | AI actions are logged within the ERP audit trail. Timestamped, permission-controlled, available to auditors in a structured format. |
| ERP integration | None. Output must be manually transferred to your financial systems. | Native. AI-generated entries, approvals, and workflow actions are executed within the ERP and reflected in your financial data immediately. |
| Compliance alignment | Not designed for accounting standards, VAT rules, or financial reporting obligations. No guarantee of regulatory currency. | Developed and maintained by software providers with regulatory obligations and customer commitments. Updated as requirements change. |
| Accountability | No defined accountability for AI output that enters your financial processes. Accountability lies entirely with the person who used the tool and applied the output. | AI actions are attributed to defined system users. Governance frameworks can assign accountability at every workflow stage. |
The right question is not which tool is better — it is which tool is appropriate for which task. A finance team using a general-purpose AI to draft a conference agenda, research an accounting standard, or prepare talking points for a board meeting is operating within reasonable risk parameters. The same team using a general-purpose AI to generate journal entries, draft statutory accounts, or process payment runs has a governance gap that no technology preference resolves.
Map your AI use cases before selecting your tools. Define which tasks require ERP integration and audit trail. Use embedded AI for those tasks. Use general-purpose AI for tasks where the output is informational and will be verified before it enters your systems of record.
Sage describes its approach to AI as “Authentic Intelligence” — a deliberate framing that positions AI not as a demonstration of technological sophistication, but as a tool that should serve genuine business outcomes in a trustworthy way.
The Authentic Intelligence philosophy has three components. First, AI should be purposeful — solving real problems for real users, not deployed because AI is available. Second, it should be transparent — users should understand what the AI is doing, what data it is working with, and how conclusions are reached. Third, it should be accountable — there should be a clear governance chain for every AI action, from configuration through to audit.
For finance leaders evaluating AI in their ERP, the Authentic Intelligence commitment is a useful benchmark. Ask any AI vendor the same three questions: Is this AI purposeful? Is it transparent? Is it accountable?
The Authentic Intelligence philosophy and the AI Trust label provide the platform-level foundation. The governance framework that Mysoft builds around Sage Intacct deployments extends and operationalises that foundation at the customer level — adding the process governance maps, tiered approval structures, audit trail protocols, and quarterly review cycles that translate platform commitments into operational practice.
The platform is trustworthy by design. The deployment determines whether that trust is realised in practice.
Most AI conversations in finance focus on capability. The more important conversation is about governance: the controls, accountability structures, and review cycles that determine whether AI creates value or creates risk.
Mysoft’s AI Governance Framework is proprietary IP developed for Sage Intacct deployments at Stages 3 and 4 of our AI maturity ladder. It has been built to meet the standards that CFOs, Finance Directors, external auditors, and regulators will increasingly require of organisations using AI in their finance operations. It is not an afterthought to deployment — it is the design foundation on which deployment is built.
The framework is built on five non-negotiable principles and delivered through four governance layers.
| Layer | Focus | What it delivers |
|---|---|---|
| 1. Process Governance | Workflow mapping and AI action classification | Every process mapped as: routine low-risk (fully automated), routine elevated value (automated + notification), or exception/anomaly (escalated to human review). |
| 2. Authority & Approval | Tiered Approval Model | Four tiers of human oversight calibrated to transaction value and risk: Tier 1 fully automated through to Tier 4 requiring CFO/FD sign-off. |
| 3. Audit Trail & Control | Complete action logging | Timestamped record of every AI action within Sage Intacct. CFO/FD have full read access. External auditors have structured access during audit periods. |
| 4. Model Review Cycle | Quarterly AI Optimisation Review (QAOR) | Regular calibration of exception rates, assessment of new Intacct AI capabilities, audit trail review, and next-stage roadmap alignment. |
The AI vendor landscape is crowded and claims are rarely backed by specifics. Finance leaders evaluating AI solutions need a consistent framework for separating substantive capability from marketing positioning.
These ten questions are designed to expose the gaps. A vendor that cannot answer them clearly has not built governance into their product. A vendor that answers them in detail has.
1
2
3
4
5
6
7
This is the question that reveals whether the vendor thinks of governance as a compliance checkbox or as an ongoing advisory responsibility. The answer should describe specific deliverables: governance documentation, audit trail formatting, board reporting templates.
8
9
10
This is the question that reveals whether the vendor thinks of governance as a compliance checkbox or as an ongoing advisory responsibility. The answer should describe specific deliverables: governance documentation, audit trail formatting, board reporting templates.
A manufacturing business processing 2,400 supplier invoices per month has moved from a paper-based approval process to AP automation within Sage Intacct. Invoices received digitally are matched against purchase orders automatically. Three-way matching — invoice, PO, and goods receipt — is completed without manual intervention for invoices within defined tolerance thresholds.
Exceptions — invoices with no matching PO, variances outside tolerance, or new supplier accounts — are routed to a human review queue with full context attached. The finance team’s time moves from data entry to exception review. Processing time falls from 14 days to 3 days. The audit trail captures every action: automated match, exception flag, human decision, and posting. No AI hallucination risk applies — this is rule-based automation operating within defined parameters.
A professional services firm using Sage Intacct Copilot has integrated AI assistance into its period-end close process. At month end, the finance manager queries the AI assistant in natural language: “Summarise the material variances in operating expenditure against budget this month and identify the top three drivers.”
The AI retrieves the relevant data from Intacct, identifies the variances, and drafts a structured commentary. The finance manager reviews, adjusts one figure, and incorporates the commentary into the board pack in under 15 minutes, rather than 2 hours. The Intacct audit trail records the query, the AI response, the human edit, and the final version. The process is faster and the human remains accountable for the output.
Learn more about our AI maturity framework and what each stage of maturity mean.
Core vocabulary, concepts, risks, and governance foundations.
A 13-question diagnostic across five dimensions. 10 minutes. No registration required.
Four articles on how finance functions scale and build the operational foundation for growth.
Use-case focused articles on AP automation, AI-assisted close, FP&A, and agentic workflows.
A conversation about where your organisation is and what a structured path forward looks like.
