Why AI Goes Wrong
Most AI failures are not caused by bad technology. They come from how it is used, and from the gap between what a probabilistic model does and what accounting, and MTD compliance specifically, demands: reliability and material accuracy.
Scattergun adoption with no strategy. Staff experiment with different tools, generate drafts for random tasks, and spend hours iterating prompts without tying any of it to a defined goal. Alistair calls this the "throw it against the wall" approach, using AI "for anything and everything" with no coherent strategy behind it. For firms in the middle of an MTD rollout, there is no time to spare on experiments that never become repeatable workflows.
Using AI for tasks that need binary accuracy. Accounting is full of yes or no outputs: calculations, reconciliations, compliance checks, filing requirements. Sebastian Triff, a developer who works closely with how these systems behave in practice, notes that AI tends to be stronger at creative generation than strict correctness. Quarterly submissions to HMRC are not a place for a confident best guess. That is precisely why the filing itself belongs with dedicated, deterministic software, like TaxCalc's MTD Quarterly Filer, while AI stays useful for what surrounds it: client reminders, summaries, and communication.
Poor context and dirty inputs. LLMs do not understand documents the way people do. They generate responses based on patterns in the tokens they are given, so feeding in irrelevant or inconsistent information produces output that looks confident while being wrong. As Seb puts it, "dirty data is dirty context, and that throws the LLM off." A bookkeeper pasting a client's entire digital records into a chatbot to ask about MTD eligibility might get an answer that sounds authoritative and is still incorrect.
Overconfidence in agentic workflows. Modern AI can take actions in other apps: calendars, inboxes, shared drives. This is the agent layer, and it is not fully reliable yet, especially when given autonomy over real business systems. Seb's warning is blunt: "it's not very reliable yet... people try to let it go full autonomous... and the quality isn't there." For a firm managing clients through quarterly MTD deadlines, an unsupervised agent handling submissions is a reputational risk waiting to happen.
Misunderstanding data security. The moment client data goes into an AI tool, it matters what happens next: storage, retention, whether it is used for training, who can access it. Seb's reality check: "the moment you give feedback, the entire conversation is pulled into the training set... if you upload a document, that gets pulled in too." With MTD bringing more client financial data into digital workflows than ever, this is not a detail to skip past.
A Simple Framework for Picking the Right AI Use Cases
Before rolling AI out across your practice, run this exercise.
- List your firm's activities: summarising client queries, drafting emails, classifying inbox items, preparing routine proposals
- Assess frequency: how often does each one happen?
- Assess the cost of a mistake: reputational, financial, compliance, or staff impact if it goes wrong
- Plot these on a simple matrix, weighing frequency against cost of error
High frequency and low mistake cost is your best starting point. Low frequency and high mistake cost should wait. As Alistair frames it, "anything in the top right quadrant is what you focus on... you tend not to touch the bottom left as a quick win."
Run MTD work through the same lens and the picture is clear. Client communication about MTD, chasing records, answering "do I need to file yet" queries, sits firmly in the low risk, high frequency zone and is well suited to AI assistance. The actual quarterly submission sits in the high risk zone. That is a job for software built for the purpose, not a general AI model asked to have a go.
Being Deliberate About Adoption
Alistair's core point is that AI decisions should flow from the firm's strategy, not the other way round: "it comes back to the firm's vision. What are they trying to achieve, rather than using technology just for the sake of technology."
For a mid market firm navigating MTD, the vision is usually straightforward: get every eligible client compliant, on time, without burning out the team. Being deliberate about AI means asking where it genuinely reduces that load, drafting client updates, triaging inbox questions about MTD, summarising webinar content into FAQs, versus where it introduces risk you do not need, anything touching the actual calculation or submission.
That means having three things in place.
- A clear objective: faster client onboarding to MTD, less time spent on repetitive client queries, more capacity for advisory work
- A boundary: AI drafts and summarises, dedicated MTD software calculates and files
- A roadmap: how you move from small pilots, like AI-drafted client emails, to firm-wide repeatable workflows
Without these, AI becomes a red herring. Lots of activity, unclear outcomes, and the occasional embarrassment, at exactly the time firms can least afford it.
Getting Better Answers Out of an LLM
LLMs respond to what you feed them, so accuracy improves when your prompts come with the right context. Context is the information you give the model before asking your question. Because LLMs generate based on the tokens they have already seen, incomplete context increases the odds of a confident but wrong answer.
A few habits worth building in, particularly for MTD client communications drafted with AI:
- State the goal clearly, for example "draft a reminder for clients whose first MTD quarter ends 5 August"
- Provide relevant background only, the specific deadline or client segment, rather than pasting an entire HMRC guidance document
- Isolate the section that matters instead of feeding in everything at once
- Build in a review loop: ask the model what it assumed, ask what is missing, then verify against HMRC's actual requirements
Why the Same Prompt Does Not Give the Same Answer
AI is probabilistic. Even asked the same question twice, it may choose a different path to the answer. There is deterministic code underneath, but there is no such thing as a deterministic model output. LLMs predict likely next tokens with weighted randomness built in.
This is another reason MTD filing should not sit with a general AI tool. A quarterly submission needs to be the same, correct answer every time, not a plausible one. Software designed specifically for MTD for Income Tax handles that determinism by design. AI's value sits upstream and downstream of that, helping clients understand what is needed, and helping your team communicate it clearly.
Data Security and GDPR: Make It a Process, Not a Panic
Security is as much an operational question as a legal one. Check the privacy policy for any tool before you use it, and understand what happens to information once it is sent.
In practice, that means:
- Using approved tools and configurations
- Having clear data classification rules covering what is safe to paste and what must never be uploaded
- Preferring enterprise or controlled environments for sensitive client material
- Treating third-party AI adoption like any other system, with due diligence first and use after
With MTD pushing more client financial detail into digital tools, this discipline matters more, not less.
The Bottom Line
AI adoption fails when it is treated like magic: copy, paste, accept the output, trust the autonomy. It succeeds when it is treated as a system component, chosen deliberately, backed by good context, and constrained by human review.
For firms managing the shift to MTD for Income Tax, that means a clear split. Let AI help with the volume of client communication and admin around MTD. Let purpose built software handle the part that has to be right every single time.
Alistair's summary is the one worth remembering: AI should accelerate the journey to your firm's destination, not drive faster in the wrong direction. Get the right tasks, feed the right context, apply the right controls, and let the software built for the job do the job.
TaxCalc's MTD Quarterly Filer is built precisely for practices managing multiple clients at different stages of digital readiness, standardising submissions and keeping your workflows scalable as the second and third waves of MTD arrive. If you would like to see how it fits into your practice, take a free trial today.
Want to hear more from Alistair and Seb? Watch the full episode:
