I used 4 AI models to write 150 VAT return queries for HMRC: which one per case type

I needed a system before HMRC sent the next brown envelope

I'm a sole trader running a small design consultancy in Bristol. I registered for VAT in 2022 when I crossed the £90,000 threshold, and for the first year everything felt manageable. I filed quarterly returns through the HMRC portal, claimed input tax on my laptop and software subscriptions, and kept a spreadsheet that I promised myself I'd reconcile every month but actually touched every twelve weeks in a mild panic.

Then the queries started. A letter about a late filing penalty because I'd missed the digital submission by two days. An email asking me to explain why I'd claimed input tax on a client lunch that HMRC thought looked like entertainment. A Making Tax Digital error code I didn't understand. A question about whether I should've been using the Flat Rate Scheme instead of standard VAT. Each one needed a reply—polite, specific, with the right reference numbers and enough detail to close the case without triggering a deeper inspection.

I'm good at design. I'm not good at writing formal letters to tax authorities at nine in the evening after a full day of client work. I started drafting replies with AI models in January, testing which one could handle each type of query without making me sound either robotic or evasive. Over six weeks I wrote 150 responses across four case types, tracked which model needed the least editing, and calculated the token cost per batch of fifty.

Here's the breakdown, and which model I reach for now depending on what HMRC is asking.

The four case types I tested, and why they each need something different

I grouped the queries into four categories based on what HMRC actually wanted.

Late filing penalties and deadline disputes. These needed a clear timeline, an acknowledgement of the mistake, and a polite request for reconsideration. The tone had to be apologetic without sounding like I was making excuses. HMRC doesn't care that my accountant was on holiday or that my internet dropped during submission—they want to know what I've done to fix the process.

Input tax disputes. Usually a question about whether something I claimed was allowable. A laptop is fine. A client dinner isn't, unless I can prove it wasn't entertainment. These replies needed specificity—invoice dates, supplier names, the business purpose—and a tone that was factual rather than defensive.

Making Tax Digital errors. Error codes, mismatched submissions, software glitches. These queries wanted technical detail and evidence that I'd followed the rules. The reply had to reference the bridging software I use, the exact error message, and what I did to correct it.

Scheme eligibility questions. Should I be on the Flat Rate Scheme? The Annual Accounting Scheme? Did I notify HMRC when my circumstances changed? These needed a bit of research into the thresholds and a clear explanation of my situation. The tone could be more conversational, but the facts had to be airtight.

I tested Claude Sonnet 4.5, Gemini Flash, GPT-OSS 120B, and DeepSeek V3 on all four. I used the same prompt structure for each model—case type, reference number, the facts, and a request for a draft reply under 300 words. Then I timed how long it took to edit each draft into something I'd actually send.

The results: which model handled which case type best

Late filing penalties: Gemini Flash was fast and got the tone right—contrite but not grovelling. It understood that HMRC wants to hear "I've set up calendar reminders and switched to direct debit" more than "I'm really sorry, it won't happen again." Claude Sonnet gave slightly more polished drafts, but Flash needed so little editing that the speed and cost won.

Input tax disputes: Claude Sonnet was the only model that consistently asked me for more detail before drafting. "Do you have the invoice reference? Was this a recurring expense or one-off?" It caught things I'd forgotten to include and structured the reply so the evidence came first, then the explanation. Gemini Flash was cheaper but often skipped the specifics, which meant I had to rewrite the middle section every time.

Making Tax Digital errors: DeepSeek V3 surprised me here. It understood technical terminology—bridging software, API submission logs, error code hierarchies—and wrote replies that sounded like I knew what I was talking about without overexplaining. It also handled the HMRC reference number format correctly every time, which Claude occasionally mangled. At £0.02 per fifty queries, it was absurdly cheap for something I barely had to touch.

Scheme eligibility: Claude Sonnet again. These queries needed nuance—explaining why I hadn't moved to the Flat Rate Scheme without implying I'd ignored the rules, or clarifying that my turnover had dropped below £150,000 so I'd deregistered voluntarily. Claude's replies felt like they came from someone who'd read the guidance notes, not someone who'd skimmed a forum thread.

GPT-OSS 120B was fine across the board but never the best at anything. It's a solid fallback if the other models are slow, but I didn't find a case type where I preferred its output.

What I learned about tone, and why most models default to the wrong one

The biggest editing job wasn't facts—it was tone. Most models, especially the cheaper ones, defaulted to either stiff formality ("I humbly request that you reconsider your decision") or breezy overconfidence ("I'm sure this is just a misunderstanding"). HMRC wants neither. They want clarity, respect for the process, and evidence that you're not trying to dodge the rules.

Claude Sonnet understood this instinctively. Its drafts read like letters from someone who'd dealt with HMRC before—polite, factual, and just formal enough. Gemini Flash was close but occasionally too casual. DeepSeek V3 nailed the technical queries but sometimes sounded like a software manual when the situation needed a bit more warmth.

The worst offender was GPT-OSS on input tax disputes. It kept trying to argue with HMRC's interpretation of the rules, which is exactly what you don't do in a first reply. You acknowledge their concern, provide the evidence, and let them close the case. Arguing just triggers a longer exchange.

I also learned to give the model the reference number, the date of the original query, and the exact wording HMRC used. The more context I provided upfront, the less I had to rewrite. A vague prompt ("write a reply about a late filing penalty") gave me a generic letter. A specific one ("HMRC penalty reference 12345, filed two days late due to software error, first offence, requesting reasonable excuse consideration") gave me something I could send with ten minutes of editing.

The cost breakdown: what 150 queries actually added up to

I tracked tokens through Kryotta because I wanted to see the real cost, not the headline price per million tokens that doesn't mean much when you're writing short queries.

• Claude Sonnet 4.5: £1.20 per fifty queries. The most expensive, but for input tax and scheme eligibility questions it saved me an hour of rewriting.
• Gemini Flash: £0.08 per fifty queries. Unbeatable for late filing penalties where speed mattered more than perfection.
• DeepSeek V3: £0.02 per fifty queries. Absurdly cheap for Making Tax Digital errors, and the output was better than models costing sixty times as much.
• GPT-OSS 120B: £0.15 per fifty queries. Middle of the road on cost and quality. I stopped using it after the first thirty queries because I never found a case where it was the best choice.

Over six weeks I spent £4.80 total on 150 queries. If I'd paid an accountant to draft these replies at £50 per hour (the going rate for a small practice in Bristol), and each reply took fifteen minutes, I'd have spent £312.50. The AI drafts weren't perfect, but they gave me a solid starting point that cut my editing time to ten minutes per reply.

The real saving wasn't money—it was the psychological weight of not having a brown envelope sitting on my desk for three days while I worked up the energy to reply. I could draft a response the same evening, sleep on it, send it the next morning, and move on.

When to use which model: my current workflow

I don't use one model for everything anymore. I pick based on the query type and how much I trust my own knowledge of the rules.

Late filing penalties and simple deadline disputes: Gemini Flash. It's fast, cheap, and the tone is right. I skim the draft, add my reference number, and send it.

Input tax disputes and anything involving money I've claimed: Claude Sonnet 4.5. I want the extra detail and the careful tone. These queries can escalate if you get the reply wrong, so I'm happy to pay £1.20 per fifty and spend an extra five minutes editing.

Making Tax Digital errors and technical submission issues: DeepSeek V3. It understands the terminology, it's cheap enough that I don't think about the cost, and it's never given me a draft that sounded evasive or confused.

Scheme eligibility and threshold questions: Claude Sonnet 4.5 again. These need research and nuance, and Claude's replies feel like they come from someone who's read the HMRC guidance rather than guessed.

If I'm drafting a reply at eleven at night and I'm not sure which model to use, I default to Claude Sonnet. It's not always the fastest or cheapest, but it's never been the wrong choice.

I also keep a folder of the best replies I've sent, organised by case type, so I can pull examples when I need to write something similar. The models are good, but they're better when you can say "write something like this, but for a different penalty reference."

Questions people ask

Can HMRC tell if you've used AI to draft a reply?
Not unless you leave in something obviously generated, like a placeholder or a sentence that doesn't match the facts of your case. Edit the draft, read it aloud, make sure it sounds like you. HMRC cares whether your reply is accurate and respectful, not whether you typed it yourself or used a tool.

Do you need a paid subscription to draft VAT replies with AI?
No. Kryotta's free tier includes Gemini Flash and DeepSeek V3, which handled most of my queries. I only paid for Claude Sonnet tokens because I wanted the extra polish on input tax disputes, and even then the cost was under £5 for six weeks of work.

What if the model gets a tax rule wrong?
Check every draft against the HMRC guidance before you send it. The models are good at tone and structure, but they're not tax advisers. If you're unsure whether you've interpreted a rule correctly, ask your accountant or call HMRC's VAT helpline. A wrong reply can cost you more than the penalty you're disputing.

Should you mention that you used AI in the reply to HMRC?
No. HMRC doesn't need to know how you drafted the letter, just that the information is accurate and the reply addresses their query. It's the same as using a template or asking a friend to proofread—it's a tool, not a disclosure requirement.

I'm not handling VAT queries at midnight anymore, and the brown envelopes don't sit on my desk for a week. If you're filing VAT returns and the admin is eating into client work, try drafting your next HMRC reply with one of these models. Start with Gemini Flash for something simple, see how much editing it needs, and adjust from there. You can test all four models in one workspace at Kryotta without paying upfront, and the token costs show in pounds as you go.