78% of billable time in a mid-sized tax advisory firm is spent on tasks that a skilled professional could delegate: classifying documents, reconciling amounts, cross-referencing accounting entries with tax returns, formatting settlements or searching for criteria from the tax authority across thousands of binding rulings. Artificial intelligence will not replace the tax advisor; it can, however, take over that 78%, allowing the advisor to focus on strategic analysis, audit defence and client counsel. This article explains which tools work today, with what real results, and how to avoid the most common mistakes when deploying AI in a tax firm.
The real problem facing tax advisory firms in 2025-2026
Regulatory pressure on tax advisory firms has not stopped growing. The Verifactu obligation (Regulation approved by Royal Decree 1007/2023, with software production deadlines throughout 2025 and 2026), the extension of the Immediate Supply of Information (SII) to new obliged parties, the changes introduced by Law 13/2023 to the General Tax Law to adapt the regulatory framework to the digital economy, or the novelties of the Multilateral BEPS Convention for groups with cross-border presence: every year the volume of regulation a firm must monitor multiplies.
At the same time, the war for talent makes hiring experienced tax technicians increasingly difficult and expensive. The result is a squeeze: more technical work and fewer available hours. The advisory firms that will thrive are those that convert that repetitive burden into automated processes, freeing the team for the tasks that truly justify their fees.
AI use cases that already work in tax firms
1. Automatic data extraction from invoices and receipts
Computer vision systems trained on invoices (a combination of classic OCR with visual language models) achieve reliability above 95% for fields such as issuer, VAT number, date, taxable base, VAT rate, tax amount and invoice number, including poorly structured formats such as cash receipts or low-resolution scanned invoices. The relevant figure for a firm: if a technician processes between 40 and 80 invoices per hour manually, a well-calibrated system processes several thousand without intervention. The human only reviews exceptions (a percentage that varies according to document quality, but typically below 8-10%).
This is especially useful for client portfolios with a high volume of transactions: distributors, hospitality businesses, retail. Integration with the firm's ERP or accounting software (a3ERP, Sage, Holded, Odoo) is the critical point; without it, the time saved turns into double the manual import work.
2. Copilot for the tax technical team
A copilot for advisory firms is not a generic chatbot. It is an assistant connected to the information sources the firm uses: the AEAT binding rulings database, the Official State Gazette, doctrine from the Central Economic-Administrative Court (TEAC), Supreme Court rulings and regional legislation where devolved taxes apply. The technician asks the question in natural language («Does the accelerated depreciation in article 12.3 of the Corporate Tax Law apply to company electric vehicles in 2026?») and the system returns the applicable criterion with the exact source citation.
The result is not a magic answer: it is a documented first proposal that the technician reviews and validates before signing off. The saving lies in the search and synthesis time, which for complex queries can exceed one hour. With a well-configured copilot, that same work can be resolved in five to ten minutes of critical review.
To see how we deploy this type of assistant for professional firms, visit our Copilot for tax and legal firms service.
3. Automatic document classification and filing
An advisory firm with 200 active clients generates tens of thousands of documents per year: contracts, notarial powers of attorney, deeds, settlements, inspection records, AEAT requests, responses to requests, filed returns. Classifying and archiving that volume coherently — so that any technician can find what they need in seconds — is a task that in many firms is left to each employee's discretion, with heterogeneous results.
AI-based document classification systems read the document, identify its type, the client it belongs to, the tax year and its status (draft, filed, notified) and file it in the correct folder. What previously required human judgement and discipline is now deterministic and auditable.
4. Personalised regulatory alerts by client portfolio
Not every update in the Official State Gazette or the Official Journal of the EU is relevant to every client. An AI-powered regulatory monitoring system can filter each official publication and cross-reference it against each client's profile (tax regime, sector, size, region, applicable taxes). The firm receives a daily or weekly summary containing only the updates that affect its specific portfolio, rather than having to read the full Gazette.
This capability is especially valuable for changes to regional personal income tax deductions, modifications to the reduced VAT rate in certain sectors, or developments in the Business Activities Tax. The risk of missing a development that affects a client is significantly reduced.
5. Review of contracts and clauses with tax implications
Lease agreements, related-party transactions, international distribution contracts or shareholder agreements carry tax implications that are not always detected before signing. Specialised language models can review contract drafts in minutes, flagging clauses with tax risk (transfer pricing, withholding-at-source clauses, image rights, taxation of capital gains in call options) and providing a first analysis to the technician, who integrates it into the client advisory.
This capability connects with the AI-powered contract review service, which can be implemented independently or integrated into the firm's workflow.
Comparison table: manual tax task vs. AI-assisted
| Task | Estimated manual time | Time with AI | AI accuracy (market reference) | Human intervention required |
|---|---|---|---|---|
| Data extraction from 500 invoices | 6-12 hours | 15-30 minutes + exception review | >95% on structured fields | Review of the 5-10% with issues |
| DGT criterion query + report drafting | 45-90 minutes | 5-15 minutes of critical review | High if the copilot is connected to official sources | Validation and sign-off always by the technician |
| Classification and filing of 1,000 documents | 8-16 hours | <1 supervised hour | 90-97% depending on document quality | Review of rejections and ambiguities |
| Tax review of a 20-page contract | 1-3 hours | 20-40 minutes (including technician review) | Detects key clauses; does not replace legal judgement | Technician validates and completes the analysis |
| Weekly regulatory monitoring (Official Gazette + OJEU) | 2-4 hours/week per technician | Reading filtered summary: 15-30 min | Depends on the filtering model and client profile | Technician decides what to communicate to the client |
Note: time ranges and accuracy figures are market estimates based on documented deployments in the sector. They vary depending on document quality, software used and the level of system customisation.
What AI cannot do (and firms need to know)
AI makes mistakes. Language models can generate plausible but incorrect statements (what is called «hallucination»). In a tax context, an erroneous regulatory reference or a wrongly calculated amount has real consequences: surcharges, penalties, loss of client trust. That is why AI in a tax firm only works well if a human always reviews the output before it reaches the client or the tax authority.
Safe use cases are those in which the AI output is an intermediate proposal (a draft, a classification, an alert) that the technician validates before acting. Dangerous ones are those that place AI directly in front of the client or the administration without human review. No responsible system omits that step.
Nor can AI replace the trust relationship with the client, the negotiation with the inspector during a regularisation, or the strategic judgement on how to structure a complex transaction. That remains exclusively human territory.
Applicable regulation: GDPR, AI Act and professional secrecy
Before deploying any AI tool that processes client data, the firm must resolve three legal questions:
- GDPR (EU Regulation 2016/679) and LOPDGDD (Organic Law 3/2018): individuals' tax data is personal data. Processing requires a legal basis (typically, execution of the advisory contract or legitimate interest), and if a cloud-based AI provider is used, it is mandatory to sign a data processing agreement guaranteeing that data is not used to train the provider's models.
- AI Act (EU Regulation 2024/1689, in force since August 2024): AI systems that assist in decision-making with legal effect on individuals (settlements, appeals, tax planning) may be subject to transparency and human oversight obligations. Firms deploying these solutions in 2026 must verify the risk category of the system with their technology provider.
- Professional secrecy: Article 93.5 of Law 58/2003, the General Tax Law, and the deontological rules of the colleges of economists and administrative managers establish the duty of confidentiality. Sending client data to cloud-based AI tools without adequate contractual guarantees may compromise that duty.
The solution for advisory firms handling sensitive data frequently involves models deployed on their own infrastructure or on European servers with no-training agreements, which is known as sovereign AI. This approach is more costly but eliminates confidentiality risks.
Practical roadmap for a tax advisory firm of 5-25 people
It makes no sense to deploy everything at once. A sensible sequence for a mid-sized firm:
- Month 1-2 — Invoice extraction: this is the use case with the highest volume of hours saved and the lowest risk, because the technician always reviews the amounts before posting. Results visible within weeks.
- Month 3-4 — Regulatory copilot: connect the assistant to official sources (AEAT, Official Gazette, TEAC) and train the team on how to ask questions and validate answers. The cultural shift is more important than the technology.
- Month 5-6 — Document classification: review the firm's folder structure, define categories and launch automatic classification for new documents. Migrate the historical archive only if there is time and budget.
- Month 7 onwards — Regulatory alerts and contract review: once the team is confident in the tools, add the more advanced modules.
At Summum IA we accompany this process from the initial diagnosis through to go-live and team training. We have been helping SMEs and professional firms adopt technology without unnecessary friction since 2007, with five offices across Castilla y León and the Canary Islands.
Frequently asked questions
Can AI file tax returns autonomously?
No, and it should not. Filing returns with the tax authority requires a digital certificate or Cl@ve, and — more importantly — the signature of a professional responsible for the content. AI can prepare the draft, cross-check the data and flag anomalies; the technician reviews, validates and files. Automating the final step without human review is a risk no responsible firm should take.
What happens if AI makes an error in a tax calculation?
Liability remains with the advisor who signs off the work. That is why the correct deployment model is always «human in the loop»: AI proposes, the technician verifies and decides. Contracts with AI providers typically do not include liability for errors in the generated content; reviewing those clauses before signing is essential.
Is it expensive to deploy AI in a small tax firm?
Cost varies greatly depending on the use case and the provider. Invoice extraction tools have per-document pricing models (market ranges between €0.005 and €0.05 per document), making them accessible for small firms. A copilot connected to legal sources may have a monthly subscription cost similar to that of practice management software. What does require investment is integration with the firm's existing systems and team training; underestimating that part is the most common mistake.
What is the difference between using ChatGPT and a copilot built for advisory firms?
ChatGPT (or any general-purpose language model in its public version) does not have access to AEAT binding rulings, the Official Gazette in real time, or up-to-date TEAC or Supreme Court doctrine. Nor does it guarantee that the data you enter will not be used to train future models. A copilot for advisory firms is configured with access to those sources, with contractually guaranteed privacy policies and with specific instructions for the Spanish tax context. The difference in practical usefulness and legal certainty is substantial.