Skip to content

AI automation9 min read

Seven business processes you can automate with AI today

Seven repeatable processes where controlled automation can create capacity today. Each example names what the system can propose and where a person approves.

Reviewed by
Félix Tarnóy
Published
Last reviewed
Updated
A hand sorts through hanging file folders lined up in cardboard filing boxes, with blank, unlabeled folder tabs.

Suppose it is Monday morning: forty-two unread emails. Two of them are real enquiries. One arrived on Friday afternoon, and the sender has since been quoted by two other suppliers. The rest are delivery notices, copies of invoices, internal circulars and a newsletter nobody asked for.

By the time somebody has read, classified and distributed them, ninety minutes have gone. Tomorrow the same. This is not a technology problem. It is a capacity problem currently funded with human attention — and it is exactly the kind of work AI-based automation is genuinely good for today. High repetition, describable steps, low stakes on any single item.

The EU enterprise survey suggests this is not yet common practice in Hungary — not even among the firms large enough to be counted.

10.4%

In 2025, 10.4% of Hungarian enterprises with 10 or more persons employed used at least one artificial intelligence technology, compared with an EU average of 20.0%.

Enterprises with 10 or more persons employed, NACE Rev. 2 Sections C to N excluding Section K and Division 75, including Group 95.1. The financial sector, agriculture and mining are outside the scope.

Eurostat · 2025 · Source

Share of Hungarian enterprises with 10 or more persons employed that used at least one AI technology in 2025. EU average: 20.0%.

On its own the figure proves nothing, and it needs reading carefully: the EU survey covers only enterprises with 10 or more persons employed, while the great majority of Hungarian firms are smaller than that. What it does show is that this is not yet routine. Your first deployment will not have an internal precedent to copy — which is why it is worth starting with one detail of how the technology actually works.

The model executes nothing. It proposes.

This is the spine of the article, and it fits in a sentence. When a language model “handles” something — sends an email, produces a quotation, updates a record — it is not the model doing it. The model proposes a tool call; your application’s own code executes it and returns the result to the model.

Commercially, that distinction decides everything. The model does not decide whether a quotation may go out; your application decides, under rules you wrote. The model’s proposal is an input, not an instruction. It follows that the reliability of an automation is not a property of the model but of your integration — and three checks sit firmly on your side of the line.

  • Shape. Where a tool is defined with a strict schema, the vendor documentation states that the tool input strictly follows the schema you supplied and the tool name is always valid. That is shape and type, not correctness: the quantity will certainly be a number, but whether it is the right number is still for your code to check.
  • Permission. Which records the process may touch at all, and on whose behalf. At database level that is expressed as row level security policies, per table and in policy form — not in the instructions given to the model.
  • Log and approval. Every proposal, every executed step and every human approval leaves a trace. When a question comes up six months later about a quotation that went out, you can retrieve what the system proposed, who approved it and when.

Without those three layers you have not built an automation. You have built an unpredictable assistant with live permissions. Most of the risk in a deployment is settled here, not at the point where a tool is chosen.

Seven processes worth starting with

A process is a candidate when it runs often enough, behaves consistently enough, and can be measured as it stands. Automating a rare exception almost never pays back. These seven belong on one list because each one is repetitive and administrative, and in each case you can say precisely where the machine has to stop.

1. Enquiry handling and triage

Today somebody reads the email or the form, decides whose it is, records it, forwards it. How much of the day that consumes differs from company to company and grows with the volume; until you measure it, you are guessing. Automated, the system extracts the company, the subject, the deadline and the estimated value, proposes a category and opens a record for the person responsible. What stays human: accepting the classification, and the tone of the first reply.

2. Lead qualification

Today the salesperson ranks enquiries in their head, largely by who wrote most recently. Automated, the system assembles what your own records and public company data already know, compares it against the ideal customer profile, and returns a proposal with its reasoning: why it thinks this enquiry matters. What stays human: overriding the order — and insisting the reasoning is readable, rather than a bare score.

3. Document and quotation preparation

Today more than one person gathers the same figures from more than one system, and the quotation often takes days. Automated, the system reads the line items out of the request, matches them against your own price list and prepares a draft in your existing template. What stays human: approving the price, the lead time and the terms, and sending it. Sending is never automatic.

4. Scheduling

Today several emails go back and forth over one appointment and the diary still clashes. Automated, the system knows what capacity is free, where the job is, how long the travel takes and which qualification it calls for, offers two or three slots and books the one that is confirmed. What stays human: the exceptions. An emergency call-out, a client of twenty years and a Saturday visit are not diary questions.

5. Reporting

Today Monday’s management report often starts on Friday and consists largely of copying. Automated, the numbers arrive from the source systems and the system drafts the narrative: what changed, by how much, and which line is out of pattern. What stays human: the explanation and the decision. A report earns its place only when somebody adds why the outlier is there.

6. Customer support with escalation

Today the same two colleagues lose their day to repeat questions, while the difficult case queues behind them. Automated, the system answers common questions from your own documentation and hands the rest over — not once it is already stuck, but at a threshold you set in advance. What stays human: every complaint, every billing dispute, and every case where a customer has asked the same thing twice.

7. Internal handover between teams

Today sales hands the job to delivery and part of the information stays behind in an email thread. Automated, the handover has a precondition: the system checks that the order, the technical scope, the deadline and the contact are all present, then writes a summary for the receiving team. What stays human: confirming that the handover really is complete.

Where the process must stop, and what breaks when it does not
ProcessMandatory human checkpointWhat breaks without it
Enquiry triageAccepting the classification and the ownerThe largest enquiry lands with the wrong person and nobody notices, because the system was confidently wrong.
Lead qualificationOverriding the rankingThe score takes on a life of its own: whoever it ranked low is never called back, and the mistake is never discovered.
Quotation preparationApproving price, lead time and termsA badly priced quotation goes out in the company’s name, and the commitment is yours, not the vendor’s.
SchedulingChecking capacity and exceptionsDouble bookings, an engineer already on the road, and a slot nobody can actually honour.
ReportingChecking the source of the numbers and the outliersA management decision is taken on data whose error nobody looked at.
Customer supportThe escalation threshold and complaint handlingThe customer receives the same automated reply a third time, and a complaint becomes a public one.
Internal handoverConfirming the handover conditions are metAn incomplete handover slips through to the next team, and the error surfaces expensively at the end of the chain.

Free assessment

Does this sound familiar? Let us discuss the best place to begin.

Your assessment is carried out personally by one of our experts.

An illustrative calculation for one process

What follows is an illustrative scenario, not a measured result and not a client reference. It is here to show which numbers you need to collect before deploying anything. Take enquiry triage.

  1. Sixty enquiries arrive each week, by email and web form.
  2. Reading, classifying, recording and forwarding one enquiry by hand takes six minutes on average.
  3. Approving a prepared classification takes one minute.
  4. Eighty per cent of enquiries are routine; twenty per cent are exceptions that still have to be worked through by hand.

On those assumptions the current effort is 360 minutes a week, or six hours. In the automated version, 48 items approved at one minute each is 48 minutes, and 12 exceptions handled manually is 72 minutes: two hours in total. The difference is four hours a week, roughly seventeen hours a month.

Those four hours are only worth something if there is somewhere to put them. If the same colleague spends the same time in the inbox, only less tired, the calculation stayed on paper. And it only holds if the sixty and the six minutes are measured. If they are estimates, so is the answer. That is why every deployment starts with two weeks of measurement rather than a software decision.

What not to automate

Some processes have a technical answer of yes and a commercial answer of no. The rule is simple: the more a single error costs, the less sense it makes to take the decision away from a person.

  • Sending advertising email without consent. Advertising may be sent to a natural person by email only with that person’s prior, unambiguous and explicit consent, and the Hungarian regulator states that publicly available contact details are not themselves permission. An automated follow-up sequence crosses into that territory in moments. This describes the statutory text and is not legal advice: the actual model needs sign-off from a qualified lawyer.
  • Pricing and payment decisions. Discounts, extended terms, credit limits. The commitment binds you, not your vendor.
  • Complaints and terminations. Where the relationship is already damaged, an automated reply enlarges the damage rather than reducing it.
  • Employment decisions. Selection, appraisal, dismissal. Human judgement here is not an efficiency question.
  • Regulated professional responsibility. Medical, legal and financial advice; technical conformity statements.
  • Any process whose input may not leave your own systems. The problem there is not the process but the architecture: the data stays put and only the strictly necessary field travels.

Three documented limits to design around

The three behaviours below come from the vendor’s own documentation, as checked in September 2026. None of them is a malfunction: they are known properties of the technology, and the process design has to price them in.

  • A missing value may be filled in for you. When a required field cannot be derived from the input, the model may infer a plausible-looking value rather than ask. The documentation is explicit that this behaviour cannot be relied on and differs between models. On a quotation, that means an invented date can appear where the delivery deadline was missing.
  • Content arriving from outside is untrusted. The documentation instructs that content returned by tools — web pages, inbound email, uploaded files — must be treated as untrusted, because instructions can be hidden inside it to redirect the model. The computer use tool page adds that such precautions remain important even with its own classifier defence layer in place. This bears directly on the processes above: the inbound email and the uploaded document.
  • Error recovery has a ceiling. On an invalid tool call the documentation says the model retries roughly two to three times before apologising to the user. If your process does not hand over to a person at that point, the task is quietly lost — which is worse than a visible failure.

That same computer use tool page recommends that a human confirm any decision with meaningful real-world consequences and any task requiring affirmative consent, naming financial transactions and agreeing to terms. That is not timidity but a design principle: the approval point is not the opposite of automation but its precondition.

What to do in the next four weeks

  1. Pick one process

    The one that runs at least twenty times a week and that you already know is consuming somebody’s day. Choose the most frequent, not the most impressive.

  2. Measure it for two weeks

    Volume, average handling time, number of errors. Without that there is nothing to compare against afterwards, and the discussion will be about impressions rather than numbers.

  3. Mark where the machine stops

    Write down in one sentence what a person approves, and what happens when the approver is away. Without an approval point automation is not faster, only unauditable.

  4. Draw the data boundary

    Which fields may go to an external provider, which system stays closed, and who can read the log. That decision belongs before the deployment, not after it.

  5. Run it in parallel, then decide

    Let the old and the new process run side by side for two or three weeks. By the end of the second week the real difference is visible, and the next process can be chosen responsibly.

This sequence looks slower than simply buying a tool, and in the short term it is. In exchange, by the second process you have a pattern and by the third a routine. If it is useful, see how we approach AI automation, or request a free digital assessment: the process map and the priority list are the first step in it.

Sources

  1. Anthropic — How tool use works

    The model proposes a tool call; the application’s own code executes it and returns the result to the model.

    https://platform.claude.com/docs/en/agents-and-tools/tool-use/how-tool-use-worksLast reviewed: 1 September 2026

  2. Anthropic — Strict tool use

    With strict mode enabled, the documentation states that the tool input strictly follows the supplied input schema and the tool name is always valid.

    Applies to custom tool definitions; the computer use and browser use toolset entries do not accept the setting.

    https://platform.claude.com/docs/en/agents-and-tools/tool-use/strict-tool-useLast reviewed: 1 September 2026

  3. Anthropic — Handle tool calls

    Tool results often carry content from outside the application — web pages, inbound email, user uploads — which the documentation instructs must be treated as untrusted; on an invalid tool call the model retries roughly two to three times with corrections before apologising to the user.

    https://platform.claude.com/docs/en/agents-and-tools/tool-use/handle-tool-callsLast reviewed: 1 September 2026

  4. Anthropic — Tool use overview — when required parameters are missing

    When the prompt does not contain enough information to fill a required parameter, the model may ask for it, but it might also infer a reasonable value; the documentation states this behaviour cannot be relied on and varies by model.

    https://platform.claude.com/docs/en/agents-and-tools/tool-use/overviewLast reviewed: 1 September 2026

  5. Anthropic — Computer use tool — security considerations

    The documentation recommends asking a human to confirm decisions that might result in meaningful real-world consequences and any task requiring affirmative consent; it also advises against giving the model access to sensitive data such as account login information, and states that these precautions remain important even with its classifier defence layer in place.

    The guidance appears on the computer use tool page, not as a general tool-agnostic policy.

    https://platform.claude.com/docs/en/agents-and-tools/tool-use/computer-use-toolLast reviewed: 1 September 2026

  6. Anthropic — API and data retention

    The documentation states that conversation content — prompts and outputs — is not retained by default, and that retained data is never used for model training without the customer’s express permission.

    Designated Covered Models are an exception and require 30-day retention. Zero data retention is not the default: it is arranged per organisation under a separate agreement, and it does not extend to every product surface.

    https://platform.claude.com/docs/en/manage-claude/api-and-data-retentionLast reviewed: 1 September 2026

  7. Eurostat — Use of artificial intelligence in enterprises (isoc_eb_ai)

    In 2025, 10.4% of Hungarian enterprises with 10 or more persons employed used at least one artificial intelligence technology, compared with an EU average of 20.0%.

    Enterprises with 10 or more persons employed, NACE Rev. 2 Sections C to N excluding Section K and Division 75, including Group 95.1. The financial sector, agriculture and mining are outside the scope.

    https://ec.europa.eu/eurostat/databrowser/view/isoc_eb_ai/default/table?lang=enLast reviewed: 1 September 2026

  8. Supabase — Row Level Security

    PostgreSQL row level security defines, per table and in policy form, which rows a given role may access.

    https://supabase.com/docs/guides/database/postgres/row-level-securityLast reviewed: 1 September 2026

  9. European Commission — Legal grounds for processing personal data

    A controller must have — and document — a valid legal basis for processing personal data.

    https://commission.europa.eu/law/law-topic/data-protection/information-business-and-organisations/legal-grounds-processing-data_enLast reviewed: 1 September 2026

  10. Nemzeti Jogszabálytár — Act XLVIII of 2008 on the Basic Conditions and Certain Limitations of Business Advertising Activity

    Section 6(1) permits advertising to a natural person by email only with that person’s prior, unambiguous and explicit consent; Section 6(3) allows that consent to be withdrawn at any time, free of charge and without reasons.

    https://njt.jog.gov.hu/jogszabaly/2008-48-00-00Last reviewed: 1 September 2026

  11. National Media and Infocommunications Authority (NMHH) — Electronic advertising

    The NMHH states that publicly available contact details do not in themselves constitute permission to send advertising.

    https://nmhh.hu/cikk/4077/Az_elektronikus_hirdeteskuldes_szabalyaiLast reviewed: 1 September 2026

Free assessment

Four areas. One clear next step.

Choose what we should look at. Our experts review it and return a specific recommendation.

Insights