Adrian Capraru, Head of AI & AutomationMost AI projects in finance die between the demo and the audit. Adrian builds the ones that survive both.Most AI projects in finance die between the demo and the audit. Adrian builds the ones that survive both.

Asked now in today's best-paid AI roles Rising to 2030 where demand and budgets grow next

1. Sht 01
   Asked now

   AI service lines, built from zero

   An Irish chartered accountancy firm offered no AI when Adrian joined. A year later it has its own client platform, a trained team, an AI Academy, published fee packages and automation clients of its own in Ireland, the UK and the US.

   Proof

   Client platform live in 10 days, 20 modules, fee packages published

   You keep

   An AI practice with its own offer, team and pricing

2. Sht 02
   Asked now

   Finance automation that moves real money

   Bills read and posted to Xero, deal-to-invoice-to-payment across three legal entities and four currencies, month-end journals matched 107 of 107 against the bookkeeper. Every write carries an approval, an audit trail and an idempotency key.

   Proof

   About 52 invoices a month under signed approvals, live since April 2026

   You keep

   Live workflows in your ledger, with approvals and a runbook

3. Sht 03
   Rising to 2030

   Governed autonomy, safe by design

   Controls live in the software, not in a policy document. Agents read freely, draft for review and write to a ledger only after a human approves. EU-only AI processing enforced in code, an AML gate before any client account opens. The same discipline that took a FinTech through its EU licence audit.

   Proof

   Romania's 4th ECSP crowdfunding licence, with Adrian as lead on the audit

   You keep

   Controls a regulator or auditor can read and test

4. Sht 04
   Rising to 2030

   Evals, observability and audit trails

   Nothing ships on a good demo. Golden test sets, mutation-tested guards and regression gates decide when a release goes out, and every action leaves a logged, idempotent trail, so a retry never doubles an invoice.

   Proof

   3,000+ checks gate one platform; 120+ checkers and about 260K assertions gate another

   You keep

   The eval suite and the logs, so quality stays measurable after handover

5. Sht 05
   Rising to 2030

   Security for agents and client data

   Hard tenant walls and nothing exposed that a job does not need. Postgres row-level security, masked PPS and IBAN numbers, CSP and HSTS, a retention purge, data processing agreements raised early and sign-in tested before anyone logs in.

   Proof

   52 of 52 sign-in tests and 5 of 5 row-level security mutation tests on a Dutch pilot design

   You keep

   A security model backed by tests, not promises

6. Sht 06
   Rising to 2030

   Built for machine customers

   AI assistants are starting to buy on behalf of people. Adrian shipped one of the first manufacturing MCP servers, so ChatGPT and Claude can quote real parts: 4 tools, keyless, rate-limited, prices calibrated on measured parts.

   Proof

   Listed in the official MCP Registry, 18 Oct 2026

   You keep

   A storefront that AI agents can read, quote and buy from

7. Sht 07
   Asked now

   Data foundations agents can trust

   An agent is only as good as the records under it. Client lists rebuilt, ledgers reconciled against the books the business already trusts and duplicates made impossible, before any agent is allowed to act on them.

   Proof

   587 clients rebuilt in 3,172 writes with 0 failures; a Xero audit joined 749 invoices and surfaced 391 unbilled clients

   You keep

   Clean, reconciled data the next system can build on

8. Sht 08
   Asked now

   Adoption that sticks

   Software changes nothing until people use it. Teams trained with live demos, an AI Academy, per-person dashboards and a daily staff brief, so the workflow changes, not just the tooling.

   Proof

   AI for Accountants session for the full team (10+ people); Academy of 8 lessons and 48 prompts

   You keep

   A trained team and the material to train the next hire

9. Sht 09
   Rising to 2030

   Built to hand over

   Fixed scope with acceptance tests, then a handover where the client owns the code, the data and the runbooks from day one. No lock-in, no black box and nobody waiting on the builder to come back.

   Proof

   Every engagement closes with a runbook, a deploy procedure and the full regression suite

   You keep

   Code, data, accounts and runbooks in your name

10. Sht 10
    Asked now

    Operations run from 0 to 100

    The builder has also run the operation. A factory founded with an EU start-up grant and five 3D printers, grown to 200+ 3D printers with a hired and trained team, SOPs and ISO 9001:2015 certification, selling into 26 of the 27 EU member states. Before that, COO of a regulated FinTech through its licence.

    Proof

    x65 turnover in three years at a 57% net margin, no outside equity

    You keep

    An operator who knows what the system has to survive on a Monday morning

1. Rev A

   Understand the process first

   Work starts inside the team that does it: their spreadsheets, their month-end, their edge cases. The process is redesigned before anything is automated, and scope is fixed with acceptance tests.

2. Rev B

   Ship behind gates

   Small releases behind evals and regression gates, a rehearsal on demo data, then a dry run on live data that changes nothing. One dry run turned 587 would-be payroll tasks into exactly 38 before a single one was created.

3. Rev C

   Hand over the keys

   Code, data and accounts sit with the client from day one. Runbooks, tests and a trained team stay behind, so the system outlives the engagement.

Numbers computed, words generated

Figures come from deterministic code; AI writes the words around them. A fee, a VAT amount or a filing deadline is never a model guess.

People approve what needs judgement

Agents act; people sign off wherever money, customers or a client ledger are involved. Drafts not postings, signed approval links, confirm-first assistant actions.

Evals and regression gates

Golden test sets, mutation-tested guards and deploy smoke tests. 3,000+ checks gate one platform; 120+ checkers and about 260K assertions gate another.

A trail an auditor can follow

Every action is logged and idempotent, so a retry never doubles an invoice. Failed jobs land in a queue a person reviews, never in silence.