There is a common pitch for AI in professional services: replace the accountant, the lawyer, the specialist. In regulated markets it rarely survives first contact with a buyer. The buyer is not only paying for the output. They are paying for someone to stand behind the output. Remove that person and you have removed what they were buying.
Three of the companies we have built made the opposite choice. They put a qualified human at a specific, designed point in the workflow and built the AI around making that person's review fast, focused and well evidenced.
FinanceCrew: a chartered accountant signs every period
FinanceCrew is a Singapore-first AI finance platform for SMEs and multi-entity groups. Specialist agents handle the operating sequence: source-document intake, ledger posting, reconciliation, adjustments, tax, and group consolidation across currencies and entities. The output is not "the AI's accounts". Every completed period is reviewed by a Chartered Accountant before sign-off.
The automation carries the repetitive work. Qualified judgement sits where financial statements are approved. That is what lets the product expand professional capacity without removing professional accountability.
LawCrew: the boundary is part of the design
LawCrew is a development-stage legal-technology platform that prepares drafts of legal documents under Singapore law. It is explicit about what it is not: it is not a law firm, and every output is a draft for a suitably qualified lawyer to review before anyone relies on it.
What makes that boundary workable is what happens before the human sees a draft. Specialist agents draft, critique their own work and then face adversarial review. Then six deterministic checks test each draft: citations, jurisdiction, personal data, advice language, risk and conflicts. Matters the automated checks cannot clear go to an independently engaged practising lawyer. The human reviews fewer, better-prepared documents, with an audit trail showing how each one moved through the workflow.
WaterDoctor: limits, then autonomy
At WaterDoctor, the agents make real treatment decisions on biofilm reactors, with engineer-defined floors and ceilings. In week six of the build, the system tuned aeration too aggressively. The guardrails caught it. Afterwards, any change of more than 15% from the rolling baseline required human approval. Engineers sign every report that goes to a regulator.
The design pattern underneath
Across all three, the same structure holds:
- Narrow agents, each with one job. Broad "do everything" agents are hard to evaluate and harder to trust.
- Deterministic checks before human review. Anything a rule can verify, a rule verifies. The human's attention goes only where judgement is needed.
- One named human, one final approval. The approval is recorded, and it belongs to a person, not a team.
- An audit trail the professional can defend. The reviewer can see what the system did, what it checked and why it escalated.
Why this is also good economics
There is a trap on the other side. If the human must carefully re-check everything the AI produces, the product can cost more than the work it replaces. Checking output is expensive, and it is where many agent projects stall. The answer is not to remove the human. It is to shrink and sharpen the review: automated checks catch the mechanical errors, escalation rules route the hard cases, and the professional spends their time on the decisions only they can make.
We don't sell "an AI that replaces the professional". We build companies where one professional can stand behind far more work.
If you have a professional workflow where trust depends on a qualified person, and you can see how much of it is repeatable, that is the kind of company we like to build. Start with the problem.