How to process business account opening applications with an AI agent

Business onboarding turns on ownership. The agent pulls the registry record, builds the ownership tree to the people at the top, verifies directors and signatories, screens every party and lists every gap. A named person accepts the customer, sets the risk rating and signs off complex structures. Only then does the agent set up the account.

What does business account opening involve, and who owns it?

Opening an account for a company, partnership, charity or trust means knowing the entity, the people who own and control it, the people who will operate the account, and what the business does and expects to put through the account. For a sole director company with one shareholder that is quick. For a group with overseas holding companies and a trust at the top, it can take weeks.

Business onboarding teams own the process, with financial crime setting policy and approving higher-risk cases. Relationship managers often bring the application in. The work differs enough from personal onboarding that it needs its own capability; processing personal account opening applications covers the individual case.

What does the agent prepare?

1. Agent, depth 1: verifies the entity. Registry record, registered address, status, filing history, incorporation date.
2. Agent, depth 1: builds the ownership tree. From the registry, the application and any group structure chart, up through each layer to the individuals who own or control the business. Each link is cited to a document.
3. Agent, depth 1: compares sources. Ownership declared in the application against the registry, and the registry against any constitutional documents. Every difference is listed.
4. Agent, depth 1: verifies people. Identity checks on beneficial owners, directors and signatories under your policy.
5. Agent, depth 1: reads screening results. Sanctions, politically exposed persons and adverse media for the entity and every related person, from your screening system.
6. Agent, depth 1: checks the business. Trading presence, stated activity against industry codes, expected turnover against the size of the business.
7. Agent, depth 2: drafts the onboarding file and requests. A summary with risk factors and a proposed rating, plus a request to the applicant for missing items.
8. Person: decides. Accept or decline, set the risk rating, require enhanced due diligence, sign off complex structures.
9. Agent, depth 3: sets up the account. Once approved, it creates the account and the signing mandate exactly as approved.

What does a named person approve, and why?

Accepting a business customer is a risk decision the bank must be able to explain. It rests on judgements the agent can inform but not make: whether a structure has a legitimate reason to be complex, whether a director's adverse media matters, whether expected turnover fits a business of that size.

The risk rating carries forward into monitoring thresholds and review frequency, so the person who sets it should be accountable for it. Setting up the account afterwards is routine work inside approved terms, which is why it is depth 3. Whether an agent can open a bank account for a customer depends on that sequence: approval first, set-up second.

Which controls apply, and what evidence do they leave?

Provenance on ownership. Registry data is external, and much of it was declared to the registry by the company itself. The application is declared. A certified copy of a share register checked by your team may be signed. The ownership tree shows each link's class.

Discrepancy list. Every difference between sources is filed with its resolution and the person who resolved it.

Mandate set-up checked against approval. The signing rules the agent configures are compared with the approved mandate before the account goes live. Any difference is refused.

Held decisions. The onboarding file is filed as held until the approver decides, with their name and reasons.

Where does it usually go wrong?

Trees that stop too early. The chain reaches an overseas holding company and stops because the registry there is harder to read. The agent must say where it stopped and why, not present a partial tree as complete.

Registry taken as proof. Registry ownership is often self-reported. Treat it as one source among several.

Nominees and shelf companies. A newly incorporated company with a nominee director and very high expected turnover is a classic warning pattern. The agent should list these risk factors plainly, not bury them.

Expected activity left vague. "General trading" tells monitoring nothing. If the application does not say what flows to expect, the agent's request should ask. A vague profile makes later monitoring blind.

Mandate errors. "Any two directors" set up as "any one director". The check against the approved mandate exists for this.

What should you measure?

• Time from complete application to decision.
• Applications needing more than one round of requests. Clear first requests are the agent's easiest win.
• Discrepancies found after account opening.
• Proposed ratings changed by approvers, by direction.
• Mandate corrections after go-live. Should be none.

Lending applications from the same businesses follow a different path, with their own steps in processing personal loan applications for individuals and credit memos for business borrowers. The general limits are in AI agents in banking.