Banking: AI agents prepare, bank staff decide. What 2026 has already shown
The announcements of 2026 no longer talk about chatbots but about agents: programs that read a file, prepare an operation, match due dates, and hand back to a person. Large French groups build or buy them, sometimes with Mistral AI or Google. Tunisian banks, for their part, must first answer a new obligation on their data. Here are the facts, and what they make possible today.


What French banks have put into production
Société Générale created a sixty-person unit, SocGen AI, which is building an agent platform. A first securities operations agent has been in production since October 2025; the unit stated, as at 14 December 2025, 84% reliability and describes an orchestrator routing customer messages to specialised agents (La Revue du Digital, 14 December 2025).
On 10 June 2026 Crédit Agricole announced a programme of close to 500 million euros over 2026-2028 and the creation of a group AI company (press release of 10 June 2026). Its head of AI says customer-facing agents are in development, and that there is no such thing as zero hallucination (IT for Business, 17 September 2026).
BNP Paribas extended its agreement with Mistral AI by three years, with know-your-customer solutions in pilot, two pilots in Belgium, where experts validate decisions (Markets Media, 27 May 2026; LeMagIT, 27 May 2026), then signed a five-year deal with Google for its internal agents, with first deployments on the preparation of credit analysis files (IT for Business, 28 September 2026).
In the case of BNP Paribas, experts validate the decisions; for Crédit Agricole's payment, the customer validates: the first agentic payment completed in France on 25 June 2026 with Worldline and Mastercard was executed only after the customer's explicit approval (press release of 25 June 2026).
What supervisors say
On 9 September 2026, Denis Beau, first deputy governor of the Banque de France, stated that almost all French banks and insurers now have artificial intelligence use cases in production, according to the ACPR's 2025 survey, and that the ACPR will supervise high-risk systems from December 2027 (Banque de France, 9 September 2026). The ACPR had opened on 1 July 2026 a consultation on algorithmic fairness in the financial sector, closed on 30 September (ACPR).
At European level, the European Banking Authority plans in its 2027 work programme a report on the provision of AI systems to the banking sector by third parties and an analysis of agentic AI in payments (EBA, 2027 work programme). Credit scoring is among the high-risk uses of the European AI regulation; that is one reason, among others, why none of the offers described below contains scoring.
And in Tunisia?
No production deployment of an AI agent is documented at a Tunisian bank. Amen Bank launched AmenBot on 28 September 2020 to inform, collect complaints and route customers to the contact centre (ilboursa, 28 September 2020), and announced on 26 January 2026 a new version based on language models and a financial coach (interview of Néji Ghandri, chairman of the executive board, L'Économiste Maghrébin, 26 January 2026).
The regulator, however, has moved. Central Bank of Tunisia circular 2025-08 of 16 May 2025 requires in its article 6 a board-approved data governance policy, a data officer, quality controls, traceability and the automation of critical data flows to limit manual errors; article 7 asks for a plan within three months (circular 2025-08). Circular 2025-06 of 28 February 2025 also frames remote customer onboarding (circular 2025-06).
The banking ombudsman gives the measure of the irritants: 506 complaints received in 2024, up 56.2%, first on credit (31.2% of admissible complaints), account operations (20%) and means of payment (14.3%); 11.4% were handled beyond the regulatory two-month deadline, against 21.7% in 2022 (2024 banking mediation annual report).
What a bank can entrust to an agent today
Prepare, not decide. Read a complaint file, sort it by reason, check deadlines, propose a reply the handler signs. Check that an onboarding file is complete and consistent, chase missing documents. Match loan instalments with receipts and prepare graduated reminders. Read loan agreements and collateral to extract clauses, rates, currencies and due dates, and alert before a date.
On the premises. For a Tunisian bank, files are in French and Arabic, often scanned, and covered by banking secrecy and the personal data law. An open model installed on a machine on the bank's premises reads those documents without any of them leaving the network. We tested this component on five supplier contracts from our own business, in French and English, native and scanned text; e-mails, loan schedules and identity documents remain to be evaluated during the trial. It has not yet been delivered in a bank, and we say so.
Measured. A trial on a batch of real files, a measure on the bank's own turnaround times, then scoping. No gain figure announced beforehand.
Where to start
For a Tunisian bank, the data and agents assessment answers two needs at once: the map of candidate processes and the data governance plan required by circular 2025-08. For a bank that already has a process overflowing, the complaints agent or the loan and collateral agents can be tested on a batch of real files.
The eight offers, including the training and, for the French market, the agentic payment readiness review, are described in the Agentic banking section. The core system itself belongs to the banking industry page.
What decision-makers ask about this topic.
Can an AI agent handle a complaint end to end?+
It can receive it, qualify it, prepare the file and propose a reply. The reply goes out signed by a handler. Regulatory deadlines, two months in Tunisia, are tracked by the agent, not decided by it.
Large banks build or buy their platforms; what is left for a provider?+
The peripheral processes of mid-sized banks, neobanks and payment providers: complaints, KYC documents, arrears, contracts. These are agents installed alongside the system, not inside the core.
Where do the models run?+
In Tunisia, on a machine installed on the bank's premises; no file leaves the network. In France, on the hosting chosen by the client, on its premises or with a European Union host.
What about credit scoring?+
We do not offer it. In France it is among the high-risk uses of the European AI regulation, applicable from 2 December 2027 (Regulation (EU) 2026/1744); the designation of the ACPR as the authority is proposed in a bill not yet passed; in Tunisia no dedicated framework exists yet; and it is not our trade. An agent prepares the file, a committee decides.
Does your program match this case?
We frame every program at Intake, with transparent budgeting. A short POC of a few weeks can be delivered before committing to the full program.
Scope an agentic banking pilot →