Agentic banking · Offer 5

Loan and collateral agents: your agreements read on site, your account managers alerted

An open AI model runs on a machine installed on your premises. It reads credit agreements, contracts, collateral deeds and releases, extracts clauses, deadlines, rates and currencies, and written rules trigger alerts to the account manager, with the source page. Once the agent is installed, no document leaves your network.

In short

Loan and collateral agents read your credit documents on site, with no upload to an online service, and turn what they read into verifiable alerts: repayment or rate review deadline, collateral about to expire, release due after repayment, currency or rate inconsistent between two documents, unusual clause. The account manager decides; the agent grants, refuses and classifies nothing. Actual status: extraction engine and rules engine tested on five supplier contracts from our own business, in French and English, native and scanned text (open model on site, deterministic rules, e-mail alerts); e-mails, loan schedules and identity documents remain to be evaluated during the trial; credit agreements, collateral deeds and the connection to your loan management system will be assessed during the trial on your documents; no delivery at a bank to date. Price set after scoping.

Typical situations for a loan portfolio manager

Four typical situations for banks, leasing and factoring companies. For each one: the problem, then what the agent would do.

Agreements sleep in the document system

Credit agreements, amendments, pledge and mortgage deeds are PDFs or scans, in Arabic and French. Finding a rate review clause or the exact scope of a guarantee depends on one person's memory.

What the agent would do : The agent reads scanned documents in French and English; Arabic is assessed during the trial. It extracts a table of clauses reviewed by the account manager, with the page number of each value.

A deadline goes unnoticed

Rate review date, expiry of a guarantee, renewal of an insurance required by the contract, maturity of a foreign currency line: the date is in the document, not in the account manager's calendar.

What the agent would do : Written rules compute the deadlines from the extracted clauses and warn the account manager at set milestones, quoting the clause.

The release comes late after repayment

The loan is repaid, the collateral stays registered, the customer complains. In the 2024 report of the Tunisian banking mediation, releases account for 6% of admissible complaints. ([2024 banking mediation report](https://www.bct.gov.tn/bct/siteprod/documents/RAPPORT_ANNUEL_MEDIATION_BANCAIRE_FR.pdf))

What the agent would do : The agent matches the loan balance, exported from your system, with the registered collateral, and flags to the account manager the releases that are due, with the deed concerned.

Documents must not leave

Agreements and collateral contain data covered by banking secrecy and personal data. Sending them to an online AI service raises a question compliance does not want to settle in a hurry.

What the agent would do : Everything runs on a machine installed on your premises, with an open model. Documents stay in your network; the people allowed to access them are named in the contract.

What the agent does, and what it leaves to the account manager

The AI model is only used to read. Alerts come from written, dated and verifiable rules, never from the model's impression.

What the agent does (tested) and would do (to assess during the trial)
  • Tested on five supplier contracts from our own business, in French and English, native and scanned text: it extracts parties, subject, dates, amounts, currencies, rates and renewal terms into a table reviewed by a person.
  • Tested on the same five contracts: deterministic rules compute the deadlines and produce e-mail alerts that state their cause and quote the clause.
  • To assess on your documents: the reading of credit agreements, collateral deeds and releases, e-mails, loan schedules and identity documents, the matching with an export of your loan management system and the consistency check between documents of the same file.
  • It would send the alerts to account managers by e-mail; the connection to your loan management system is built at scoping, it does not exist off the shelf.
  • It keeps a log of every reading and every alert, for review, internal audit and compliance control.
What the agent does not do
  • It grants, refuses, scores and classifies no loan: the credit and classification decision stays with the bank.
  • It writes nothing into your management system without validation: the clause table is reviewed before being used.
  • It does not replace the lawyer: an unusual clause is flagged, not interpreted.
  • It issues no release: it flags those that appear due, the account manager checks and signs.

The deliverables

An installation on your premises, readable rules, alerts that can be checked.

Machine and model installed on your premises

An open model installed on a machine in your network, sized at scoping according to your volume. We propose a trial on an existing workstation before any hardware purchase.

Clause table per file

Parties, subject, amount, currency, rate and review terms, repayment schedule, collateral and its scope, release conditions, each value with its source page. Reviewed and validated by the account manager.

Written alert rules and log

The milestones, the matching with your exports and the consistency checks, written in plain language and editable by your teams; one e-mail per alert with its cause, what to do and the clause quoted; a log of all readings and alerts.

When these agents are not useful

Three cases where we point elsewhere.

Your loan management system already does it

If your agreements and collateral are entered as structured data and your system already manages deadlines and releases, the agent would duplicate it. It is only useful where information sleeps in documents.

A very small portfolio

A few dozen files are followed in a spreadsheet. Scoping an agent would take longer than the reading itself.

Your concern is the system itself

If the problem is an old loan management system to modernise or connect, see the banking industry page: the agents come afterwards.

How it works

Three steps, from the trial on your real files to scoping.

  1. 1

    Trial on a batch of files

    A batch of real agreements, collateral deeds and releases, anonymised if needed, is read on a workstation in your network or on a machine we bring and that you keep under your control; nothing is copied off your premises. We compare the clause table with what your account managers know of these files.

  2. 2

    Review with your account managers

    Correct clauses, partial clauses, missed clauses: we review the result with the people who manage these files, and we set the useful alert rules together.

  3. 3

    Scoping

    Free first scoping conversation. Volume, formats, languages, machine, rules, exports to match, alert recipients, access rights: everything is written down. The price and the timeline are fixed at this stage.

Frequently asked questions

Questions before installing loan agents in a bank

Do the documents really stay with us?

Yes. The AI model runs on a machine in your network and calls no online service to read. In Tunisia, as we read organic law 2004-63 (article 52), the transfer of personal data abroad is subject to the INPDP's authorisation; your compliance team confirms, and on-site processing avoids that transfer. The qualification under banking secrecy and outsourcing rules remains with your compliance team; we write in the contract what the tool does, where, and who has access.

How mature is this offer?

Extraction engine and rules engine tested on five supplier contracts from our own business, in French and English, native and scanned text, with an open model on an office machine and e-mail alerts; e-mails, loan schedules and identity documents remain to be evaluated during the trial. No delivery at a bank to date; the trial on your real files measures the reading quality on your agreements and deeds before you commit.

Are scans and Arabic read?

Scanned documents go through text recognition before reading. Our in-house trials covered contracts in French and English; Arabic, notarial deeds and your agreement formats are assessed during the trial, as quality depends on fonts and scans. We promise no reading rate before the trial.

Does the core banking system need to be connected?

Not for the trial nor for the pilot. A file export (balances, repayment schedules, registered collateral) is enough to match the clauses read with the file's situation. The deep connection to the loan management system comes afterwards, as a classic integration project, if you decide so.

What about a leasing or factoring company?

The need is the same: contracts, collateral, deadlines, currencies, end of contract and return. The alert rules are adapted to your portfolio. The same engine also serves the policy and claims agents of our insurance section and companies' supplier contracts in our agentic solutions.

Page reviewed and updated on · Access International, a Franco-Tunisian IT group founded in Tunis in 1999, with a subsidiary in Paris.

Try the reading on a batch of your real agreements

Tell us the type of files and the volume: we propose an on-site trial, then scoping with your account managers.

Discuss your project

See all agentic banking offers →