ChatGPT or a specialist tool for Transaction Services?

The question seems simple: should we use ChatGPT or invest in a specialized tool for Transaction Services? However, I would start by reformulating the need. Preparing a management question, extracting information from a contract and maintaining a register of evidence in a moving data room are not three versions of the same task. They involve different sources, responsibilities and control requirements. A useful answer for writing is not necessarily enough to conclude on the quality of the results.

A general assistant can work in an approved professional environment, with functions that depend on the offer and its configuration. A specialized tool can combine a comparable model with workflow, calculation and document management mechanisms. Specialization is therefore not a guarantee of better intelligence, and the name of an assistant does not describe all of his abilities. I would compare the environments actually offered, not a general public version against a commercial promise.

For a buyer, the challenge is to be able to use the product file. If the manager must find the supporting documents, redo the allocations and rewrite the reservations, the speed of the first draft matters little. I would look for a reproducible workflow: request, source, calculation, discussion, decision and restitution. The right choice depends on which part of this chain your team wants to improve.

I would select three tasks from the mandate. The first could be the preparation of an IRL adapted to the scope. The second, the examination of a candidate for EBITDA normalization. The third, the reopening of a conclusion after replacing a source. These tests involve writing, financial reasoning and documentary control. They avoid choosing a solution based on a single, carefully prepared demonstration.

It is also necessary to distinguish the data used. An anonymized example, a confidential contract and a payroll export cannot be deposited indiscriminately into any account. Before testing, I would have the environment, data categories and users approved. A security check occurs before the functional rating: a weakness in confidentiality is not compensated for by a good dashboard.

A useful comparison does not assign all functions to one category in principle. Some general offers have connectors, workspaces and administrative controls. Some specialized solutions do not retain all calculations or require manual exports. Request a demo on your setup. The following table describes the questions to ask, not a predetermined verdict.

I would also look at the system output. An excellent result in a conversation can become difficult to transmit if the references disappear on export. Conversely, a less dramatic workflow can be valuable if it allows the manager to quickly understand the sources and reserves. The level of control must be proportionate to the task: a wording suggestion does not require the same procedures as a price adjustment.

Let's imagine an invoice for 200 k€ included in the EBITDA for the period. Management presents it as an exceptional cost. The general ledger v3 and the invoice support the amount, but not necessarily the non-recurrence. A first draft therefore describes a candidate of 200 k€, subject to the professional nature examination. The ledger v4 then arrives with a credit note of 40 k€ attributable to this same service.

The candidate amount becomes 160 k€ if the credit note relates to the expense and period under review. I would ask each environment to find the write, check the allocation and identify the conclusion that used 200 k€. A silent change is insufficient: the manager must know why the amount has changed, which version is authentic and whether the exceptional qualification remains open.

An answer may correctly recalculate 200 minus 40 while failing to flag which report paragraph to reopen. This is precisely the difference between calculation capacity and workflow control. Conversely, detecting the new version without understanding that the credit note concerns another service would lead to a bad correction. The professional review must cover both dimensions.

An assistant can come up with an excellent list of supporting documents. I would then test a December receivables request for two entities, with detail by invoice, accounting value adjustments or provisions for charges and subsequent collections. If the dataroom only contains a consolidated PDF, the good result is partial: the document is relevant, but it does not respond to the entire request. An answer that declares the complete line confuses file accounting reconciliation and sufficiency of proof.

The benefit of a specialized workflow can lie in monitoring this discrepancy: reminder owner, missing entity, missing calculation and reviewer validation. But these functions must be demonstrated. If your approved general purpose tool provides them in a controlled configuration, they are also included in the comparison. What matters is the file obtained and its maintenance when the supporting documents evolve.

I would never drop a transaction file into an untrusted account to do a quick comparison. It is necessary to examine the use of documents and requests, model providers, processing locations, conservation, backups and support access. Lack of data training is important, but it doesn't cover all privacy issues.

I would ask for the conditions actually applicable to the successful offer: processing agreement, subcontractors, authentication, logging and deletion at the end of the mission. A certification can support the assessment, without guaranteeing all configurations. Security, data and legal managers must assess the requirements based on the parties, jurisdictions and information involved. A general compliance sentence does not replace this work.

I would prepare a reference set established by the team and kept away from suppliers: bad periods, duplicates, missing supporting documents, simple calculations and ambiguous allocations. Both environments would receive the same authorized tasks. The protocol would distinguish correct answers, significant errors, unanswered requests and minutes of professional review. An accuracy note alone can hide a serious error in a material finding.

We must measure preparation, control, correction and restitution. Assume a manual process of 40 hours and an assisted process of 28 hours, including 14 preparation, 10 professional examination and 4 correction. The gain is 12 hours, not the 26 saved on preparation alone. These fictitious figures are used to explain the calculation; they constitute neither a product benchmark nor a promise of savings.

I would set the acceptance criteria before the test. For non-confidential writing, a stylistic error can be easily corrected. For a recommendation on a debt amount or a protected source, the risk is different. A good solution can be accepted for one task and rejected for another: deployment does not need to be all or nothing.

For one-off formulation tasks in an approved environment, a generalist assistant can meet the need. If the objective is the sustainable monitoring of hundreds of requests, version control and transmission of a revisable file, workflow mechanisms become decisive. A specialist solution can provide these, but the demonstration must confirm that they perform at the expected level.

A combination is possible if the responsibilities and data boundaries are clear. It can also multiply copies, accounts and versions. I would ask who has the reference source and where the final decision is kept. No draft should become a conclusion because it has circulated between several tools. Investment in technology replaces neither the scope of the mandate nor the judgment of the team.

Solenor presents workflows for documentary accounting reconciliation, transcription, professional review of evidence and preparation of deliverables. I would apply the same controls to them as to other solutions: findable sources, verifiable calculations, demonstrated rights and measured net gain. The correct conclusion from the comparison is not that one category always wins, but that the chosen workflow meets your needs without hiding its limitations.