MMOOZZGG.com

AI due diligence

AI vendor due diligence: how we check a solution before the contract

5 min read2 sources

How do you check an AI vendor when every website makes the same promises? AI vendor due diligence tests not the presentation but the work of the solution on your examples, its connections, the contract and the way out of it. Gartner estimated in June 2025 that there are only about 130 real AI agent vendors among the thousands that call themselves that.

On this page · 5 sections

What the solution does without a person

First we find out which part of the work the solution does by itself and where a person stands behind it. Under the words “AI agent”, vendors sell both a scripted chatbot and manual work by their own operators.

We ask to see the solution not on the vendor’s examples but on yours: real requests, leads or documents from last month.

We count three numbers: the share of cases the solution closed without a person, how many errors it made and how many it passed to people. If the vendor refuses a test on your examples, that is already an answer.

We repeat the test on fresh examples a week later: this shows whether the solution was tuned for the demo or works on any case.

We write the results into a table next to the answers of other vendors, so we can compare solutions on one scale.

In AI due diligence, a full check of a company before AI decisions, we check both what the company has already bought and what it plans to buy.

Which systems it connects to

A solution that is not connected to your systems adds work for staff: they must move the result by hand. So we check which systems it connects to now, not “in the plans”.

We ask to see the connection to the CRM, the accounting system or the phone line on a live example: a lead from a messenger appeared in the CRM, a document was posted in accounting. We clarify who builds the connection, whether it is in the price and what happens when the system is updated.

We check the access rights separately: which access the solution asks for and why. A solution that answers customers does not need access to all of accounting.

If a connection does not exist yet, we ask for the term and the price in writing, before the contract.

Result of this step: a list of connections that work today and those the vendor still has to build, with the price of each.

The metric in the contract

We look at which measurable number the vendor is ready to write into the contract before the start: time to first reply, the share of leads entered in the CRM, the share of requests handled with no human.

We compare the promise with your current number: if the vendor promises replies many times faster than your manager answers now, we ask to see how.

We check how the number will be counted: from which system, over what period and who verifies it. A metric that only the vendor counts in its own panel is not a metric.

We read what happens if the number is not reached: rework at the vendor’s cost, a partial refund or an exit from the contract. If the contract text has no answer, we ask to add one.

The full list of questions for a vendor is in the article on how to choose an AI vendor. Here we describe how we check the answers ourselves.

Where the data is

We find out which servers and services the information passes through: where the conversations are stored, where the model runs and which outside services read the text of requests.

Under 152-FZ, Russia’s personal data law, customer personal data must be processed on servers in Russia. We do not accept an answer without the country and the company that hosts the servers: we ask to write them into the contract.

We check whether your conversations are used to train the vendor’s models and whether the contract can forbid it.

We look at who on the vendor’s side sees your data: developers, operators, subcontractors. Each extra person with access raises the risk of a leak.

Result of this step: a map of where each part of the information sits and who has access to it, with notes on what to change in the contract.

What happens if the vendor leaves

91% of Russian companies work with a single AI vendor, according to a December 2025 survey by a major Russian telecom’s cloud unit. If that vendor raises the price, shuts down or falls under restrictions, the company has no backup.

So we check the exit in advance: can you export the data, history, settings and rules of the solution, in what format and how fast. We clarify who owns the accounts and access.

We check whether the solution is tied to one model or the vendor can change the model without a rebuild. A solution that works only on one foreign model carries the risk of being cut off.

Result of this step: a one-page exit plan: what we take, in what form, how fast and what it costs.

How to reduce vendor lock-in is covered separately. If the company already has several contractors, their estimates and deadlines are checked under the rules for control of contractors and the IT budget.

Sources and fact-checking

Checked by the editors on 13 October 2026.

  1. ComNews: only 26% of Russian companies have an AI development strategy, 22 December 2025 (comnews.ru)
  2. Gartner: over 40% of agentic AI projects will be canceled by the end of 2027, 25 June 2025 (gartner.com)

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min
Contact us