12 flags
The flags fall into three groups of four. None of them is a reason to refuse at once, but each is a reason to get a written answer before you sign.
Price and payment.
1. No metric. The commercial proposal says what the agent can do, but gives no number the vendor will put in the contract: response time, share of leads logged in the CRM, share of requests solved without a person.
2. Payment for hours. The estimate lists developer hours with no result at the end: the longer the work goes, the bigger the bill.
3. “Our own neural network”. A model is not trained from scratch for an agent. DeepSeek R1 came out in January 2025 with open weights under the MIT license, and the Qwen 3 family in April 2025 under Apache 2.0. If a vendor charges separately for “its own neural network”, ask which model is inside.
4. Connections as a “separate stage”. Integration with your CRM, accounting system (in Russia often 1C) or telephony is left outside the price, and its cost appears after the start.
What is inside the agent.
1. An agent that is not an agent. Renaming chatbots as agents is called agent washing, a swap of the concept. Ask the vendor to show where the agent picks the next step and calls a tool, and where it follows a fixed script with buttons.
2. A demo only on someone else’s examples. The vendor shows the agent on its own questions and refuses to test it on your requests from last month.
3. No handoff to a person. The proposal does not say what the agent does with a hard question: passes it to a manager with the dialog history, or answers at random.
4. No answer log. If you cannot see what the agent answered each client and why, you will find an error from a complaint, not from a report.
Data and contract.
1. Data abroad or with no address. The proposal does not name the country and the company whose servers process the correspondence. Under 152-FZ, Russia’s personal data law, the personal data of Russian citizens is collected and stored in databases in Russia, and before a transfer abroad the operator notifies Roskomnadzor, the Russian data regulator. More on the data security page.
2. The system stays with the vendor. The scripts, rules and dialog history sit in the vendor’s account, and when you part, you lose the agent.
3. One model with no replacement. The agent runs only on one foreign model, and the vendor cannot switch it without rebuilding.
4. No price for the month after launch. The proposal has a one-time sum but no monthly one: the model, the server, support.
Questions for the vendor
Ask these questions in writing and compare the answers of different vendors on one scale. General questions about price, timing and ownership of the system are in the article on choosing an AI contractor; here the questions are about the agent itself.
1. Which model is inside, under which license, and can it be replaced?
2. Which steps does the agent take on its own, and which follow a script? Show it on one live request.
3. Which systems is the agent connected to today, and which connections are still to be built, at what price?
4. Where is the correspondence processed: country, company, servers?
5. Where is the answer log, and who reads it every week?
6. Which number will you put in the contract, and what happens if we do not reach it?
7. What does a month after launch cost: the model, the server, support?
Gartner counts only about 130 real agent vendors out of thousands that claim it, so question 2 matters most. If a vendor refuses to show the agent on your requests, that is also an answer. How we check a solution ourselves before the contract is described on the vendor check page.
What the price should consist of
A clear estimate for an agent splits into lines, and you can check each one:
1. Process review and metric: what the agent does and how we count the result.
2. Connections: CRM, accounting system, telephony, messengers, each on its own line.
3. Instructions and knowledge base: rules, price list, answers to common questions.
4. A test before launch and the first weeks under control.
5. Monthly costs: the model, the server, support.
A typical agent’s estimate needs no lines for a “neural network license” or “model training”: open models such as DeepSeek R1 and Qwen 3 are free to download, and you pay for setup for your process. When development is really needed and when a ready product is enough is covered separately.
For comparison, our range: a pilot of one agent from $1,499 in 14 days, support from $480 a month, and the price is fixed before the start. All price levels are covered on the page about what an agent costs.
What to write into the contract
You can still fix a proposal, but not a signed contract. Before you sign, check six points:
1. The metric, how it is counted, the data source and the date by which it is reached.
2. What happens if the metric is not reached: rework at the vendor’s cost, a partial refund or an exit from the contract.
3. The full list of connections with the price of each.
4. The country and the company whose servers process client data, and a ban on training other models on your correspondence.
5. Who owns the system, the rules, the dialog history and the accounts after the contract ends.
6. The exit procedure: what you take, in what format and within what time.
With us, the pilot metric is in the contract before the start, and the contract sets what happens if it is not reached. You can check another vendor’s proposal against these points with AI due diligence, a full check of a company and its decisions about AI.