Blog · Published · Updated
Counterparty research file: how the workflow works
Short answer
A team of agents in CrewAI is worth choosing only when it gives a better result than one agent on the same tasks, criteria and tools. In counterparty research, three roles usually use several times more model calls: at 200 files a month that is about €77.11 against €25.70 for a single agent. Scenario.
How we workDetails ↓
- 01Call30 minutes with an engineer, free
- 02PrototypeA working system on your process and data, free
- 03ImplementationTarget and fixed price in the contract; 50% at signing
- 04Acceptance50% once the agreed target is met
- analysis with an engineer and a free prototype
- €0
- team in ChatGPT or Claude, up to 10 people
- from €1,150
- automation, agent or application for one process
- from €3,000
We belong to the OpenAI and Anthropic partner programmes and learn through their official courses. What this means ↓
Author: Syntalith · 3 min read
Statistics
AI agents in companies: statistics 2026
3×
more for the model in a team of three agents than in one agent doing the same research, when each role makes 6 calls: €77.11 against €25.70 a month at 200 files. Scenario.
40%
According to Gartner (2025), over 40% of agentic AI projects will be cancelled by the end of 2027 because of costs, unclear value or weak risk controls.
130
According to Gartner (2025), of the thousands of agentic AI vendors only about 130 offer real agentic solutions.
392,716
According to GitHub (2026), the OpenClaw repository has 392,716 stars and 82,549 forks, and the project was created on 24 November 2025.
90%
According to the European Commission (2026), 90% of companies using AI say that people take the decisions, and 2% of decisions are fully automated.
Source: European Commission, Flash Eurobarometer 578 (2026) ↗
Figures checked: 22 October 2026
The same research in two architectures
Preparing a company profile means joining sources and spotting gaps: register data, public procurement, the company website and the news. In CrewAI you can split this work into roles: a researcher, a verifier and a report writer. One agent does the same in one loop with the same tools. Anthropic recommends starting with the simplest architecture and adding complexity only when the result justifies it.
Whatever the architecture, one rule stays: an unconfirmed name match must never turn into a certain assignment of data to a company. A person confirms the identity of the entity.
- 01Tasksexamples with an agreed correct result
- 02Toolsthe same sources and the same access for both versions
- 03Modelthe same model and the same limits
- 04Criteriaone scoring rubric, set before the trial
- 05Measurementresult, time, model cost, handovers to a person
What a team of agents costs
- One agent: 6 calls per file
- €0.13 per file, €25.70 a month
- Team of 3 agents: 18 calls per file
- €0.39 per file, €77.11 a month
The team costs 3 times more, a difference of €51.41 a month
The number of calls and files are the scenario's assumptions. Model: 8,000 input tokens and 800 output tokens per call, Claude Sonnet 5.5 at Anthropic's rates, converted at the ECB rate of 1.1205 USD per euro on 22 October 2026. The model cost is a small part of the labour cost, so the quality of the result decides.
The difference in model cost rarely decides. Response time, the number of places where something can go wrong and how easy it is to test matter more. A team of agents has more handovers between roles, and so more points where one role's error passes further down.
When a team of agents makes sense
Splitting into roles helps when parts of the work really need different tools, permissions or models. For example, the researcher has access only to public sources, and the verifier to a paid register and the history of work with the company. It also helps when one part of the work runs for a long time and can run in parallel with others. A team of agents does not help when all the roles use the same tools and data, because then one job is split into several conversations that pass summaries to each other and lose detail.
The same tools and data for every step
one agent with well-described tools
simpler tests
Roles with different permissions or sources
a team with separated access
more handovers
Independent, long parts of the work
parallel calls in one flow
shorter response time
Every one of these decisions rests on the same measurement: the result on tasks with an agreed correct outcome, the time and the number of cases handed to a person. The framework matters less than whether the architecture fits the work.
What our demonstrator shows
- The data snapshot was small and deliberately incomplete
- No financial statements and no full procurement documentation
- Automatic entity resolution was unconfirmed
- The rubric scored a specific contract, and says nothing about research quality in general
Our counterparty file demonstrator from August 2026 shows how gaps and confirmations work. It was not run again during the rebuild, and its result is no framework benchmark. For a first trial in your company, prepare examples with an agreed result, a decision owner and the data sources. We build research agents as part of our AI agents, and for the free analysis a list of the questions the research has to answer is enough.
Worked example · Scenario
How we count the payback: counterparty verification
We run the same count for your research in the free analysis.
€1,047
a month for checking 60 counterparties by hand
Payback: 3.4 mo.
30 minutes to check a counterparty and checking in 15% of the old time are our assumptions. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat, for companies with 10 or more employees. We measured the model cost of the simpler profile, €0.009, in our counterparty profile system, converted at the ECB rate of 1.1205 USD per euro on 22 October 2026. Avoided losses are a scenario: we assume that 1 in 800 counterparties checked is a company that would not have paid, and that the check protects €6,900 on average. In the review we calculate it on your receivables.
See the full calculation →
