# ChatGPT for account research: fine until you need to keep track

> A $20 subscription is great at researching one company before a call. Once you have two hundred accounts to watch every week, the manual work starts. Here is exactly where.

Source: https://signal.granek.agency/en/blog/chatgpt-for-account-research
Published: 2026-10-22

A rep opens ChatGPT or Claude and types: "Tell me about Alpha Industrial: what they do, who runs sales, what has been in the news". A minute later there is a clear answer with links. The subscription costs [$20 a month](https://claude.com/pricing). For five companies before five calls, that is more than enough.

The trouble starts when there are two hundred companies and you need to keep track of them.

## Where the subscription is enough

To be fair, for many jobs a chat is handier than any specialised tool.

- Getting up to speed on a company before the first call.
- Understanding an unfamiliar industry in half an hour.
- Drafting an email once you already have the reason to write.

If you have a dozen target accounts and come back to each one once a quarter, you can stop reading here. Get the subscription.

## What changes at two hundred accounts

A sales team owns 200 target accounts. The job sounds simple: know every week where something happened. They opened sales roles, announced a tender, changed their CIO.

In a chat that means: paste the name, wait for the answer, read it, decide whether there is news, copy it into the CRM. Even at 3 minutes per company, that is 600 minutes, or 10 hours a week. And in a week packed with calls, nobody checks the list at all.

![A chat at the two hundredth company: the same question, and 10 hours a week of copy and paste](/blog/figures/chatgpt-week-en.jpg)

## Five places where it turns into hell

1. **No comparison with last time.** Today's answer about Alpha Industrial and last month's answer have to be compared by eye. The chat tells you what is true now. Deciding what is new is on you.
2. **Scheduling within tight limits.** On the Plus plan, ChatGPT allows [5 active scheduled tasks](https://help.openai.com/en/articles/10291617-scheduled-tasks-in-chatgpt). The result arrives as a chat message, and someone copies it into the CRM by hand again.
3. **Facts nobody checked.** The model confidently names a head of procurement who left the company a year ago. Not every claim comes with a source, and checking stays with the rep. We covered how to verify people in the article on [finding decision makers](/en/blog/how-to-find-decision-makers).
4. **Research lives in personal chats.** Two reps research the same company and write to the same person. A rep leaves, and their history leaves with the account.
5. **The answer is prose.** You cannot filter paragraphs for "companies that changed their CIO this month" or sort the list by how fresh the reason is.

![Two answers about the same company a month apart: finding what changed is up to you](/blog/figures/chatgpt-diff-en.jpg)

## "Then we will script it on the API"

The next logical step is to connect the API and run the checks automatically. Search in the [OpenAI](https://developers.openai.com/api/docs/pricing) and [Anthropic](https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool) APIs costs $10 per 1,000 searches. On top of that you pay for the text of the pages found, which the model reads as input. How much one company costs, you learn only after the run.

Calling the model takes an evening. Then comes the list of what you have to build and maintain:

- scheduled runs, and retries when search or the model does not answer;
- a store of what you knew about the company last time;
- a comparison with the previous check, so you see only what is new;
- merging duplicates: the same news reprinted on ten sites;
- a separate check that the person really holds that title;
- a spending limit per run, so one company does not eat the budget;
- shared access for the team and an export to the CRM.

And models do not last forever. OpenAI shut down its deep research models on July 23, 2026, as their [deprecations page](https://developers.openai.com/api/docs/deprecations) shows. Anyone whose script was built on them rewrote it over the summer.

![An API script: the model call is done in an evening, the rest of the list remains](/blog/figures/chatgpt-script-en.jpg)

## "We will vibe-code it over the weekend"

These days nobody has to write that script by hand. Claude Code or Codex will put together a first version over a weekend, and it will even work. In the demo it looks great.

Then comes the part the vibe-coding posts leave out:

- search changes its response format, and half the companies quietly drop out of the report;
- the model gets shut down, and the prompts have to be rewritten for the new one;
- the monthly API bill turns out several times higher than expected, and nobody knows why;
- a rep asks why the report lists someone who left the company in the spring.

Sorting this out is on you. First in the evenings, then at weekends. A couple of months in, it becomes clear that the founder or the head of sales is fixing a script instead of selling. So you hire someone to look after it.

Let us do the maths. The median pay of a software developer in the US was $134,040 a year in 2025, according to the [Bureau of Labor Statistics](https://www.bls.gov/ooh/computer-and-information-technology/software-developers.htm). A Signal subscription is $299 a month.

| | A year |
|---|---|
| An in-house developer, salary only | $134,040 |
| A Signal subscription | $3,588 |
| The difference | $130,452 |

The developer costs 37 times more, before benefits, taxes and the API bill. And all of their work goes into an internal tool that brings in no revenue by itself.

Ask yourself a simple question: are you building your business, or an account research service? If it is the service, you now have a second product, and it needs managing too. If it is the business, your time and your sales team's time are worth more than a ready-made tool.

## Side by side

| | Chat subscription | Your own API script | Signal |
|---|---|---|---|
| One company before a call | Good | Overkill | Good |
| Watching 200 accounts | By hand | Has to be built | On a schedule |
| What changed since last time | You compare | Has to be built | Shows what is new |
| A source for each fact | Not always | As you build it | For every fact |
| Price of a run | Part of the plan | Known after the run | Shown before the run |
| Team work | Personal chats | Has to be built | Shared lists |
| Who fixes it when it breaks | OpenAI or Anthropic | You or a developer you hire | We do |

## When each one is enough

If you have fewer than twenty accounts and do not need to watch them, a ChatGPT or Claude subscription is enough. If you own a list that should tell you by itself when there is a reason to call, you need a system with memory, checks and a schedule. You can vibe-code one and keep it running yourself. Or use one that exists and get on with selling.

To work out how much time research takes from your team, take the quiz on the home page.