Using AI to prioritize sales accounts

Sales teams often have more accounts than they can work. Someone still has to decide where the team should spend its time.

Projects like this often start with a discussion about tools. Which data provider should we use? Which model? Should the workflow run in Clay, code, or the CRM?

Those choices can wait. First, define what should improve for sales: less time spent researching poor-fit companies and a shorter list of accounts worth contacting.

The reasons and sources still need to be there, but they support the decision. They are not the result.

It does not need to find every company in the market. It does not need to identify contacts, write emails, or update the CRM. Those may become later parts of a larger outbound system. First, prove that the agent can help someone choose where to spend time.

My article on choosing the first job for a GTM AI agent explains why account selection often works well: the result can stay read-only and a salesperson can inspect it before anything reaches a prospect. Here I cover how to build it and where to stop.

Begin with the decision sales already makes

Do not give the agent a vague instruction such as “find good accounts.” Give it the decision a salesperson is already trying to make.

For a first version, I like three answers:

• Work now: the available evidence supports the target criteria.
• Skip: the company clearly misses an important criterion.
• Take a closer look: the company might fit, but a fact is missing, two sources conflict, or the company itself is unclear.

Without the third answer, the agent has to turn weak evidence into a confident yes or no. That produces tidy spreadsheets and bad decisions.

If the list feeds an automated campaign, only the first group should continue. “Take a closer look” stays out until a person resolves it. The spreadsheet can have three queues even when the downstream system accepts only a binary decision.

The salesperson still decides which accepted accounts to work first. The agent is removing research and first-pass sorting, not taking ownership of the sales strategy.

Turn the target criteria into questions

The target brief has to become a set of questions that public evidence can answer.

Suppose a team wants established B2B software companies selling to finance leaders in the UK and Germany. “Looks like our ICP” is not enough. The agent needs questions closer to these:

These questions are examples, not a universal scorecard. A different campaign needs different evidence.

Company size can often be checked from structured data. Business model and buyer fit usually cannot. An industry label might help narrow a list, but it does not prove what the company sells or who buys it. For that, the agent needs to read the company's own product and customer material.

Any disagreement about the target criteria has to be resolved before the agent uses them. If one salesperson includes agencies and another excludes them, the agent cannot quietly decide which interpretation wins.

Resolve the company before researching it

Company identity looks like an administrative detail until it goes wrong.

Names collide. A brand may belong to a larger group. A careers page may sit on a recruitment platform. A company may have changed its domain after an acquisition. A LinkedIn profile can point to a similarly named business in another country.

If the agent researches the wrong company, everything after that can look convincing. The website is real. The quotes are real. The conclusion is still useless.

Make identity the first check:

1. Start from a sourced domain, company profile, or CRM identifier.
2. Confirm that the website describes the intended operating company.
3. Keep the original identifier with the research.
4. Stop when the domain, company profile, and supplied name point to different businesses.

Research only what the decision needs

Once identity is clear, the agent can collect evidence for the target criteria. This should be a short route, not a general company dossier.

If the question is whether the company sells software to finance teams, the useful pages are likely the product, solution, pricing, and customer pages. A founder interview from four years ago may be interesting, but it is not the first place to establish the current offer.

I use a simple order:

1. Reuse trusted fields already attached to the company.
2. Open the known official website.
3. Read the specific official page that should answer the criterion.
4. Search for another official page only when the needed page cannot be found.
5. Use another provider when the missing fact genuinely belongs there.

The order keeps the work cheaper, but the larger benefit is clarity. Every research action answers a named question. When the agent cannot answer it, you know exactly what remains unknown.

Keep the evidence with the result. At minimum, each important finding should include the source URL, the fact or passage used, when it was checked, and which criterion it answers. A final verdict without those pieces asks the salesperson to redo the research.

Search snippets are discovery aids. They are not strong evidence for a semantic decision. The agent should open the selected page and use what the page actually says.

Keep research and scoring separate

AI is useful when a company page has to be interpreted. Fixed rules are better for arithmetic and exact conditions.

For example, an AI step can read a product page and decide whether the product is clearly sold to finance teams. A formula can then apply the agreed rule that every required criterion must pass. There is little benefit in asking the model to add five numbers or remember that one hard exclusion overrides the rest.

That separation made the Vanta buying-window model I built for the Clay Cup easier to inspect. AI handled research and classification. Formula columns calculated the score. If an account's score changed, I could trace the change to an input rather than wonder whether the model had simply answered differently.

The same principle applies without a numerical score. Let the model interpret evidence. Let ordinary logic enforce exact exclusions, required fields, and the final output shape.

Return something a salesperson can scan

The first result does not need a dashboard. A table is enough if it carries the decision.

The reason should be short enough to scan and specific enough to challenge. “Strong fit” is not a reason. “The product page names finance teams, the company sells recurring software, and the current employee count is within the agreed range” is much more useful.

The unresolved column is equally important. It tells the reviewer where judgment is still needed instead of hiding the uncertainty inside a paragraph.

Test it on an awkward list

A list of obvious fits proves very little. The first test should make the agent uncomfortable.

Include:

• several clear target companies;
• several obvious non-fits;
• adjacent companies that use similar language;
• a company with a confusing name or domain;
• a company whose website does not answer an important question;
• at least one case that previously caused disagreement in the team.

Use a list the sales team was going to review anyway. Give the agent the same target brief a person would receive, then ask the salesperson who needed the work to use the output.

Look closely at the mistakes that could waste sales time:

• Did it research the wrong company?
• Did it call a company a fit from an industry label alone?
• Did it cite a source that does not support the reason?
• Did it convert missing evidence into a positive decision?
• Did it mistake a possible signal for buying intent?

One average score can hide all five. Record the important false positives and false negatives separately. Then measure the practical result: how long the original research took, how long review now takes, and whether the salesperson can use the list without reopening every company.

The agent is helping only when the time saved is greater than the time spent checking it.

The first version is smaller than most teams expect

A useful first version can be described in six lines:

You provide: target criteria and a real company list.
The agent checks: company identity and the exact evidence needed for each criterion.
It returns: work now, skip, or take a closer look.
Every decision includes: a short reason and the supporting sources.
It stops: when identity is unclear, sources conflict, or required evidence is missing.
A person decides: which accepted accounts sales works first and how to approach them.

Keep broader sourcing outside this version. Leave contact research, copy, sending, and automatic CRM changes for later. Each one introduces another kind of error and another person who depends on the result.

Once the account list is consistently useful, the next addition should follow the work. If sales needs relevant contacts next, add contact research. If the list exists because a signal appeared, carry the signal and its source into the account record. If the result goes into the CRM, make sure the reason for the decision travels with it.

The account-selection agent can then become a reliable input to a larger sales system.

Start with next week's list

Take the list someone already needs next week. Write down the target criteria. Decide what counts as enough evidence. Add the awkward cases. Then see whether the agent can return a list the salesperson is willing to use.

Extend it only after the salesperson is willing to use the output. If they are not, fix the decision or evidence rules before adding another integration.

Building one for your team?

Send me the company list and the criteria you're using. I'm happy to take a look and tell you where I'd start.

Email me