How to give an AI agent a reason, not just a number
A score tells an agent where to start. The factors behind it tell the agent what to say, what to check, and when to hold back.
A score tells an AI agent where to start. It does not tell the agent what to say, what to check, or when to stop. Those come from the reasons behind the score, and most setups never pass them along. The agent gets “7.8” and is left to work out the rest.
This is a how-to. It covers what a reason is made of, three before-and-after examples of the same agent with and without one, and the few lines of instruction that make an agent use what it is given.
What a reason is made of
“The reasons behind the score” is easy to say and vague to build. For an agent to use them, each factor needs to carry five things.
| Part | What it is | Why the agent needs it |
|---|---|---|
| The factor | Its name, in plain words | So it knows what the number is about |
| The value | How the record rated on that factor, out of 10 | So it can tell a strong point from a weak one |
| The weight | How much that factor counts | So it does not open with a detail that barely matters |
| The contribution | Value times weight | So it can find the top reason without doing math |
| Empty or not | Whether the factor is 0 because the field is blank | So it can tell “unknown” from “poor match” |
Add one thing for the record as a whole: the time the score was last calculated. A reason from this morning and a reason from last quarter should not be treated the same.
Example 1: what to say
The agent’s job is to write the first line of an email to a contact at a demo account.
| What the agent is given | First line it writes | |
|---|---|---|
| Before | Score: 7.8 | “I came across your company and was impressed by your recent growth.” |
| After | Score: 7.8. Top factor: industry, 9.3 at 30%. Next: contacts in buying roles, 8.1 at 25%. | “We work with a lot of regional freight brokers, and I can see three people on your team who run carrier onboarding.” |
The first line is what an agent writes when it has to invent a reason. It is polite and it could be sent to anyone. The second is specific because the agent was told what is specific about this account.
Example 2: what to check
A demo account scores 5.2. The agent’s rule is to send a short email to anything from 5.0 to 6.4.
| What the agent is given | What it does | |
|---|---|---|
| Before | Score: 5.2 | Sends the short email. The rule says to |
| After | Score: 5.2. Tech stack: 0, field is empty, weight 20%. Other factors: 7.4 and up. | Does not send. Adds the account to the enrichment list and names the empty field |
Without the reason, 5.2 reads as a middling match. With it, the account turns out to be a strong match with one unknown. If the tech stack comes back as a fit, this account is near the top of the list. The number alone would have spent it on a throwaway email.
Example 3: same number, different instruction
Two demo contacts both score 6.4. An agent with only the number treats them the same.
| Contact A | Contact B | |
|---|---|---|
| Score | 6.4 | 6.4 |
| Account fit | 9.2 | 4.8 |
| Role and seniority | 3.9 | 9.3 |
| Department | 6.1 | 5.7 |
| Region | 8.8 | 8.8 |
| Personal email address (disqualifying) | 1.8 | 1.8 |
| What it means | Right company, wrong person | Right person, weaker company |
| What the agent should do | Do not pitch this contact. Look for a more senior contact at the same account | Write to this contact with a lighter ask. Do not hand to a rep yet |
- Account fit (from the company record) 35% 9.2
- Role and seniority 30% 3.9
- Department 20% 6.1
- Region 10% 8.8
- Personal email address 5% -1.8
The math for contact A: 9.2 at 35% is 3.22, 3.9 at 30% is 1.17, 6.1 at 20% is 1.22, and 8.8 at 10% is 0.88. That adds to 6.49. The disqualifying factor subtracts 1.8 at 5%, or 0.09, for 6.4.
Contact B gets to the same total a different way: 4.8 at 35% is 1.68, 9.3 at 30% is 2.79, 5.7 at 20% is 1.14, and 8.8 at 10% is 0.88, which adds to 6.49, less the same 0.09.
This is the clearest case for reasons over numbers. A 6.4 is not one situation. It is at least two, and they call for opposite moves.
The instruction to give the agent
Passing the reasons is half of it. The agent also has to be told to use them. Four lines cover most cases.
- “Before you act on a record, read its score and every factor with its value, weight, and contribution.”
- “Use the factor with the largest contribution as the reason for outreach. Put it in plain words. Do not mention the score or that the record was ranked.”
- “If any factor is 0 because its field is empty, do not contact the record. Report the empty field.”
- “If the weakest factor is about the person and the strongest is about the company, look for a better contact before writing.”
Keep the list short. Each rule you add is one more thing to test, and one more rule that can pull against another. With four, you can read a sent email and tell which rule produced it. With a long list, you cannot.
Put the reasons where the work happens
A reason only helps if it shows up where the decision is made. For an agent, that is wherever it reads the record. For a rep, it is wherever they already spend the day.
| Who needs the reason | Where it should show up |
|---|---|
| An agent working records in the CRM | A field on the HubSpot record |
| A rep working a queue in HubSpot | The same field, on the record they have open |
| A rep who lives in Slack | A message when a record crosses a threshold, with the top factors in it |
| A manager reviewing the week | An email with the records worked and the reason for each |
The agent and the rep should be reading the same reasons. When a rep takes over a record an agent started, the first question is why it was contacted. If the answer is already in front of them, the handoff takes a minute.
Four mistakes to avoid
Passing everything. Do not hand the agent every property on the record and hope it finds the reasons. Give it the factors and their contributions. More fields make the reason harder to find.
Passing words in place of numbers. A tag that calls a record a good match throws away the detail the agent needs. Give it 9.2 and 3.9, not an adjective.
Letting the reason reach the buyer raw. “Your account fit is 9.2” is not an email. The agent should turn the factor into something a person would say.
Ignoring the date. If the score was calculated before the contact changed jobs, the reason is about someone who no longer exists. Have the agent check the calculation time and skip anything too old to trust.
Where BoostIdeal fits
BoostIdeal produces the first half of this: a score out of 10 for each record and the factors behind it. It can share those reasons wherever your rep works, in a HubSpot field on the record, in Slack, by email, and so on, so agents and people read the same thing. You build the model and set the weights, with AI helping you analyze your data and draft. It can score a contact using the fit of its account, as in example 3, and an empty field contributes 0 without counting against the record. Scores recalculate hourly by default, and in near real time with webhook sync.
It does not write the outreach or run the agent. Those stay in your own tools. And it shows a score’s current value and the time it was last calculated, not a history of past values.
If the model is not written yet, start with your ideal customer profile and the five-step framework for building an ICP scoring model.
Frequently asked questions
What is the difference between a score and a reason?
The score is the total. The reasons are the factors that produced it, each with a value and a weight. An agent can sort by the score. It needs the reasons to decide what to say or do.
How much of the breakdown should I give an agent?
The score, each factor with its value, weight, and contribution, whether any factor is empty, and when the score was calculated. Leave out fields that are not part of the model.
Where should the reasons live?
Where the work happens. A field on the HubSpot record for agents and for reps who work in the CRM, and Slack or email for people who work there. BoostIdeal can share reasons in each of those places.
Should the agent tell the buyer why they were chosen?
It should use the reason, not recite it. Lead with what makes the account a match, in ordinary language, and never mention a score.
What if two records have the same score?
Read the factors. Two records can reach the same total for opposite reasons, and the right action differs. That is the main argument for passing reasons along with the number.
Can an agent work out the reasons on its own?
It can guess from the record, and the guess will read well. It will not be the reason your model scored the record, so you will not be able to check it.
The takeaway
Give an agent a number and it can put records in order. Give it the factors, their sizes, which ones are empty, and when they were calculated, and it can write the right first line, hold back the records that need data, and tell two identical scores apart. The reasons are already in your model. The work is getting them to where the agent and the rep can read them, and telling the agent to use them.