How to Build an Explainable AI Account Scoring Agent (2026 Guide)
Black-box AI scores waste sales time. An evidence-first ICP agent scores 1,000 accounts for about $63: Gemini 4 Hydrogen handles bulk research at $1.50/M input tokens, Carbon reviews the risky 10% at $4/M. Every score ships with source quotes and confidence.
Stop Buying Black-Box AI Account Scores
At published rates, researching 1,000 accounts could cost about $63 in model fees. That excludes search, storage, and CRM costs.
An ICP fit agent researches accounts, verifies claims, and applies your scoring rules. It also shows why each account earned its score.
That last part matters.
A mystery score of 87 isn't useful. A score of 87 backed by six sources is useful.
Step 1: Turn Your ICP Into Testable Claims
Most account scoring starts with fuzzy criteria.
"Growing company."
"Modern tech stack."
"Strong market fit."
Those aren't scoring rules. They're opinions wearing business casual.
Your ICP fit agent needs criteria it can prove or reject.
Start with five categories:
| Category | Weight | Testable question |
|---|---|---|
| Firmographics | 30 | Does the company meet your employee and revenue range? |
| Use-case fit | 25 | Does it have the problem your product fixes? |
| Buying signals | 20 | Did it announce funding, hiring, expansion, or a new leader? |
| Technology | 15 | Does it use tools related to your product? |
| Disqualifiers | 10 | Is it a competitor, bad geography, or wrong business model? |
Every criterion should return one of three states:
- Verified true
- Verified false
- Unverified
Never convert "unverified" into "false."
That mistake can remove good accounts because the model couldn't find one page.
Keep fit and confidence separate.
An account may score 92 on verified facts. Its confidence may still be 54% because half its criteria are unverified.
Use fixed scoring rules after research:
> Fit score: Sum of verified criterion weights. > Confidence: Supported weight divided by applicable weight. > Evidence quality: Average source grade across supported claims.
Don't ask Gemini to make up a score from instinct. Ask it to extract evidence.
Your rules calculate the score.
Credit scoring taught this lesson years ago. A number without reason codes creates arguments.
Revenue teams need reasons they can inspect.
Expected outcome: A written scorecard with fixed weights, pass rules, and disqualifiers.
Step 2: Make Every Claim Carry a Receipt
Evidence-first AI treats every claim as a record with rules around it.
The evidence manifest is the source of truth. The account summary is a report built from it.
Each claim should include:
| Field | Purpose |
|---|---|
| `claim` | The exact fact being evaluated |
| `status` | Verified true, verified false, or unverified |
| `source_url` | The page supporting the claim |
| `source_title` | Human-readable source name |
| `source_date` | Publication or update date |
| `quote` | Exact supporting text |
| `source_grade` | First-party, trusted third-party, or weak |
| `confidence` | Confidence in the evidence match |
| `checked_at` | When the agent checked it |
| `failure_reason` | Why verification failed |
The quote matters.
A URL alone turns verification into a scavenger hunt.
Your agent also needs source rules.
Use this order:
1. Company website and regulatory filings 2. Named news outlets and official databases 3. Trusted software directories 4. Job boards and employee profiles 5. Search snippets and scraped lists
Search snippets are leads, not evidence for final scoring.
The verification layer should run fixed checks:
- Did the page return a successful status?
- Does the saved quote exist on the page?
- Is the source recent enough?
- Does the quote support the claim?
- Did another source contradict it?
- Is the page about the correct company?
Store the raw page text with a timestamp. Pages change.
A 2026 governance paper calls this version-pinned lineage. Approval should attach to an exact artifact version instead of the latest output.
Your pipeline should look like this:
> CRM or account list > ↓ > n8n research queue > ↓ > Browser and search tools > ↓ > Gemini 4 Hydrogen > ↓ > Schema and citation checks > ↓ > Evidence store > ↓ > Fixed score calculation > ↓ > Gemini 4 Carbon exception review > ↓ > Human approval or CRM update
We use n8n for workflows like this. Zapier gets painful when branching and retries matter.
Expected outcome: Every scored criterion includes evidence or a clear verification failure.
Step 3: Use Hydrogen for Cheap Batch Research
Gemini 4 Hydrogen is the worker model.
Google prices Hydrogen at $1.50 per million input tokens and $7.50 per million output tokens. That price covers its full one-million-token input window.
Google also says Hydrogen supports up to one million output tokens. Keep outputs short. They're cheaper and easier to check.
Give Hydrogen one narrow job:
> Research this company against the supplied ICP criteria. > Return only claims supported by quoted evidence. > Mark missing facts as unverified. > Never infer revenue, employee count, technology, or intent. > Attach one source URL and one exact quote per verified claim. > Report contradictions without resolving them. > Return the required schema.
Use `low` thinking for simple extraction.
Use `medium` for multi-page research.
Use `high` when the company has naming conflicts or thin data.
Reserve `xhigh` for difficult exceptions. Google offers all four levels in Antigravity.
Don't send every account to `xhigh`. It wastes time and tokens.
Run accounts in batches through n8n. Save each research result before scoring.
For 1,000 accounts, assume:
- 20,000 Hydrogen input tokens per account
- 2,000 Hydrogen output tokens per account
- 20 million total input tokens
- 2 million total output tokens
The estimated model cost is:
- Hydrogen input: 20 million × $1.50 = $30
- Hydrogen output: 2 million × $7.50 = $15
- Hydrogen total: $45
That estimate excludes browser tools, search APIs, storage, and CRM fees.
Google says Hydrogen ran 2 to 2.8 times faster in partner tests. Token use fell 12% to 23% compared with Gemini 3.8 Flash.
Those results came from Google's tests. Your account research may perform differently.
Expected outcome: Most accounts receive a cited first-pass score for pennies.
Step 4: Send Only Exceptions to Carbon
Gemini 4 Carbon is the reviewer.
Google built Carbon for deep reasoning and long verification loops. Use it when Hydrogen finds conflicts or incomplete evidence.
Route an account to Carbon when:
- Confidence falls below 75%
- A must-have criterion remains unverified
- Two credible sources conflict
- The company name matches multiple entities
- A citation quote fails validation
- A high-value account sits near your qualification threshold
- A disqualifier depends on unclear evidence
Give Carbon the evidence manifest and raw sources.
Don't ask it to restart research without a reason.
Its job is to challenge Hydrogen:
> Review the evidence manifest against the raw sources. > Reject claims without direct support. > Resolve conflicts only when one source has stronger authority. > Preserve unresolved conflicts. > Explain every changed status. > Return the corrected manifest and review notes.
Google announced Carbon at an introductory $2 per million input tokens. Output cost $10 per million tokens.
After November 30, 2026, public Carbon pricing rises to $4 per million input tokens and $20 per million output tokens.
Assume Carbon reviews 10% of 1,000 accounts:
- 100 exception accounts
- 30,000 input tokens each
- 3,000 output tokens each
- 3 million total input tokens
- 300,000 total output tokens
At public pricing:
- Carbon input: 3 million × $4 = $12
- Carbon output: 300,000 × $20 = $6
- Carbon total: $18
Hydrogen plus Carbon costs about $63 in this example.
Google's Orchestrator uses the same model split. Carbon plans and handles unclear cases. Hydrogen handles faster execution.
On Terminal-Bench 4.1, Orchestrator used 24% less quota and finished 19% faster than Carbon-heavy work. Account research differs from coding, but the routing idea still applies.
Expected outcome: Carbon reviews the risky 10% instead of every routine account.
Step 5: Test the Evidence Before Shipping Scores
Don't judge your agent by how polished the summary sounds.
Judge the receipts.
Build a labeled test set of 200 accounts. Include good fits, bad fits, private companies, subsidiaries, and companies with common names.
Track these metrics:
| Metric | Calculation | Suggested launch gate |
|---|---|---|
| Citation validity | Working citations ÷ total citations | 98% |
| Claim support | Supported claims ÷ verified claims | 95% |
| Evidence coverage | Criteria with evidence ÷ applicable criteria | 90% |
| Unsupported claim rate | Unsupported claims ÷ all claims | Under 1% |
| Exception rate | Carbon reviews ÷ total accounts | Under 20% |
| Human agreement | Matching human and agent decisions | 90% |
| Cost per account | Total research cost ÷ accounts | Track weekly |
These are suggested gates, not published Gemini benchmarks.
Test the whole pipeline after every prompt or model change. Models often change faster than CRM rules.
Salesforce surveyed 24 Sales Cloud customers in 2026. Those customers reported six saved hours per rep each week.
Salesforce also found that 63% reported better CRM accuracy. Two-thirds reported more thorough account research.
The sample was small and run by the vendor. Treat it as directional evidence.
Bad research has a clear cost.
Paligo surveyed 411 business leaders. Seventy percent reported a materially wrong AI result at least once a month.
Among respondents who found errors, 51% spent five or more employee hours fixing them. Only 22% could always identify the source version used.
GWI surveyed 1,883 senior professionals. Fifty-eight percent acted on AI insight that later proved wrong.
Only 36% verified AI output most of the time.
The problem is usually the system around the model.
Calling it a hallucination makes the failure sound unavoidable. Citations, checks, and exception routing give you ways to catch it.
Expected outcome: A score you can defend, audit, and improve.
FAQ
What is explainable AI account scoring?
Explainable AI account scoring ranks companies with visible rules and cited evidence. Each score includes supporting claims, source quotes, confidence, and unverified facts.
What is Gemini 4 Carbon, and how much does it cost?
Gemini 4 Carbon is Google's deeper reasoning model for long-running and difficult tasks. Public pricing is $4 per million input tokens and $20 per million output tokens after November 30, 2026.
How do I make AI account scoring explainable and auditable?
Store each claim with an exact quote, source URL, source date, confidence, and timestamp. Keep unverified facts separate, then calculate scores with fixed rules outside the model.
How should I configure Gemini 4 thinking levels for agents?
Use `low` for extraction, `medium` for normal research, and `high` for conflicting evidence. Use `xhigh` only for difficult reviews because deeper reasoning adds latency and token use.
Why use Hydrogen and Carbon together?
Hydrogen handles high-volume account research at $1.50 per million input tokens and $7.50 per million output tokens. Carbon reviews conflicts, weak citations, and high-value exceptions at $4 input and $20 output per million tokens.
Related Reading
How much does it cost to AI score 1,000 accounts using Gemini 4 Hydrogen and Carbon?
Scoring 1,000 accounts costs about $63 in model fees. Hydrogen handles 90% of accounts for $45. Carbon reviews the riskiest 10% for $18.
What is the difference between Gemini 4 Hydrogen and Carbon pricing?
Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens. Carbon costs $4 per million input tokens and $20 per million output tokens after November 30, 2026.
How do you make AI account scores auditable?
Store each claim with an exact source quote, URL, source date, and timestamp. Calculate scores with fixed rules outside the model. Keep unverified facts separate from verified ones.