Working capital · Native AI · IDP + 22 surfaces

22 moments where AI operates inside your treasury programme

This is not "AI everywhere", it is AI applied at the points where a decision needs more than a dashboard. Catalogued, auditable and honest about what is Claude and what is deterministic heuristic.

Try it at app.advanta.pt → See the 22 surfaces ↓
22
distinct surfaces, in production
~30%
calls to Claude Sonnet 4
~70%
deterministic heuristics
€15/mo
typical LLM cost
Interactive demo

Test the rate-suggestion engine in real time

Adjust the 3 parameters on the left; the logistic model computes the optimal offer, with no network calls, 100% deterministic.

Inputs
€100,000
€5k€500k
D+75
15d120d
Medium · 65% acceptance
Cold · 30%Warm · 95%

Click above for an analysis of why these numbers make sense.

AI recommendation
Pay the supplier at D+30 instead of D+75. With a discount of 1.40% (€1,400) on the invoice.
Net yield for your company
€1,190
after 15% Advanta
Equivalent APR
11.4%
vs 2% at the bank
Prob. of acceptance
71%
logistic model
Supplier receives
€98,600
instead of €100,000 at D+75
Your company captures
€1,400
45 fewer days of idle capital
Advanta charges
€210
15% of yield · €0 fixed

Deterministic algorithm · logistic regression with platform priors. See the full maths →

Full catalogue

Each surface does one thing, well

Grouped by moment in the flow: what comes in, what gets decided, what goes out and what gets reported. Each card marks whether the AI is LLM (Claude) or det. (deterministic).

01Input · turning documents into dataLLM dominates
1.1LLM

Invoice extraction

Drag in a PDF. Claude vision reads supplier, tax ID, totals, IBAN, ATCUD, line items and dates. Returns confidence per field + detected anomalies.

In /invoices/review
1.2LLM

Contract intelligence

Master agreements read by the AI. Extracts payment terms, early-discount clauses, penalties, jurisdiction and red flags for legal.

Tab on each supplier
1.3det.

Supplier de-duplication

Exact match on (tax ID + country), Jaccard on company names, email domain. Suggests merges at upload, avoiding yield fragmented across duplicates.

In the upload flow
1.4det.

Spend categorisation

15 canonical categories (COGS, OpEx-IT, Logistics, CapEx…). Rules compiled over vendor name + sector + amount band. Manual override always available.

In /invoices
02Decision · how much, to whom, whendeterministic + LLM explanation
2.1det.

Rate-suggestion engine

Logistic regression anchored in each supplier's history. Grid-search over (discount, days early) that maximises expected yield × p(acceptance).

On each invoice
2.2det.

Opportunity scanner

Applies the same model to every eligible invoice. Ranked by expected €. Hero card on the dashboard: "€420k of yield available this week".

In / (Overview)
2.3det.

Acceptance forecast

Each pending offer gets a probability. Aggregate: "8 of 12 projected to accept · €42k of expected yield". Risk surfaced early.

In /offers
2.4det.

What-if simulator

"What if I raise the target discount to 1.5%?" Interactive sliders · logistic model from your history · projection of yield, acceptance and cash deployed.

In /policy
2.5LLM

Policy recommender

A 4-question mini-flow (industry · AP volume · typical term · risk). Claude proposes initial parameters with a rationale. Deterministic fallback with no LLM.

In /policy
2.6det.

Risk score on approvals

Each four-eyes offer gets 0–100. 11 weighted factors: sanctions, recent IBAN, outlier amount, new supplier, decline streak, fraud signals. Claude explains why.

In /approvals
03Security · catching what goes wrong, earlySQL + LLM narration
3.1det.

Anti-fraud graph

4 SQL detectors: IBAN changes, IBAN collisions across suppliers, threshold-gaming, disproportionate first invoices. Each signal with an idempotent dedup_key.

In /risk
3.2LLM

Alert explainer

Each banner has an "Explain" button → Claude reads the contextual payload and returns 2–3 sentences on the cause + 1–3 actionable recommendations. Server-side cache avoids re-tokens.

On any alert
3.3det.

Invoice anomalies

Duplicate doc numbers, 10× outliers, tax-ID mismatch, out-of-hours uploads, round figures. Surfaced in the upload modal before promoting to AP.

In invoice upload
3.4det.

Treasury pacing

Linear extrapolation over day-of-month vs cap. Urgent banner when the projection exceeds the cap before day 25. Button to adjust the policy.

In /treasury
3.5det.

Supplier health

0–100 score per supplier. Compares the recent 6 weeks vs the previous 6. Trend (improving / stable / degrading / dormant) with visible drivers.

In /suppliers
3.6det.

Re-quote signal

When a supplier's accepted median is ≥30bps above what we typically offer, it suggests a re-quote. Yield left on the table, made visible.

In /suppliers
04Output · the AI writes, the human sendsLLM dominates
4.1LLM

Email writer

4 types: concrete offer · re-quote · follow-up · onboarding. PT / ES / EN. Uses the supplier's real history (typical rate, last decline). Editable output.

In send-offer + supplier detail
4.2LLM

CFO reports

Quarterly review · CSRD/ESG narrative · weekly audit summary. ~500 words, structured sections, anchored in real data. Board-ready.

In /reports
4.3LLM

Audit-log summary

Weekly summary of platform events for the compliance officer. ~250 words. Identifies changes, risk signals and recommended actions.

In /audit
4.4LLM

Natural-language search

Command palette ⌘K · "invoices due this week above €10k" · "top 10 suppliers by yield". Deterministic heuristics + Claude on the long tail.

Global · ⌘K
4.5LLM

Spend insights

Breakdown by category with Δ vs the previous quarter · top vendors · an "Analyse with AI" button produces a 2-3 sentence synthesis identifying the changes.

In /reports
4.6det.

Review-queue triage

"Approve all with high confidence" in a batch. "Reject suspicious ones" in a batch. The decision is always human, but the AI pre-filters. Audit log preserved.

In /invoices/review
How we think about AI

AI earns its place when a deterministic alternative is genuinely impossible or 10× worse

Most of what looks "AI-shaped" is a SQL aggregation in disguise. The list below is the subset where Claude wins on merit.

LLM only where rules fall short

Unstructured text, narrative multi-factor reasoning, natural-language generation or fuzzy pattern-matching at scale, that is where Claude dominates.

Everything else is logistic regression, SQL rules, moving averages. Faster, cheaper, reproducible.

Every call is auditable

Every call to Claude is logged in the corp_ai_drafts, corp_ai_explanations and corp_ai_narratives tables with the full prompt, tokens consumed and model used.

If a narrative comes out wrong, it is forensically reconstructible.

The human decides, the AI accelerates

The AI suggests, classifies, narrates, never executes autonomous financial decisions.

Creating an offer, approving a payment, changing a policy: all require human confirmation. Decisions above the four-eyes threshold require two distinct approvers.

Deterministic fallback always

Every LLM surface has a fallback path. If the Anthropic key fails, or if you want to disable it, the platform keeps working with heuristics.

Zero vendor lock-in. AI is an acceleration, not a critical dependency.

Cost

Single-digit euros a month on LLM

For an active corporate, in practice.

Typical volume: 500 invoices/mo · 200 offers sent · 10 alert explanations/week · 50 emails drafted/mo
€8/mo
invoice extraction
~$0.01 / invoice
€2/mo
email drafting
~$0.005 / draft
€5/mo
explanations + narratives
cached server-side
Total: €15–30/mo in Anthropic calls, less than 0.5% of the platform's typical revenue.

Ready to see AI in production?

5 minutes to create an account · no card · no minimums · AI active from second zero.