Hesabi AI insights: what we ship vs roadmap

Hesabi AI insights are assistive features built on your own ledger data: short cash flow hints, suggested expense categories, and anomaly flags when a pattern looks unusual. They do not replace your expert-comptable, auto-file VAT, or clear TTN invoices without human review. This guide separates what we ship today from roadmap experiments so Tunisian PME owners and accountants can set expectations before turning features on.

We describe production behavior from DarDev's Hesabi stack on hesabi.tn. Hesabi is Tunisia-only—built for local VAT, TEJ-oriented workflows, and TTN El Fatoora paths where your plan supports them. If you are evaluating AI in finance software, start with the same question your cabinet asks: who approves changes, and where does data leave Tunisia?

What shipped today (and what it is not)

Three insight types are in production for eligible Hesabi workspaces. Each is optional, logged, and reversible—you accept or dismiss suggestions rather than having the system silently rewrite books.

• Cash flow hints: rolling outlook based on open invoices, recorded payments, and recurring expenses already in Hesabi
• Categorization suggestions: proposed account or tag for bank lines and expenses when text or counterparty matches learned patterns
• Anomaly flags: alerts when amounts, timing, or duplicate references deviate from your recent baseline

Cash flow hints in plain language

Cash flow hints summarize liquidity pressure over the next few weeks using data you already entered: customer invoices awaiting payment, supplier bills with due dates, and payroll or CNSS-related outflows you scheduled in Hesabi. The model does not pull credit scores or external lending data—it reads your workspace.

For a Tunisian retailer or services PME, the useful output is often boring and valuable: a warning that three large B2B receivables cluster in the same week as a VAT payment, or that recorded expenses exceed typical monthly burn based on the last ninety days. Hints appear as dashboard cards and optional notifications; they are not binding forecasts.

If your team still tracks cash primarily in a notebook or parallel spreadsheet, hints will be incomplete until invoicing and payment recording live in one system. Our first-week dashboard walkthrough covers the setup that makes hints trustworthy—master data, roles, and consistent payment marking matter more than any algorithm.

Categorization suggestions—not magic autopilot

When you import or enter bank movements, Hesabi may propose a category or ledger mapping based on prior choices: same supplier name, similar memo text, or recurring amount bands. Suggestions inherit your history; a new supplier with no precedent returns low confidence or no suggestion.

Accountants should treat proposals like a junior clerk's draft. Accept when correct; correct and accept when wrong so the model learns your firm's chart—not a generic international template. Mis-accepted categories propagate into VAT summaries and TEJ-related exports, so finance should own approval rules, not sales.

1. Enable suggestions for finance roles first; keep sales on draft-only entry
2. Review the first twenty accepted mappings with your cabinet in month one
3. Reject ambiguous proposals instead of guessing—silence is better than wrong TEJ lines
4. Document category decisions in internal notes when regulatory context is non-obvious

Anomaly flags: signal, not accusation

Anomaly detection compares new transactions to your workspace baseline: duplicate invoice numbers, payments twice the usual supplier average, weekend bulk edits, or clearance retries that fail repeatedly. Flags are heuristic—they reduce surprise at month close; they do not prove fraud.

Seasonal Tunisian businesses (tourism, agriculture, pre-Ramadan retail) will trigger false positives until the baseline window includes a full cycle. Finance leads should tune sensitivity with your accountant rather than disabling alerts after the first noisy week.

Accountant trust: how cabinets should evaluate AI

Expert-comptables rightly ask whether AI changes audit trails or exports. In shipped Hesabi insights, every accepted suggestion writes a normal ledger event with user attribution—who accepted, when, and what changed. Rejected proposals leave no footprint beyond an optional dismiss log for admins.

We recommend aligning AI usage with the same monthly rhythm you use for manual exports: week-one close, accountant review, adjustments only inside Hesabi. Our accountant collaboration guide describes RACI and export formats; add one line to your engagement letter stating that AI suggestions require finance approval before they affect declarations.

Cabinet plans on hesabi.tn serving multiple PME clients should not assume cross-client learning: models are scoped per workspace unless we publish otherwise. Your Client A's categories must not bleed into Client B.

Data privacy and residency expectations

Insight features run against data stored for your Hesabi tenant. We do not sell workspace transaction feeds to third-party ad networks, and we do not train global models on identifiable Tunisian taxpayer content without explicit contractual scope. Operational details evolve—check hesabi.tn security pages and your data processing terms for current subprocessors.

Practical PME checklist: restrict admin roles, disable shared owner passwords, export monthly audit logs, and ask your cabinet what they require before enabling any new automation toggle. If a feature sends data outside Tunisia in the future, we will label it in release notes and product settings—not hide it in generic "improvements" copy.

Roadmap (not shipped—subject to change)

Roadmap items below are engineering direction, not commitments with dates. They may shrink, merge, or ship under different names after pilot feedback from Tunisian accountants.

• Natural-language questions over your own reports ("unpaid B2B over 30 days") with cited rows
• Stronger TEJ-oriented categorization templates co-designed with cabinet partners
• Seasonality-aware cash hints that understand Tunisian fiscal calendar peaks
• Batch review UI for accountants managing many PME workspaces
• Optional on-prem or VPC deployment paths for firms with strict data policies

We will not ship autonomous "AI accountant" modes that file VAT or issue cleared TTN invoices without a named human approver. Regulatory risk and accountant liability are not solved by marketing adjectives.

Adoption playbook for the first 30 days

1. Stabilize inputs

   Complete dashboard setup, TTN test clearance if applicable, and two weeks of consistent payment recording.

2. Enable one insight type

   Start with categorization or cash hints—not both—so your team learns review discipline.

3. Weekly finance review

   Spend fifteen minutes accepting, correcting, or rejecting suggestions; log patterns for your cabinet call.

4. Accountant sign-off

   Share sample exports after AI-assisted month close; confirm TEJ and VAT lines match expectations.

5. Expand or pause

   Add anomaly flags only when categorization error rate is acceptable; pause features during audit crunch.

Teams new to cash discipline should read our SME cash flow basics for Tunisian retailers before relying on hints—understanding receivable delay and supplier terms beats any dashboard card.

When to ignore AI features entirely

Skip or disable insights if you are mid-migration from legacy spreadsheets, during your first TTN enrollment week, or while your chart of accounts is still negotiated with a new cabinet. Garbage in produces confident-looking garbage out.

Also ignore vendor demos that promise "zero-touch accounting" for Tunisian PMEs—local VAT, TEJ, and clearance rules need human judgment. Hesabi AI is built to reduce tedium, not eliminate professional oversight.

Explore plans and enable features from hesabi.tn. For implementation help, contact DarDev via dardev.net—product updates land on news.dardev.net when we ship material changes to insight behavior.