Vithali GroupIlluminating ExcellenceTwo clocks are running on your margin. One of them is legislated.
We install an air-gapped AI workforce inside the ERP you already run. It reads manufacturer statements and credit notes, reconciles them against your contract tiers, and drafts the claim files your finance team signs — while the same data layer captures the batch, expiry and serialisation record SAHPRA now requires. Zero external API calls. Nothing written to your ledger without a named approver.
15 minutes. No deck to sit through, and nothing to prepare — no ERP login, no statements.
Pharmaceutical margin is not lost in the warehouse. It is lost in the reconciliation — and the deadline for fixing the data layer is already in law.
Two things are moving at once, and they need the same plumbing. One is a regulatory deadline with published dates. The other is a contract cycle that closes whether or not you were ready.
The phase you are in now. A systems question disguised as a labelling one.
Every receiving, put-away, pick and dispatch event must capture structured product data.
Aggregation between packaging levels, with data exchange agreed with trading partners.
Pallet and carton identity, not just product identity.
Unit identity at the item, not the batch.
Source: SAHPRA Guideline on Traceability of Health Products (SAHPGL-INSP-RC-04, v1) and the HPA16-2025/26 implementation roadmap, September 2025. A distributor whose data layer is spreadsheets and manual capture does not arrive at 2027 compliant — and the fix is the same data layer the rebate and invoicing work needs. That is why the compliance question and the margin question are one project.
Your margin on the medicine is capped by the Single Exit Price framework and a regulated logistics fee. Everything you keep beyond that is earned in the layers the regulations do not price: recovery, data accuracy, and the cost of the people rekeying documents.
Five places the work sits. None of them is a percentage we can put on your business without seeing it.
Every distributor runs some version of these five processes. Whether they leak, and by how much, is a question about your contracts, your portals and your team — not a benchmark we can quote at you.
| Process | What actually happens | What it looks like when it goes wrong |
|---|---|---|
| Rebate & chargeback reconciliation | Contract tiers, growth incentives and chargeback credits tracked across several manufacturer portals, each with its own statement format and cycle. | A credit note arrives in month four for a claim nobody raised in month one. It is matched by hand, or it is not matched at all, and the window closes. |
| Invoice & document intake | Supplier invoices, packing lists, batch certificates and delivery notes keyed into the ERP by people reading PDFs and scans. | An expiry date keyed as the batch date. Found three weeks later in a reconciliation, with stock already picked against it. |
| Short-dated & expired stock | Expiry lives in the WMS or on the carton, not in a layer that compares it to forward demand. | The write-off is discovered at picking, when redistribution to another site is no longer possible and the return window has passed. |
| Replenishment & working capital | Safety stock held against supplier lead time and unpredictable ordering across sites. | Stock-out on a chronic line at one branch and overstock at another, in the same week, both justified as caution. |
| Traceability & audit evidence | GDP, POPIA and traceability records assembled by hand when an auditor, a client or SAHPRA asks. | Three weeks of someone's December, reconstructing what the system should have captured as it ran. |
We do not publish a sector leakage percentage. We have not measured your business, and neither has anyone else quoting you one. Phase 0 produces your number from your data — and if it does not justify the programme, we will tell you that in writing.
Three objections end this conversation before it reaches procurement. We built the system around them.
“POPIA will not allow it.”
Your risk committee is right to refuse a cloud service that sends price files and patient-linked dispensing records to a public API. So we do not use one. Models run on your infrastructure or in an in-country private cloud. There is no egress path to remove later — it does not exist in the architecture.
Answered in the architecture“No AI writes to my ledger.”
Correct. Every output is a document a named human signs: a claim file, an exception queue, a suggested purchase order. Nothing touches the general ledger, the price file or the stock ledger without a recorded approver — and the governance layer can stop any agent mid-task.
Human gate on every write“Nobody has proven this in our sector.”
Fair, and we will not pretend otherwise. That is why week one is a paid diagnostic that either produces a rand-denominated model from your own statements or tells you not to proceed. The guarantee is written against what we control — identification, evidence and drafting — not against a manufacturer's decision.
Proven on your data firstWhat lands on your desk is a document with a person's signature on it.
Not a dashboard you have to interpret. A claim file per manufacturer, and an exception queue for everything the system refused to guess at. This is the format — illustrated with neutral values, not client data.
| MANUFACTURER | CONTRACT TIER | STATEMENT CREDITS | CLAIMED TO DATE | VARIANCE | EVIDENCE | STATUS |
|---|---|---|---|---|---|---|
| Manufacturer A | Tier 3 · 4.5% | R— | R— | unmatched | 3 credit notes | Draft for approval |
| Manufacturer B | Tier 2 · 3.0% | R— | R— | reconciled | statement + PO | Ready to submit |
| Manufacturer C | Growth incentive | R— | R— | no statement | portal export | Exception queue |
It shows its working
Every line points back to the statement, the contract clause and the purchase record it came from. If a number cannot be evidenced, it does not go in the claim — it goes in the queue.
It fails loudly, not quietly
Schema-constrained extraction means a document that does not parse is rejected and queued. The exception queue is a first-class surface in the product, not an error page.
It improves every cycle
Each approved claim and each cleared exception teaches the model your manufacturer's format. By the second quarter, the queue is mostly the genuinely ambiguous documents.
Five layers. None of them is a chatbot.
Each layer has a job the others cannot do. Read the right-hand column if you only want to know what it produces.
Workflow automationDeterministic orchestration
API triggers, SQL calls, file watchers and ERP/WMS synchronisation. All arithmetic and matching logic lives here, where it is testable and repeatable — not in a model that might answer differently tomorrow.
Worker agentsPersistent task execution
Agents that log into manufacturer and distributor portals, pull statements and transaction history, reconcile against contracted tiers, and draft the claim. This is the three days a month nobody in finance wants to spend downloading PDFs from five portals.
The intelligence, on your floorLocal open-weight models
Models served on your hardware or in an in-country private cloud. Pricing sheets, supplier contracts and scheme claims never leave the perimeter, because there is nowhere for them to go.
The domain trainingFine-tuning on your schema
Adapted to your SKU master, manufacturer contract language, SAHPRA registration data and ERP table structures, with schema-constrained output. A generic model guesses at a product code. A tuned one lands in a valid field or fails into the queue.
The controlsGovernance portal
Per-role token budgets, approval queues, a full audit trail of every agent action and every human decision, and the switch that stops any agent cold. This is the layer that lets an IT director sign, because it proves what the system did.
Layers 1 and 2 produce the recoverable value. Layer 3 removes the POPIA objection. Layer 5 removes the fear of autonomous software touching the ledger. Layer 4 is what makes the output usable rather than impressive.
“Your data never leaves the building” is a policy until it is a network diagram.
Here is ours. Every arrow runs inside your perimeter, and the last step is a human, by design.
NO PROMPT, PRICE FILE OR ORDER DATA CROSSES THIS BOUNDARY. THE MODEL ENDPOINT IS NOT ROUTABLE FROM THE PUBLIC INTERNET.
Models and orchestration on your own hardware, inside your firewall. Maximum control, no external dependency, GPU capacity required.
AWS Africa, Cape Town (af-south-1), in a dedicated tenancy. Same isolation, no hardware to buy, data residency in South Africa.
Orchestration on-premise, inference in-country, with the egress rule enforced at the network layer rather than in a policy document.
The five controls
- Zero external inference calls. Not a promise in a contract — an absent route in the firewall. Verified in the security audit and evidenced in the architecture note.
- Pseudonymisation at ingest. Where patient-linked data appears — script-level dispensing, scheme claims — identifiers are stripped before the model sees the document.
- Schema-constrained output. A failed extraction is rejected and queued, never guessed.
- Named approver on every write. No general ledger, price file or stock ledger entry without a recorded human decision.
- Full audit trail. Every action logged with input, model version, output, confidence and approver — exportable for internal audit, SAHPRA inspection or client due diligence.
What we will not do
- Replace your ERP, or take custody of your data.
- Send a price file, a contract or a patient-linked document to a public API.
- Write autonomously to your ledger, price file or stock ledger.
- Promise a forecast accuracy figure before we have seen your history.
- Present a validation pack as a substitute for your Responsible Pharmacist's obligations under GMP/GDP.
Delivery includes a POPIA architecture note and processing agreement, plus a validation pack — user requirement specification, installation and operational qualification templates, change-control procedure — mapped to your existing quality system.
Bounded, sequenced, and reversible at every phase.
Each phase has an exit criterion and a stop point. Starting one does not commit you to the next.
Phases 2 and 3 are where the next rand comes from. The sequence is set by which leak your own data shows to be largest — not by what is easiest for us to build.
Twice your implementation fee, identified and drafted, in 90 days.
recovered or refunded
If we cannot, we keep working for free — or you get the fee back.
If our agents and automation do not identify, evidence and draft at least twice your implementation fee in recoverable rebate, chargeback and credit-note value within 90 days — R560,000 to R900,000 depending on scope — we continue engineering and optimising at no additional fee until they do, or we refund the implementation fee in full.
What this deliberately excludes: value that depends on a manufacturer accepting a claim we have correctly prepared and submitted, or on your team clearing the approval queue. We control identification, evidence and drafting. We do not control third-party decisions or internal capacity — and a guarantee that pretends otherwise is marketing, not risk transfer.
We report what we found. Not what we projected.
- Week 1The baseline is your numberTwelve months of purchase data, read against your contract tiers, produces a rand value per leak type. You approve that baseline before any build starts — it becomes the number the programme is judged against.
- WeeklyThe working fileClaims drafted, value submitted for approval, exceptions cleared, queue age. Four lines, not a dashboard you have to interpret.
- Day 90The value reportWhat was identified, what was evidenced, what was approved and submitted, and what remains in the queue — signed off by your finance manager, not by us.
Why there is no percentage on this page
Every vendor pitching this sector quotes a leakage rate — two percent, four percent, sometimes a rand figure pulled from a turnover assumption. None of them have measured your business, and a number you cannot trace to your own statements is a number your CFO will rightly discard.
So the page sells the mechanism and the diagnostic sells the number. That is the honest sequence, and it is also the faster one: the baseline exists in week one, before you have committed to an implementation.
Eight questions. The answers set scope, price and which agent goes first.
You do not need the answers to book the audit. You need one of them — question 5 — to make the call worth having.
One bundle, itemised at what we would charge for it separately.
Automation layer, ERP connectors, worker agents for statement reconciliation, and local POPIA-compliant model deployment on your infrastructure.
Technical and compliance assessment confirming zero leakage paths, plus the architecture note your risk committee will ask for.
Models trained on your SKU database, manufacturer contract language and ERP table structures.
Dashboards, per-role token budgets, approval gates and audit export.
Reads inbound RFQs and quote requests from email — including a photographed specification — extracts the line items, matches them against your ERP price file, and drafts the quotation with the compliance pack attached. Your desk reviews and sends. The agent never issues a price on its own.
Bundled into the implementation fee below.
Fixed fee, two weeks, credited against implementation if you proceed.
One-off build, scoped by entities, ERP instances, agents and data volume.
Governance portal, model monitoring, automation maintenance, monthly value reporting.
In place of part of the retainer, linked to verified recoveries.
Two distributors per quarter. That is a staffing fact, not a countdown timer.
Implementation requires an engineer embedded in your data, your ERP schema and your exception queue. We will not dilute that across four simultaneous builds, so we do not take more than two. Onboarding windows open on the first business day of each quarter, because integration work needs uninterrupted engineering time.
2 OF 2 OPEN
WAITLIST
Confirmed on the call — the counter moves when a build starts.
Who this is for, and who it is not.
- Pharmaceutical wholesalers, healthcare logistics operators and distributors with manufacturer rebate or chargeback agreements
- Syspro, Sage Evolution, SAP Business One or Infor as the system of record
- Multiple manufacturer portals, and a finance function that suspects it is losing credits but cannot prove how much
- A risk committee that has blocked a cloud AI proposal in the last 24 months
Not for you if
- You are a single pharmacy or a retailer buying from a wholesaler
- You want a staff chatbot, not a reconciliation engine
- You need the general ledger written to autonomously, with no human gate
- You will not provide read-only ERP access and one finance person for two hours a week
Why Vithali Group is the entity you sign with.
Because procurement has to sign it. Vithali Group is a 2013-registered supplier with a CSD number, a current tax compliance PIN, CIDB registration and a B-BBEE Level 1 certificate — the paperwork your vendor onboarding process actually checks. Engineering delivery sits with Vithali Group: the orchestration, worker agents, local inference and governance portal described above are built and supported by the team you contract with. You sign once, with one accountable party who owns the delivery end to end, rather than a reseller standing between you and the engineers.
There is a commercial reason too. As a 100% black woman-owned EME at Level 1, an engagement placed with Vithali Group carries 135% procurement recognition — so a services spend of this size also moves your preferential procurement line, which is more than the software alone was going to do.
The guarantee sits with the entity that signs the agreement, not with a reseller.
At a glance
Fifteen minutes. One number. Then a decision.
- We look at one thing: which manufacturer agreements you hold, and who reconciled them last quarter. No ERP login needed for this call.
- You leave with a written view of the three questions we would need answered to put a rand figure on it — and our opinion on which leak to attack first.
- If it looks worth a diagnostic, we book two weeks. If it does not, we tell you on the call and you keep the notes.
No slide deck. One of the two founding engineers on the call, not a salesperson.
Book the margin audit
15 minutes · video or phone · confidential