AI Platform
The numbers are computed.
Only the words are generated.
Every figure LekhaPro shows you comes out of a pure, unit-tested engine — before any model is involved. The AI layer is handed that finished result and asked to explain it in plain language. Turn the AI off and every number is still there.
That console is the real request path, running. Flip its AI switch off — the routing, the engine and all four figures carry on exactly as they are.
An illustrative demonstration: a question is routed to a copilot by keyword score, a tested engine computes the figures, and only then is a language model asked to describe them. With AI switched off the figures are unchanged.
- 15
- context-routed copilots
- 22
- personas — 9 built-in, 13 from packs
- 25
- registered narration kinds
- 7
- statistical forecasting engines
- 0
- figures written by a language model
One answer, split down the middle
On the left, what a tested engine computed. On the right, what the model was allowed to write about it. Every figure in that paragraph is one of the four beside it — because narration is only ever handed the finished result, never the question.
- Closing cash & bank
- ₹ 24,18,640
- Receivables overdue 90+
- ₹ 6,42,110
- Gross margin, this quarter
- 31.4%
- Projected shortfall, week 6
- ₹ 2,90,000
Pure functions over your own ledger. No network call, no model, no variance between two runs of the same data.
Cash is comfortable today, but the shape of it is not. ₹6.42L has aged past ninety days and is carrying the whole receivable book. Margin held at 31.4%, so this is a collection problem, not a pricing one. On current run-rate week six is short by ₹2.90L — chase the four oldest accounts and the gap closes.
Every figure in this paragraph is one of the four on the left. The model was given the computed result and asked to explain it — never to work it out.
Illustrative sample data. Switch the right-hand panel off and the left-hand panel is unchanged.
An assistant that is allowed to say no
Ask a question in plain English and it is resolved against a governed semantic model — your datasets, your measures, your row scope — before a single line of query is built. Three outcomes are possible, and one of them is a refusal.
Ask about payroll costs and it will not improvise. LekhaPro holds no payroll data, so the question resolves to a dataset that does not exist and the request is rejected with the reason attached. That exact case is a regression test in the suite, which means the refusal is a shipped guarantee rather than today’s behaviour.
An assistant that answers everything is easy. One that declines the question it cannot ground is the one you can put in front of an auditor.
See the analytics platform →- Accept
“Sales by customer for July, top ten by value”
Resolved against the sales dataset. Runs as a governed query — row scope applied, figures computed, not guessed.
- Clarify
“Show me the total”
Runnable in shape but ambiguous: “total” exists in more than one selected dataset, and there is nothing to group by. It asks which, instead of picking one.
- Reject
“What were our payroll costs last quarter?”
unknown dataset: payroll_costs — refused before anything is compiled. LekhaPro holds no payroll data, so there is no honest answer to give and none is invented.
The refusal case is pinned by a test, so it cannot regress into a plausible-sounding answer in a later release.
What actually happens when you ask
Five steps. Four of them are arithmetic and run with the AI switched off. Only the last one involves a language model, and it never sees your question without the answer already attached.
- 01
Route
Your question is scored against each copilot’s keyword list and the highest match wins, ties broken by registry order. Choosing the lens is arithmetic, so the same question always lands in the same place.
No model involved - 02
Scope
Only the lenses belonging to packs you have switched on are even considered, and row-level scope is applied at the database. A lens you do not own is unreachable by construction, not by instruction.
No model involved - 03
Compute
A pure, unit-tested engine produces the structured result — balances, ageing, variance, cover, forecast. Same inputs in, same numbers out, every time.
No model involved - 04
Validate
For analytics questions the proposed query is checked against your governed semantic model before anything is compiled or run. It can accept, ask for clarification, or refuse outright.
No model involved - 05
Narrate
Only now does a model see anything — the finished result, under an instruction to use those figures and invent none. This is the one optional step, and the only one that leaves your tenant.
Optional · uses the gateway
The lens is chosen by arithmetic, not by inference
You do not pick a copilot from a dropdown and no reasoning budget is spent working out which one you meant. Your question is scored against each copilot’s keyword list and the best match wins — instantly, identically, every time. Five are always on. The other ten appear only when you switch on the industry pack they belong to.
Business Copilot
The default lens when nothing more specific matches.
AI CFO
Cash, profit, margin, GST, loans, expenses, budget.
Sales Copilot
Pipeline, deals, quotes, follow-ups, upsell, churn.
Procurement
Vendors, requisitions, RFQs, GRNs, spend analysis.
Budget & Planning
Variance, cost centres, scenarios, budget vs actual.
Real Estate
Projects, units, bookings, instalments, possession, RERA.
Service & AMC
AMC cover, job cards, technicians, warranty, breakdowns.
Support Ops
Tickets, SLA clocks, CSAT, incidents, desk load.
Recurring Revenue
MRR, ARR, dunning, renewals, plan mix.
Contracts
Clauses, obligations, renewals, signatures, NDAs.
Tender & Bid
Tenders, BOQs, EMD, eligibility, L1 position.
Manufacturing
Production, work orders, BOM, MRP, OEE, shop floor.
Distribution
Beats, van sales, schemes, primary vs secondary sales.
Retail & POS
Tills, counter sales, footfall, basket size, loyalty.
Construction & Projects
RA bills, WBS, subcontractors, EVM, measurement books.
A copilot whose pack is switched off is filtered out before scoring even begins, so it cannot be reached by phrasing a question cleverly.
Registries, not prompts
Every AI behaviour in the product is a typed entry in a registry that ships with tests — which is why we can tell you exactly how many there are, and what each one is permitted to do.
22 personas
The role it answers as
Nine built in — accountant, auditor, GST practitioner, CFO, inventory planner, collections manager, procurement advisor, compliance officer and sales copilot — plus thirteen contributed by the industry packs. A persona only frames the lens; it never changes a figure. Pack personas are gated with their pack, so a tenant never sees a role for a module it does not run.
25 narration kinds
Where words are allowed
Executive briefs, budget variance, KPI scorecards, cash-flow forecasts, daily sales briefs and one advisor per module. Each is a registered kind with its own instruction, its own token ceiling, and the same authoritative-data header: use only these figures, never invent. Narration is not a free-text channel into the product — it is an enumerated list.
10 agent tools
Where actions are allowed
Ten registered actions, each carrying a risk level, and nine of the ten requiring a person to approve before anything runs. A planned step naming an unknown tool, or one outside that agent’s allow-list, or one with invalid arguments, is dropped at validation — refused before the database is touched, not attempted and reversed. Nothing skips your permissions or the audit trail.
Seven forecasts, none of them guessed
Prediction is the place most products quietly hand the work to a model. Ours are ordinary statistics you could reproduce in a spreadsheet — which is precisely the point, because a forecast you cannot reproduce is a forecast you cannot defend in a board meeting.
Cash-flow
Running-balance roll-forward over weekly inflow and outflow buckets — lowest balance and the first deficit week.
Demand
Moving average, exponential smoothing or seasonal index; safety stock from service level, deviation and lead time.
Revenue
Least-squares linear trend, multiplied by monthly seasonal factors once there are two years of history.
Sales
Weighted pipeline — amount by probability — across five scenario multipliers, with run-rate as an independent baseline.
Stock
Sales velocity into days of cover, stockout date and reorder quantity from target cover plus lead time.
Rolling
Blends elapsed actuals with the remaining forecast. It only blends — it never fabricates the remainder.
Renewal
Date arithmetic over renewal anchors into monthly buckets, flagging anything without auto-renew inside the risk window.
The checks an auditor would run
Seven deterministic rules over the books you already keep. No model, no training data, no score you cannot interrogate — every finding names the rule it broke and the rows that broke it, and re-running on the same books returns the same list.
- Two vendor masters sharing a GSTIN, PAN or exact name
- The same vendor and amount paid twice inside a short window
- A bill number entered more than once — the double-ITC trap
- Suspiciously exact round-figure amounts at high value
- A voucher entered long after the date it carries
- Purchases, payments or expenses dated on a Saturday or Sunday
- ITC claimed with no supplier GSTIN, or at an impossible rate
Findings are guidance for a human to review — never an automated block, and never phrased as an accusation.
Twenty-four tools that never phone home
The strongest proof that deterministic-first is real rather than rhetorical: an entire industry AI suite with no gateway call in it at all. Expiry radar and rescue actions, Schedule-H and controlled-drug audits, scheme and landed-cost optimisation, same-salt substitution when an item is short, drug-licence expiry across your own books and your parties’ — all computed on device from your own data.
It works in a shop with the router unplugged, and nothing leaves the machine.
On device
Pure functions over your own batch, expiry, scheme and licence data. No network call exists in the code path.
Offline
The counter keeps its intelligence when the connection does not. Nothing degrades, nothing queues.
Deterministic
Ranked and bucketed by rules you can read. The same stock position always produces the same worklist.
Yours only
No third-party service, no shared corpus, no data leaving the premises to make a suggestion.
Turn the AI off. Keep the business.
This is the test every AI-first product should have to pass and most would fail. AI can be switched off by plan or by a single setting on your deployment. When it is off, the narration call returns nothing — and the computed result it would have described comes back exactly as before.
Around fifteen screens are built for this state on purpose. The variance analysis, the budget advice, the construction dashboard: the figures render, and where the written explanation would sit there is a line telling you it is off. Nothing breaks, nothing empties, nothing needs a fallback because the numbers were never coming from the model.
Every AI-assisted response is also recorded with whether it was served with AI on or deterministically — so you can audit after the fact how much of what you were told was arithmetic.
On the Cloud Suite
AI is a plan entitlement. Below Enterprise the copilot is a locked panel rather than an error, and every deterministic insight, report and dashboard behaves identically.
On the Desktop Suite
The assistant is the one feature in an otherwise fully offline app that reaches the internet — it proxies to the same gateway, holding no keys of its own. Everything else, including all of Pharma AI, runs unplugged.
On your terms
Provider keys never touch your machine or your install. Requests are authenticated, quota-checked and metered against your plan, and the model behind the gateway stays our responsibility.
What we deliberately don’t claim
Four things you will find on almost every competing AI page and will not find on ours. Each omission is a decision recorded in the code, and each one is worth more to you than the claim would have been.
We don’t name a model
Everything reaches you through one managed LekhaPro Cloud AI gateway. Keys live there and never touch your machine, usage is metered to your plan, and the model behind it is our problem to keep current — not a dependency we hand you.
We don’t claim vector search
Retrieval over your documents and knowledge hub is keyword and full-text based. We could ship embeddings and say the word; we would rather ship retrieval you can argue with, because a keyword match can be explained and a vector cannot.
There is no AI HR, recruiter or payroll
Not a roadmap answer — a data answer. LekhaPro holds no payroll or employee-compensation data anywhere in either edition, so there is nothing for such a feature to reason over and we do not sell one.
It’s anomaly detection, not “AI fraud detection”
Seven deterministic auditor’s rules over your own books. Findings are ranked guidance for a human to review, never an automated action and never a machine-learned accusation.
Questions
Can the AI invent a number?+
Not structurally. Figures are produced by tested engines before any model is involved, and the model is handed the finished result under an instruction to use only those figures. The narration layer restates arithmetic it was given; it never performs it.
What happens if I switch the AI off?+
Every number stays exactly where it was. AI can be disabled by plan or by a single environment setting, and the narration call then returns nothing at all while the deterministic result it would have described is returned as normal. Around fifteen screens are built for this: the analysis renders, and only the written explanation is replaced by a line telling you it is off.
Which AI model or provider do you use?+
We deliberately do not publish one. AI reaches you as a single managed LekhaPro Cloud AI gateway that is authenticated, quota-enforced and metered against your plan. Provider keys stay on the gateway and never touch your machine or your Desktop install.
Is this semantic or vector search?+
No. Retrieval is keyword and text based — term overlap over your own document chunks, plus Postgres full-text indexes for the document library, contracts and the HSN catalogue. There are no embeddings and no vector store, and we would rather say so than claim a technique we do not use.
Can an agent change my data on its own?+
Ten actions are registered, each with a risk level, and nine of the ten require a human to approve before anything runs. A step naming a tool that does not exist, or one outside that agent’s allow-list, or one with invalid arguments, is dropped at validation — it is never attempted and then rolled back. Nothing bypasses your normal permissions or the audit trail.
Does Pharma AI need the internet?+
No. The Pharma AI tools are deterministic computations over your own batch, expiry, scheme, salt and drug-licence data, and they run on device with nothing leaving the machine. They are the part of the AI story that works in a basement chemist shop with no connection at all.
Does the Desktop edition need to be online for AI?+
Yes — the assistant is the one feature in an otherwise fully offline Desktop app that reaches the internet, because it proxies to the same cloud gateway. Everything else, including the entire deterministic layer and all of Pharma AI, runs with the network unplugged.
Why does the analytics assistant sometimes refuse to answer?+
Because it is governed. A question is resolved against your semantic model, and if it names data you do not have or a measure that does not exist, it is rejected before any query is built rather than answered approximately. We consider a refusal a correct answer, and there is a test in the suite that keeps it that way.
Which plans include the AI?+
The deterministic layer — every computed insight, forecast, anomaly check and dashboard — is in every plan, because none of it involves a model. Conversational AI such as the Copilot and the AI CFO is an Enterprise entitlement with a token quota included. On plans below Enterprise the copilot appears as a locked panel rather than an error, and nothing deterministic behaves any differently.
Why is there no AI for HR or payroll?+
Because there is no payroll data for it to reason over. LekhaPro holds no payroll or employee-compensation data in either edition, so an AI HR feature would have nothing behind it — and the governed analyst treats that as a hard boundary. Ask it about payroll costs and the question resolves to a dataset that does not exist and is refused, with the reason attached. That refusal is a regression test in the suite, not a marketing position.
See it answer — and see it refuse
Start on your own data. Ask it something it knows, then ask it something it cannot possibly know, and judge it on the second answer.