How lamp keeps your business safe

LAMP uses AI agents to do parts of a business's work, such as reading a photo of a bill, drafting a quotation or reminding a customer about a payment. This is how they work, in plain words. The full statement is Part B of our Acceptable Use Policy and AI statement.

AI never decides what happens next

A piece of work, such as one sales order moving from enquiry to payment, follows a fixed path through the business's processes. The next step is chosen by that path and by rules applied to the business's data. An agent works inside one step at a time. The only way it affects what happens next is by producing data, and the rules then decide.

Autonomy Levels: the business chooses how much agents do alone

For each kind of Task, an Owner chooses one of three Autonomy Levels:

• Auto: the agent does the work and reports it. People review the reports and the Audit Log.
• Ask: the agent prepares the work. A person checks it and confirms, edits or rejects it.
• Manual: the agent reminds and records. A person does the work.

Autonomy is earned. A kind of Task may run at Auto or Ask only once it has passed LAMP's tests at the required level. If a business's own record shows many corrections, LAMP lowers the level for that business. An Owner can always lower it.

Output Contracts: every result is checked before work moves on

Every Task has an Output Contract: the conditions its result must meet, such as required fields, allowed values, arithmetic that must add up and statutory checks like a valid GSTIN. If a result fails, the Task does not complete: the agent tries again or a person is asked. Money and stock move only when a Document such as an invoice is issued and posted by fixed rules, and issued Documents are never edited. Mistakes are corrected with a new Document, and each Document records which template version produced it.

Deviations: a step outside the rules waits for a person

When an agent sees that a Run needs a step outside the business's rules, for example shipping part of an order or giving an unusual discount, it can only propose it. That proposal is a Deviation. It goes to the person the business's Approval Rules name, and once approved it applies to that one Run only. The business can then choose to keep it as a rule. Agents never approve their own Deviations.

Kill Switches: agents can be stopped at once

LAMP can stop agents immediately, for one kind of Task, one business, one group of servers or all of LAMP. An Owner can stop all agents for their business at any time from the Cockpit or with the stop command in a Channel. While agents are stopped, the business keeps running at Manual: people can still do their Tasks, see their data and issue Documents.

Talk to a person

In every conversation an agent handles, anyone can ask for a person by typing "talk to a person" or tapping the button where the Channel shows one. Staff reach the person responsible under the business's Roles, and finally an Owner. A customer or supplier reaches the person the Owner has named for that Channel, and the agent stops replying until a person hands the conversation back. If someone sincerely asks whether they are talking to an assistant, the agent says so.

AI makes mistakes. The checks catch many, not all. Messages go out in the business's name, and the business is responsible for them.

LAMP runs on servers in India. Some other companies process data for LAMP: they host it, run the AI models agents use, carry messages on Telegram and WhatsApp, send email and collect LAMP's own fees.

• AI model providers receive only the text a Task needs, with names and numbers masked where the Task allows. Some of them process data outside India, and our list says where.
• We use provider terms under which a business's data is not used to train their models.
• Anything that came to LAMP through WhatsApp is never used to improve LAMP for other businesses.
• Account Holders hear from us at least 30 days before we add a company or change what one receives or where it processes it.

Every company, what it receives and where it processes it is on our list of sub-processors.

How LAMP keeps business data safe. Our threat model covers every place data crosses a boundary, and each threat in it has at least one attack scenario that our tests play against LAMP.

Each business's data is kept apart

Businesses share servers, but the database itself keeps each business's rows apart, and a leak test tries to read one business's data as another on every change. Agents never have more access than the person they act for, and those limits are enforced in the database, not only in the agent's instructions.

What an agent reads never gives it authority

Purchase orders, photos of invoices and supplier messages are treated as data, never as instructions. An agent has only the tools its Task needs, messages to customers and suppliers are prepared at Ask unless an Owner chooses otherwise, and no agent can change Roles, Approval Rules or the business's processes without an Owner's confirmation.

Data is encrypted, and sensitive fields have their own keys

Data is encrypted at rest. Sensitive fields such as bank details have a separate key per business, kept apart from the database, so destroying a business's key makes its copies in old backups unreadable. PINs are stored only as hashes.

Backups only count once they have been restored

The database is backed up continuously, with 35 days of point-in-time recovery, and the backups are encrypted with keys held apart from the data. A restore drill regularly restores a backup into a separate copy of LAMP and checks that every ledger still balances, stock never goes negative and no Document number is missing. We publish the recovery a drill has actually achieved, not a target we have not yet met.

People at LAMP

Nobody at LAMP sees a business's unmasked data in the normal course of work. Access is allowed only for an incident, a support request the business raised, a legal order or work the business has agreed to in writing. It is time-limited, and each access is logged in the business's own Audit Log.

Incidents

If something goes wrong, we tell the affected Owners in their own Channel. For the most serious incidents we also publish a short note here, with no personal data, and a readable account of what happened and what we changed.

Every period we publish a report with totals across every business on LAMP, never broken down by business: Kill Switch uses and why, security incidents reported to CERT-In or to businesses, complaints and how quickly they were resolved, requests from authorities for business data, how often agents' results failed their checks, changes to Autonomy Level limits and changes to our AI model providers and sub-processors. Any number under 5 is given as "fewer than 5", so that no business can be identified.

We follow the India AI Governance Guidelines published by MeitY in November 2025 on grievance handling and transparency. They are guidelines, not a certification.